Building scraping pipelines with Dagster in 2026

Building scraping pipelines with Dagster is the right answer when scraping is one part of a larger data engineering pipeline. Dagster reframes orchestration around data assets rather than tasks: instead of writing “fetch URLs, parse, store” as separate steps, you declare “products dataset” as a software-defined asset that depends on “raw HTML pages” which depends on “URL list,” and Dagster figures out the execution graph. For pure scraping with no downstream analytics, this is overkill. For scraping that feeds into ML models, dashboards, or warehouses, Dagster’s asset-aware approach pays off in lineage, partitioning, and freshness tracking.

This guide covers Dagster 1.7+ in 2026 for scraping pipelines: software-defined assets, jobs, sensors, partitioning by time and host, IO managers, and a complete end-to-end example. Code is Python 3.12. By the end you will know whether Dagster fits your scraping use case and how to deploy a working pipeline.

Why Dagster for scraping

Dagster’s strengths for scraper-heavy pipelines:

• Software-defined assets: declare data products instead of orchestration steps
• Lineage tracking: see how data flows from raw HTML to final dashboard
• Partition support: per-day, per-host, per-customer partitions with backfill
• IO managers: separate “what to compute” from “where to store it”
• Sensors: trigger pipelines based on external events (file arrival, API webhook)
• Strong type system: Pydantic-style configs and asset metadata
• Asset checks: data quality assertions per asset
• Mature: 6+ years in production, used by Robinhood, Duolingo, etc.

For Dagster’s official documentation, see docs.dagster.io.

Where Dagster does not lead

• For simple scheduled scraping with no downstream pipeline, Prefect or cron are simpler
• For very high-frequency tasks (1000s/sec), the asset model adds overhead
• For teams without data engineering experience, the asset abstraction takes adjustment

For Prefect comparison, see building scraping pipelines with Prefect 3.

Installing Dagster

pip install dagster dagster-webserver
dagster --version  # 1.7+

# Initialize a project
dagster project scaffold --name my-scraper
cd my-scraper

Start the dev server:

dagster dev
# UI at http://localhost:3000

A first asset: scraping a URL

# my_scraper/assets.py
from dagster import asset, AssetExecutionContext
import httpx

@asset
def homepage_html(context: AssetExecutionContext) -> str:
    resp = httpx.get("https://example.com", headers={
        "User-Agent": "Mozilla/5.0 ...",
    })
    resp.raise_for_status()
    context.log.info(f"Fetched {len(resp.text)} bytes")
    return resp.text

@asset
def homepage_titles(homepage_html: str) -> list[str]:
    from selectolax.parser import HTMLParser
    tree = HTMLParser(homepage_html)
    return [n.text(strip=True) for n in tree.css("h2")]

Two assets: homepage_html (the raw HTML) and homepage_titles (parsed titles, depends on homepage_html). Dagster figures out the dependency from the function signature.

In the Dagster UI, you see a graph with two nodes connected by a line. Click “Materialize all” and Dagster fetches the page and parses it.

Partitioned assets: scrape per day, per host

For real scraping, you usually want partitions: one materialization per day, per host, or per customer. Dagster partitions track which slices have been materialized and which need backfilling.

from dagster import asset, DailyPartitionsDefinition, AssetExecutionContext
from datetime import datetime
import httpx

daily_partitions = DailyPartitionsDefinition(start_date="2026-01-01")

@asset(partitions_def=daily_partitions)
def daily_news_html(context: AssetExecutionContext) -> str:
    date_str = context.partition_key  # YYYY-MM-DD
    url = f"https://news.example.com/archive/{date_str}"
    resp = httpx.get(url)
    resp.raise_for_status()
    return resp.text

Now daily_news_html has one materialization per day. Run today’s: Dagster fetches today’s URL. Backfill last 30 days: Dagster fetches each day’s URL.

For multi-dimensional partitions (per-day AND per-host):

from dagster import (
    asset, MultiPartitionsDefinition, StaticPartitionsDefinition,
    DailyPartitionsDefinition,
)

hosts = StaticPartitionsDefinition(["example.com", "news.example.com", "shop.example.com"])
date_partitions = DailyPartitionsDefinition(start_date="2026-01-01")

multi_partitions = MultiPartitionsDefinition({"date": date_partitions, "host": hosts})

@asset(partitions_def=multi_partitions)
def host_daily_pages(context: AssetExecutionContext) -> str:
    keys = context.partition_key.keys_by_dimension
    date_str = keys["date"]
    host = keys["host"]
    url = f"https://{host}/archive/{date_str}"
    resp = httpx.get(url)
    return resp.text

This creates a 2D grid of partitions. Backfill specific cells (date X host) without re-running everything.

IO managers: separating compute from storage

In simple cases, the asset function returns a Python value and Dagster pickles it to local disk. For production, use IO managers to write to Postgres, S3, or your warehouse.

from dagster import IOManager, io_manager, OutputContext, InputContext
import boto3
import json
import io

class S3IOManager(IOManager):
    def __init__(self, bucket: str):
        self.bucket = bucket
        self.s3 = boto3.client("s3")

    def _key(self, context):
        # asset_key is a list of strings
        path = "/".join(context.asset_key.path)
        if context.has_partition_key:
            path = f"{path}/{context.partition_key}"
        return f"{path}.json"

    def handle_output(self, context: OutputContext, obj):
        key = self._key(context)
        body = json.dumps(obj).encode() if not isinstance(obj, bytes) else obj
        self.s3.put_object(Bucket=self.bucket, Key=key, Body=body)
        context.log.info(f"Wrote to s3://{self.bucket}/{key}")

    def load_input(self, context: InputContext):
        key = self._key(context)
        resp = self.s3.get_object(Bucket=self.bucket, Key=key)
        return json.loads(resp["Body"].read())

@io_manager(config_schema={"bucket": str})
def s3_io_manager(init_context):
    return S3IOManager(bucket=init_context.resource_config["bucket"])

Wire it in:

from dagster import Definitions

defs = Definitions(
    assets=[homepage_html, homepage_titles],
    resources={"io_manager": s3_io_manager.configured({"bucket": "my-scraper-data"})},
)

Now every asset’s output goes to S3 instead of local disk. Same code, different storage.

Browser scraping with Playwright as an asset

from dagster import asset, AssetExecutionContext
from playwright.sync_api import sync_playwright

@asset(partitions_def=daily_partitions)
def dynamic_page_html(context: AssetExecutionContext) -> str:
    date_str = context.partition_key
    url = f"https://example.com/dynamic?date={date_str}"

    with sync_playwright() as p:
        browser = p.chromium.launch(headless=True)
        page = browser.new_page()
        page.goto(url, wait_until="networkidle")
        html = page.content()
        browser.close()
    return html

Same pattern: Playwright runs inside the asset function. Dagster orchestrates.

Jobs: grouping assets

Jobs run sets of assets together. Define a job that materializes a related group:

from dagster import define_asset_job, AssetSelection

scrape_news_job = define_asset_job(
    "scrape_news",
    selection=AssetSelection.assets("daily_news_html", "daily_news_articles"),
)

Schedule it:

from dagster import ScheduleDefinition

daily_news_schedule = ScheduleDefinition(
    job=scrape_news_job,
    cron_schedule="0 6 * * *",  # 6am daily
)

Sensors: event-driven materialization

Sensors trigger asset materialization based on external events:

from dagster import sensor, RunRequest, SkipReason
from datetime import datetime
import httpx

@sensor(job=scrape_news_job)
def new_url_sensor(context):
    # Check an API for new URLs
    resp = httpx.get("https://api.example.com/new-urls")
    new_urls = resp.json()

    if not new_urls:
        return SkipReason("No new URLs")

    return RunRequest(
        run_key=str(datetime.utcnow().timestamp()),
        run_config={"ops": {"daily_news_html": {"config": {"urls": new_urls}}}},
    )

Dagster runs the sensor function periodically (default 30 seconds). When it returns a RunRequest, the job triggers.

Asset checks: data quality

Built-in data quality assertions per asset:

from dagster import asset, AssetCheckResult, asset_check

@asset
def products_data(homepage_html: str) -> list[dict]:
    # Parse and return products
    return [{"title": "...", "price": "..."}]

@asset_check(asset=products_data)
def check_min_products(context, products_data):
    count = len(products_data)
    return AssetCheckResult(
        passed=count >= 10,
        metadata={"count": count, "expected_min": 10},
    )

@asset_check(asset=products_data)
def check_no_missing_prices(context, products_data):
    missing = sum(1 for p in products_data if not p.get("price"))
    return AssetCheckResult(
        passed=missing == 0,
        metadata={"missing_count": missing},
    )

After every materialization, Dagster runs the checks and shows pass/fail in the UI. Fail an SLA, fail a build, alert the team.

Comparison: Dagster vs Prefect for scraping

Pick Dagster when scraping feeds analytics, ML, or BI. Pick Prefect when scraping is the end product or when async-heavy.

Complete production example

# my_scraper/definitions.py
from dagster import (
    asset, AssetExecutionContext, Definitions,
    DailyPartitionsDefinition, define_asset_job, ScheduleDefinition,
    AssetSelection, IOManager, io_manager, Resource,
)
from dagster_aws.s3 import S3PickleIOManager, S3Resource
import httpx
from selectolax.parser import HTMLParser
import json
from typing import Optional

daily_partitions = DailyPartitionsDefinition(start_date="2026-01-01")

@asset(partitions_def=daily_partitions, group_name="ingestion")
def raw_pages(context: AssetExecutionContext) -> dict[str, str]:
    """Fetch all configured pages for the partition date."""
    date_str = context.partition_key
    urls = [
        f"https://example.com/products?date={date_str}&page={i}"
        for i in range(1, 11)
    ]
    pages = {}
    for url in urls:
        try:
            resp = httpx.get(url, timeout=30)
            resp.raise_for_status()
            pages[url] = resp.text
            context.log.info(f"Fetched {url}: {len(resp.text)} bytes")
        except Exception as e:
            context.log.error(f"Failed {url}: {e}")
    return pages

@asset(partitions_def=daily_partitions, group_name="parsing")
def parsed_products(context: AssetExecutionContext, raw_pages: dict) -> list[dict]:
    """Parse products from all raw pages."""
    products = []
    for url, html in raw_pages.items():
        tree = HTMLParser(html)
        for el in tree.css("article.product"):
            title_el = el.css_first("h2")
            price_el = el.css_first(".price")
            link_el = el.css_first("a")
            products.append({
                "title": title_el.text(strip=True) if title_el else None,
                "price": price_el.text(strip=True) if price_el else None,
                "url": link_el.attributes.get("href") if link_el else None,
                "source_url": url,
                "scraped_date": context.partition_key,
            })
    context.log.info(f"Parsed {len(products)} products")
    return products

@asset(partitions_def=daily_partitions, group_name="storage")
def products_in_db(context: AssetExecutionContext, parsed_products: list) -> int:
    """Write parsed products to Postgres."""
    import asyncpg
    import asyncio

    async def write():
        conn = await asyncpg.connect("postgresql://user:pass@localhost/scraper")
        try:
            await conn.executemany(
                """INSERT INTO products
                   (title, price, url, source_url, scraped_date)
                   VALUES ($1, $2, $3, $4, $5)
                   ON CONFLICT (url, scraped_date) DO NOTHING""",
                [(p["title"], p["price"], p["url"], p["source_url"], p["scraped_date"])
                 for p in parsed_products],
            )
            return len(parsed_products)
        finally:
            await conn.close()

    count = asyncio.run(write())
    context.log.info(f"Stored {count} products to DB")
    return count

@asset_check(asset=parsed_products)
def check_minimum_products(context, parsed_products):
    return AssetCheckResult(
        passed=len(parsed_products) >= 50,
        metadata={"count": len(parsed_products), "expected_min": 50},
    )

scrape_job = define_asset_job(
    "daily_scrape",
    selection=AssetSelection.assets("raw_pages", "parsed_products", "products_in_db"),
)

scrape_schedule = ScheduleDefinition(
    job=scrape_job,
    cron_schedule="0 6 * * *",
    execution_timezone="UTC",
)

defs = Definitions(
    assets=[raw_pages, parsed_products, products_in_db],
    asset_checks=[check_minimum_products],
    jobs=[scrape_job],
    schedules=[scrape_schedule],
    resources={
        # Add IO managers, DB connections, etc.
    },
)

This pipeline:

1. Fetches raw pages daily, partitioned by date
2. Parses products from raw HTML
3. Stores to Postgres with idempotent insert
4. Asserts minimum product count
5. Runs daily at 6am UTC

In the Dagster UI, you see the asset graph, materialization history, and check pass/fail per partition.

Deployment

For production:

• Dagster Cloud (managed): easiest, paid per user + execution
• Self-hosted on Kubernetes: dagster-helm chart, full control
• Self-hosted on a single VM: docker-compose with PostgreSQL, fine for small teams

For Kubernetes:

# values.yaml for dagster-helm
dagsterDaemon:
  enabled: true
runLauncher:
  type: K8sRunLauncher
postgresql:
  enabled: true

Dagster spawns a Kubernetes pod per run, isolating resources.

Operational checklist

For production Dagster scraping in 2026:

• Dagster 1.7+ with Python 3.11+
• Dagster Cloud or self-hosted on Kubernetes
• Partitions defined for backfill capability
• IO managers for production storage (S3, Postgres)
• Asset checks for data quality
• Sensors for event-driven workflows
• Schedules for cron-style runs
• Type hints on all asset functions
• Resource configs separate from code

For broader pipeline patterns, see building scraping pipelines with Prefect 3 and distributed scraping with Apache Kafka.

Common pitfalls

• Asset name conflicts: two assets with the same key error out. Use key_prefix to namespace.
• Partition explosion: multi-dimensional partitions can produce millions of cells. Plan partitioning carefully.
• Sync code in async contexts: Dagster is sync-first; for async, use asyncio.run inside the asset.
• Long-running assets: Dagster default timeout is short. Configure execution timeout per asset.
• IO manager mismatches: changing IO managers mid-flight can cause input/output type mismatches.

FAQ

Q: should I use Dagster or Prefect for pure scraping?
For pure scraping, Prefect’s simpler model is usually a better fit. Dagster pays off when scraping is part of a larger ETL or ML pipeline.

Q: can I migrate from Airflow to Dagster?
Yes, with effort. The asset model is different from DAGs. Most teams migrate one pipeline at a time over months.

Q: does Dagster work with Scrapy?
Yes. Wrap Scrapy invocations as assets. The orchestration value is at the pipeline level, not within individual spiders.

Q: how does cost compare to Prefect?
Self-hosted: similar (just compute). Cloud: Dagster Cloud is more expensive per user, but pricing for execution is comparable.

Q: is Dagster overkill for a 5-spider project?
Probably yes. For 5 unrelated scrapers, cron + Python scripts is fine. For 5 scrapers feeding a unified analytics warehouse, Dagster shines.

Common pitfalls in production Dagster scraping

The first failure mode is the partition backfill blast radius. Dagster’s backfill UI lets you select multiple partitions and re-materialize them. A typo in the partition selector (selecting “all 2026 partitions” when you meant “this week’s partitions”) triggers 365 simultaneous materialization runs, which queues 365 Kubernetes pods, exhausts your cluster’s pod limit, and starves all other pipelines for hours. The fix is to enable backfill concurrency limits in the Dagster instance config, capping simultaneous backfill runs at a sane number:

# dagster.yaml
run_coordinator:
  module: dagster._core.run_coordinator
  class: QueuedRunCoordinator
  config:
    max_concurrent_runs: 25
    tag_concurrency_limits:
      - key: dagster/backfill
        limit: 5

This caps backfills at 5 concurrent runs while leaving 20 slots for scheduled and ad-hoc runs. Without this, a single mis-clicked backfill can take down your entire pipeline cluster.

The second pitfall is asset definition import-time side effects. Dagster imports your asset definitions on every code-server reload to build the asset graph. If your asset definition file contains import-time code that fetches from an external service (loading site configs from a database at module import), every code reload triggers that fetch. With 50 developers reloading code 100 times per day, you generate 5000 daily fetches that the upstream service was never designed for, and you sometimes get throttled by your own configuration store. The fix is to defer all external calls to inside the asset function or use a Dagster Resource that lazy-loads:

from dagster import resource, asset, ConfigurableResource

class SiteConfigResource(ConfigurableResource):
    config_url: str
    _cache: dict | None = None

    def get_configs(self) -> dict:
        if self._cache is None:
            self._cache = httpx.get(self.config_url).json()
        return self._cache

@asset
def raw_pages(context, site_configs: SiteConfigResource):
    configs = site_configs.get_configs()  # only fetched when asset runs
    # ...

The third pitfall is the IO manager memory bloat on large asset values. Dagster’s default IO managers (S3PickleIOManager, FilesystemIOManager) serialize the entire asset value to storage between assets. A raw_pages asset returning 10,000 HTML pages of 50KB each is 500MB of data. The pickle serialize-deserialize cycle runs in a single process, peaking at 1GB+ resident memory. Workers OOM-kill, the run fails, and Dagster retries it, blowing through your retry budget. The fix is custom IO managers that stream rather than buffer, or asset partitioning that keeps individual asset values under 100MB:

class StreamingS3IOManager(IOManager):
    def handle_output(self, context, obj):
        # obj is an iterator, not a list
        s3_key = f"{context.asset_key.path[-1]}/{context.partition_key}.jsonl"
        with self.s3_client.open(s3_key, "wb") as f:
            for item in obj:
                f.write(json.dumps(item).encode() + b"\n")

    def load_input(self, context):
        s3_key = f"{context.asset_key.path[-1]}/{context.partition_key}.jsonl"
        return (json.loads(line) for line in self.s3_client.iter_lines(s3_key))

Then design your assets to yield items rather than return lists. Memory usage drops from O(N) to O(1) per asset, regardless of partition size.

Real-world example: Dagster + dbt for ecommerce intelligence

A scraping team built a Dagster pipeline that scraped 30 ecommerce sites daily and fed a dbt-managed warehouse for downstream analytics. The architecture treated scraped data as Dagster assets that materialized into Snowflake, and dbt models as downstream Dagster assets that ran SQL transformations against those tables:

from dagster import asset, AssetExecutionContext, DailyPartitionsDefinition
from dagster_dbt import DbtCliResource, dbt_assets, get_asset_key_for_model
from pathlib import Path

DBT_PROJECT_DIR = Path(__file__).parent / "ecom_dbt"
daily = DailyPartitionsDefinition(start_date="2026-01-01")

@asset(partitions_def=daily, group_name="raw_scrape")
def raw_amazon_products(context: AssetExecutionContext):
    date_str = context.partition_key
    products = scrape_amazon(date=date_str)
    write_to_snowflake("raw.amazon_products", products)
    context.add_output_metadata({
        "row_count": len(products),
        "preview": products[:5],
    })

@asset(partitions_def=daily, group_name="raw_scrape")
def raw_walmart_products(context: AssetExecutionContext):
    date_str = context.partition_key
    products = scrape_walmart(date=date_str)
    write_to_snowflake("raw.walmart_products", products)
    context.add_output_metadata({"row_count": len(products)})

@dbt_assets(
    manifest=DBT_PROJECT_DIR / "target" / "manifest.json",
    partitions_def=daily,
)
def ecom_dbt_assets(context: AssetExecutionContext, dbt: DbtCliResource):
    yield from dbt.cli(["build"], context=context).stream()

The dbt project contained models like staging_amazon_products (cleans the raw scrape), mart_competitive_pricing (joins all sites for cross-comparison), and ml_features_pricing (rolling windows for the pricing prediction model).

In the Dagster UI, the asset lineage showed the full chain: scrape job -> raw tables -> staging models -> mart models -> ML feature tables. When the pricing prediction model returned anomalous results, the team traced the lineage backward and discovered the root cause was a 24-hour gap in the Walmart scrape (the site had pushed an HTML structure change that broke the parser). The lineage view turned a 4-hour debugging session into a 15-minute one.

Asset checks at each layer caught quality issues before they propagated:

@asset_check(asset=raw_amazon_products)
def amazon_row_count_check(context, raw_amazon_products):
    rows = query_snowflake(
        f"SELECT COUNT(*) FROM raw.amazon_products WHERE scraped_date = '{context.partition_key}'"
    )[0][0]
    return AssetCheckResult(
        passed=rows >= 5000,
        severity=AssetCheckSeverity.ERROR if rows < 1000 else AssetCheckSeverity.WARN,
        metadata={"row_count": rows, "expected_min": 5000},
    )

A check failure marked the partition as unhealthy in the UI and prevented downstream dbt models from running until the issue was resolved. Over six months in production, asset checks caught 47 distinct data quality regressions before they reached the analytics layer.

Comparison: Dagster patterns by scraping scale

A reference table of Dagster patterns that work well at different scraping volumes:

For most scraping workloads, the 1K-100K URLs/day band with DailyPartitionsDefinition is the right starting point. Migrate up the table only when partitioning becomes the bottleneck. Migrate down (to a simpler tool) if Dagster’s overhead exceeds your benefit.

Detection: when Dagster is the wrong choice for scraping

Five signals that your scraping workload should NOT live on Dagster:

1. Pure scraping with no downstream ETL: if all you do is scrape and dump to JSON files, Dagster’s asset model is overhead. Use Prefect or cron + scripts.
2. Heavy async workload (1000+ concurrent requests): Dagster is sync-first. Heavy async work runs better in Prefect or Celery.
3. Sub-minute scheduling: Dagster’s minimum schedule interval is one minute. Sub-minute requires sensors or external triggers.
4. Fluid pipeline shape that changes per run: Dagster’s asset model assumes a stable graph. Per-run dynamic shape is awkward.
5. Real-time event-driven only: pure event-driven workloads work better in event-streaming systems (Kafka, NATS) than in orchestrators.

If three or more of these apply, Prefect or a simpler tool is the right answer.

Performance tuning: parallel asset execution

Dagster’s default behavior runs assets sequentially within a job. For scraping where assets are independent (different sites), enable multi-process or k8s execution:

from dagster import multiprocess_executor, define_asset_job

scrape_job = define_asset_job(
    "all_sites",
    selection="*",
    executor_def=multiprocess_executor.configured({
        "max_concurrent": 8,
    }),
)

This runs up to 8 site-scraping assets in parallel within one job run. For Kubernetes-based execution, use k8s_job_executor which spawns one pod per asset:

from dagster_k8s import k8s_job_executor

scrape_job = define_asset_job(
    "all_sites",
    executor_def=k8s_job_executor.configured({
        "max_concurrent": 30,
        "container_config": {
            "resources": {
                "requests": {"memory": "1Gi", "cpu": "500m"},
                "limits": {"memory": "4Gi", "cpu": "2000m"},
            },
        },
    }),
)

Per-asset isolation prevents one runaway scrape from OOM-killing the entire job. The trade-off is pod startup latency (5-15 seconds per asset), which is fine for hourly or daily jobs but punitive for sub-minute schedules.

Wrapping up

Dagster excels when scraping is one node in a larger data pipeline. The software-defined asset model gives lineage, partitioning, and freshness tracking that pure orchestrators do not. For teams that already think in terms of data products and downstream consumers, it pays off. Pair this with our building scraping pipelines with Prefect 3 and distributed scraping with Apache Kafka writeups for the full pipeline picture, and browse the dev-tools-projects category on DRT for related infrastructure deep-dives.

Building scraping pipelines with Prefect 3 in 2026

Building scraping pipelines with Prefect 3 is the right answer when your scraping has grown beyond a single Scrapy project but does not yet need full Apache Airflow’s complexity. Prefect 3 (released 2024) is the third major version of Prefect’s orchestration platform, and it brought meaningful changes from Prefect 2: simpler async support, faster task execution, work pools that decouple scheduling from execution, and a redesigned dashboard. For scraping specifically, Prefect 3 fits the pattern of “fetch URLs, parse them, store results, retry failures, schedule it all” with much less ceremony than Airflow.

This guide covers Prefect 3 for scraping pipelines in 2026: flows and tasks, work pools and workers, retries and concurrency control, scheduling, and a complete production example that fetches URLs, runs them through Playwright when needed, parses results, and stores to Postgres. By the end you will have a working pipeline pattern you can adapt to your specific targets.

Why Prefect for scraping

Prefect’s strengths for scraper pipelines:

• Python-native: flows and tasks are decorated Python functions, no DAG syntax to learn
• Async-friendly: tasks can be async def and run concurrently
• Built-in retries: configurable per task with exponential backoff
• Work pools and workers: decouple “what to run” from “where to run it”
• Cron, interval, and event-based scheduling: flexible
• Observability: dashboard, logs, run history all built in
• Cloud or self-hosted: free OSS or hosted Prefect Cloud
• Mature: 6+ years in production, rich ecosystem

For Prefect’s official documentation, see docs.prefect.io.

Where Prefect does not lead

• For very high frequency tasks (1000+ tasks per second), Prefect adds latency
• For data-engineering-heavy workloads with many integrations, Dagster has more native connectors
• For purely scheduled cron jobs without orchestration, simpler tools (cron itself, GitHub Actions) suffice

For scraping, Prefect’s mid-frequency, retry-rich, observable model fits well.

For Dagster comparison, see building scraping pipelines with Dagster.

Installing Prefect 3

pip install prefect>=3.0
prefect --version  # 3.x

Start the local Prefect server:

prefect server start
# Dashboard at http://localhost:4200

For production, use Prefect Cloud (managed) or self-host the server with PostgreSQL.

A first scraping flow

# scraper_flow.py
from prefect import flow, task
import httpx

@task(retries=3, retry_delay_seconds=10)
async def fetch_url(url: str) -> dict:
    async with httpx.AsyncClient(timeout=30) as client:
        resp = await client.get(url, headers={
            "User-Agent": "Mozilla/5.0 ...",
        })
        resp.raise_for_status()
        return {"url": url, "status": resp.status_code, "html": resp.text}

@task
def parse_titles(html: str) -> list[str]:
    from selectolax.parser import HTMLParser
    tree = HTMLParser(html)
    return [n.text(strip=True) for n in tree.css("h2.title")]

@flow(log_prints=True)
async def scrape_titles(urls: list[str]):
    for url in urls:
        result = await fetch_url(url)
        titles = parse_titles(result["html"])
        print(f"{url}: {len(titles)} titles")
    return "Done"

if __name__ == "__main__":
    import asyncio
    asyncio.run(scrape_titles([
        "https://example.com/page1",
        "https://example.com/page2",
    ]))

Run it:

python scraper_flow.py

The flow appears in the Prefect dashboard with full task history, retries, and logs.

Concurrency: parallel task execution

For parallel scraping, use Prefect’s task mapping or asyncio.gather:

from prefect import flow, task
import httpx
import asyncio

@task(retries=3, retry_delay_seconds=10)
async def fetch_url(url: str) -> dict:
    async with httpx.AsyncClient(timeout=30) as client:
        resp = await client.get(url)
        return {"url": url, "status": resp.status_code, "html": resp.text}

@flow(log_prints=True)
async def scrape_parallel(urls: list[str], concurrency: int = 10):
    semaphore = asyncio.Semaphore(concurrency)

    async def fetch_with_semaphore(url):
        async with semaphore:
            return await fetch_url(url)

    results = await asyncio.gather(
        *[fetch_with_semaphore(url) for url in urls],
        return_exceptions=True,
    )
    return results

Or using Prefect’s .map() for static parallelism:

@flow
def scrape_parallel_static(urls: list[str]):
    results = fetch_url.map(urls)
    return results

fetch_url.map(urls) schedules one task per URL and runs them via Prefect’s task runner. The default ConcurrentTaskRunner runs tasks in threads or processes; for async tasks use the ThreadPoolTaskRunner.

Work pools and workers

Prefect 3 splits the runtime model into:

• Flows / deployments: the work to run
• Work pools: configuration for runtime environments (Docker, Kubernetes, process)
• Workers: processes that pick up flow runs from work pools

This decouples scheduling from execution. You can have one Prefect server scheduling flow runs, with workers in different environments (laptop, Kubernetes cluster, EC2) picking up runs from the same work pool.

Create a work pool:

prefect work-pool create my-scraper-pool --type process
prefect worker start --pool my-scraper-pool

Deploy a flow to the pool:

from prefect import flow, serve

@flow
async def scrape_titles(urls: list[str]):
    # ...
    pass

if __name__ == "__main__":
    scrape_titles.serve(
        name="scraper-deployment",
        work_pool_name="my-scraper-pool",
        cron="0 * * * *",  # hourly
    )

Now the flow runs hourly on whatever worker picks it up.

Retries and error handling

Per-task retry config is the most common pattern:

@task(
    retries=5,
    retry_delay_seconds=[10, 30, 60, 120, 300],  # exponential backoff
)
async def fetch_url(url: str) -> dict:
    async with httpx.AsyncClient(timeout=30) as client:
        resp = await client.get(url)
        if resp.status_code >= 500:
            raise httpx.HTTPStatusError(
                f"Server error: {resp.status_code}",
                request=resp.request, response=resp,
            )
        if resp.status_code == 429:
            # Wait longer for rate limits
            await asyncio.sleep(60)
            raise httpx.HTTPStatusError("Rate limited", request=resp.request, response=resp)
        return {"url": url, "status": resp.status_code, "html": resp.text}

For fine-grained retry policies (only retry on certain exceptions, not others):

from prefect.tasks import task_input_hash

@task(
    retries=5,
    retry_delay_seconds=30,
    retry_condition_fn=lambda task, run_state: (
        run_state.is_failed() and "5" in str(run_state.message)
    ),
)
async def fetch_url(url: str) -> dict:
    # ...
    pass

This retries only on 5xx errors, not on 4xx (which usually means permanent failure).

Persistent state: caching task results

For tasks that should not re-run on the same input (deduplication), use Prefect’s caching:

from prefect.tasks import task_input_hash
from datetime import timedelta

@task(
    cache_key_fn=task_input_hash,
    cache_expiration=timedelta(hours=24),
)
async def fetch_url(url: str) -> dict:
    # ...
    pass

Now if you call fetch_url("https://example.com") twice within 24 hours, the second call returns the cached result without running. Useful when re-running flows during development.

Browser scraping with Playwright in Prefect

For pages requiring JavaScript:

from prefect import flow, task
from playwright.async_api import async_playwright

@task
async def fetch_with_browser(url: str) -> dict:
    async with async_playwright() as p:
        browser = await p.chromium.launch(headless=True)
        page = await browser.new_page()
        await page.goto(url, wait_until="networkidle")
        html = await page.content()
        await browser.close()
        return {"url": url, "html": html}

@flow
async def scrape_dynamic_pages(urls: list[str]):
    results = []
    for url in urls:
        results.append(await fetch_with_browser(url))
    return results

Each task call spawns a browser, which is expensive. For high-volume browser scraping, batch URLs per browser:

@task
async def fetch_batch_with_browser(urls: list[str]) -> list[dict]:
    results = []
    async with async_playwright() as p:
        browser = await p.chromium.launch(headless=True)
        for url in urls:
            page = await browser.new_page()
            try:
                await page.goto(url, timeout=30000)
                html = await page.content()
                results.append({"url": url, "html": html})
            except Exception as e:
                results.append({"url": url, "error": str(e)})
            finally:
                await page.close()
        await browser.close()
    return results

This amortizes browser startup over many URLs.

Scheduling

Prefect supports cron, interval, and event-based scheduling:

# Cron
scrape_flow.serve(
    name="hourly-scraper",
    cron="0 * * * *",
    work_pool_name="my-pool",
)

# Interval
scrape_flow.serve(
    name="every-15min",
    interval=timedelta(minutes=15),
    work_pool_name="my-pool",
)

# Event-based (triggered by webhooks, file events, etc.)
from prefect.events import DeploymentEventTrigger

scrape_flow.serve(
    name="on-event",
    triggers=[
        DeploymentEventTrigger(
            expect={"my.custom.event"},
            parameters={"urls": "{{ event.payload.urls }}"},
        )
    ],
)

Event-based scheduling is useful for “scrape this URL when a webhook fires” patterns.

A complete production example

End-to-end scraping pipeline:

# pipeline.py
from prefect import flow, task, get_run_logger
from prefect.tasks import task_input_hash
from datetime import timedelta
import asyncio
import httpx
from selectolax.parser import HTMLParser
import asyncpg
import json

DB_DSN = "postgresql://user:pass@localhost/scraper"

@task(retries=3, retry_delay_seconds=[10, 30, 60])
async def fetch_url(url: str) -> dict:
    async with httpx.AsyncClient(timeout=30) as client:
        resp = await client.get(url, headers={
            "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 ...",
        })
        resp.raise_for_status()
        return {"url": url, "status": resp.status_code, "html": resp.text}

@task
def parse_products(html: str) -> list[dict]:
    tree = HTMLParser(html)
    products = []
    for el in tree.css("article.product"):
        products.append({
            "title": el.css_first("h2").text(strip=True) if el.css_first("h2") else None,
            "price": el.css_first(".price").text(strip=True) if el.css_first(".price") else None,
            "url": el.css_first("a").attributes.get("href") if el.css_first("a") else None,
        })
    return products

@task(retries=3)
async def store_products(products: list[dict]):
    logger = get_run_logger()
    if not products:
        return
    conn = await asyncpg.connect(DB_DSN)
    try:
        await conn.executemany(
            """INSERT INTO products (title, price, url)
               VALUES ($1, $2, $3)
               ON CONFLICT (url) DO UPDATE SET
                 title = EXCLUDED.title,
                 price = EXCLUDED.price""",
            [(p["title"], p["price"], p["url"]) for p in products],
        )
        logger.info(f"Stored {len(products)} products")
    finally:
        await conn.close()

@flow(log_prints=True)
async def scrape_pipeline(urls: list[str], concurrency: int = 5):
    semaphore = asyncio.Semaphore(concurrency)

    async def process_one(url: str):
        async with semaphore:
            try:
                result = await fetch_url(url)
                products = parse_products(result["html"])
                await store_products(products)
                return {"url": url, "ok": True, "count": len(products)}
            except Exception as e:
                return {"url": url, "ok": False, "error": str(e)}

    results = await asyncio.gather(*[process_one(url) for url in urls])

    success = [r for r in results if r["ok"]]
    failed = [r for r in results if not r["ok"]]

    print(f"Scraped {len(success)} successfully, {len(failed)} failed")
    return {"success": len(success), "failed": len(failed)}

if __name__ == "__main__":
    URLS = [f"https://example.com/products?page={i}" for i in range(1, 21)]
    asyncio.run(scrape_pipeline(URLS))

Deploy with cron:

if __name__ == "__main__":
    URLS = [f"https://example.com/products?page={i}" for i in range(1, 21)]
    scrape_pipeline.serve(
        name="hourly-products",
        cron="0 * * * *",
        parameters={"urls": URLS, "concurrency": 5},
        work_pool_name="my-pool",
    )

Comparison: Prefect vs other orchestrators

For scraping pipelines specifically, Prefect’s async support and simpler API give it an edge. Dagster’s data-engineering features (asset-based, type system, integrations) are overkill for pure scraping but useful when you mix scraping with downstream ETL.

Deployment patterns

Production Prefect 3 deployments:

• Prefect Cloud + workers on EC2/GCE: hosted control plane, your compute
• Self-hosted Prefect server + Kubernetes workers: full self-host
• Prefect Cloud + serverless workers (ECS, Lambda): pay-per-use compute
• All-in-one VM: server + worker on same machine for small teams

For containerized workers:

FROM python:3.12-slim

RUN pip install prefect playwright httpx selectolax asyncpg
RUN playwright install chromium

COPY . /app
WORKDIR /app

CMD ["prefect", "worker", "start", "--pool", "my-scraper-pool"]

Run multiple worker containers behind the same pool for horizontal scaling.

Operational checklist

For production Prefect 3 scraping in 2026:

• Prefect 3.x with Python 3.11+
• Prefect Cloud or self-hosted server with PostgreSQL
• Work pools for environment isolation
• Multiple workers for horizontal scale
• Configure retries with exponential backoff
• Use task caching for development iteration
• Async tasks for network-bound work
• Browser scraping in batches per task
• External storage (Postgres, S3) for results
• Monitor flow run success rate, task retry counts
• Alert on flows that have not completed within SLA

For broader pipeline patterns, see building scraping pipelines with Dagster and distributed scraping with Apache Kafka.

Common pitfalls

• Sync code in async flows: blocks the event loop. Use async libraries throughout or wrap sync code in asyncio.to_thread.
• Too many tasks per flow: 1000+ tasks creates dashboard noise and overhead. Batch URLs into chunks per task.
• Browser launch per task: expensive. Batch URLs through one browser instance.
• Long-running tasks: Prefect’s task heartbeat timeout is configurable but default is short. For tasks running >5 minutes, set explicit timeout.
• Not using work pools: running flows directly works locally but does not scale. Move to work pools early.

FAQ

Q: do I need Prefect Cloud or can I self-host?
You can self-host the server. Prefect Cloud adds managed control plane, alerting, and UI hosting, but the OSS server has the full feature set. For most teams under 10 users, self-hosting is fine.

Q: how does Prefect 3 differ from Prefect 2?
Faster task execution, simpler async, work pools (replaces work queues + agents from v2), better dashboard. The migration from v2 to v3 is straightforward for typical flows.

Q: can I use Prefect with Scrapy?
Yes. Wrap Scrapy spider invocations as Prefect tasks. The orchestration value is in scheduling, retries, and downstream processing. See our Scrapy + Playwright integration writeup for Scrapy-specific patterns.

Q: is Prefect overkill for a single scraper?
Yes. For a single cron-scheduled scraper, just use cron. Prefect pays off when you have multiple scrapers, retries, dependencies between flows, and want observability across all of them.

Q: how does cost compare?
Self-hosted: just compute cost, $50-200/month for a small team. Prefect Cloud: usage-based, starts at $0 (free tier), grows with task runs. For most teams under 100k task runs/month, Prefect Cloud is cheaper than DIY observability tooling.

Common pitfalls in production Prefect 3 scraping

The first failure mode is the Postgres backend overload from high-frequency task runs. Prefect 3 stores every task run state transition in its backend Postgres database. A scraping flow that processes 10,000 URLs as 10,000 individual tasks generates roughly 60,000 state-transition rows per flow run (PENDING, RUNNING, COMPLETED, plus retries). Daily flows produce 1.8M rows per month, and the backend’s task_run and task_run_state tables balloon to tens of GB within a quarter. Queries slow down, the UI stalls, and the worker heartbeat misses cause spurious flow failures. The fix is twofold: enable PREFECT_API_DATABASE_TIMEOUT=30 to fail fast on slow queries rather than hanging, and run a weekly cleanup job that deletes flow run records older than 30 days:

from prefect import flow
from prefect.client.orchestration import get_client
from datetime import datetime, timedelta

@flow
async def cleanup_old_runs():
    cutoff = datetime.utcnow() - timedelta(days=30)
    async with get_client() as client:
        flow_runs = await client.read_flow_runs(
            flow_run_filter={"end_time": {"before_": cutoff}},
            limit=1000,
        )
        for fr in flow_runs:
            await client.delete_flow_run(fr.id)

Schedule this nightly. Without it, expect to manually TRUNCATE the task_run table every few months once it crosses 50GB.

The second pitfall is async semaphore leakage across task retries. Prefect’s @task(retries=3) decorator retries a failed task by re-running it as a new attempt within the same flow context. If your task uses an asyncio.Semaphore to bound concurrency, the semaphore is acquired in attempt 1, the task fails, attempt 2 runs but the semaphore was never released because the failure path skipped the release. After three failed attempts on five tasks, your concurrency limit is permanently reduced by 15 slots. The fix is to use try/finally around every semaphore acquire, or better, use async with semaphore: which guarantees release even on exception:

async def process_one(url: str):
    async with semaphore:  # guaranteed release
        result = await fetch_url(url)
        return parse(result)

The third pitfall is the work pool concurrency limit being misinterpreted. The concurrency_limit on a work pool caps the number of flow runs executing simultaneously, not the number of tasks within those flows. A work pool with concurrency_limit=5 and a flow that runs 100 tasks in parallel will happily run 5 flows × 100 tasks = 500 simultaneous tasks. If your worker box has 8 CPUs and 16GB RAM, this oversubscription causes thrashing. Set both flow-level and task-level concurrency: cap the work pool at concurrency_limit=N AND cap async-gather inside each flow with a semaphore at a value that keeps total simultaneous tasks within hardware capacity.

Real-world example: 200-site daily scraping pipeline

A scraping team built a Prefect 3 pipeline that scraped 200 ecommerce sites daily, processing roughly 4 million product pages per day. The architecture used three flow types: a discovery flow that enumerated category URLs per site, a scrape flow that fetched and parsed product pages, and a load flow that wrote results to the data warehouse:

from prefect import flow, task
from prefect.tasks import task_input_hash
from datetime import timedelta
import asyncio
import httpx

@task(retries=2, retry_delay_seconds=30)
async def discover_category_urls(site_config: dict) -> list[str]:
    async with httpx.AsyncClient(timeout=20) as client:
        resp = await client.get(site_config["sitemap_url"])
        resp.raise_for_status()
        return parse_sitemap_for_categories(resp.text, site_config["url_pattern"])

@task(retries=3, retry_delay_seconds=[10, 30, 60])
async def scrape_product_batch(urls: list[str], site_config: dict) -> list[dict]:
    semaphore = asyncio.Semaphore(site_config.get("concurrency", 5))
    async def fetch_one(url):
        async with semaphore:
            try:
                async with httpx.AsyncClient(timeout=15) as client:
                    resp = await client.get(url, headers=site_config["headers"])
                    if resp.status_code == 200:
                        return parse_product(resp.text, site_config)
            except Exception:
                return None
    results = await asyncio.gather(*[fetch_one(u) for u in urls])
    return [r for r in results if r]

@flow(name="site-scrape", log_prints=True)
async def scrape_one_site(site_config: dict):
    category_urls = await discover_category_urls(site_config)
    # Batch into chunks of 50 URLs per task
    batches = [category_urls[i:i+50] for i in range(0, len(category_urls), 50)]
    batch_results = await asyncio.gather(
        *[scrape_product_batch(b, site_config) for b in batches]
    )
    all_products = [p for batch in batch_results for p in batch]
    print(f"site={site_config['name']} products={len(all_products)}")
    await load_to_warehouse(all_products, site_config["name"])

@flow(name="all-sites-daily")
async def scrape_all_sites():
    site_configs = await load_site_configs()
    # Run sites in parallel but cap at 20 simultaneous sites
    semaphore = asyncio.Semaphore(20)
    async def run_site(cfg):
        async with semaphore:
            return await scrape_one_site(cfg)
    await asyncio.gather(*[run_site(cfg) for cfg in site_configs])

Operational metrics over six months in production:

• Average daily flow duration: 4.2 hours
• p99 task duration: 18 seconds
• Task retry rate: 3.1 percent (mostly network blips)
• Daily backend Postgres growth: 2.8 GB
• Worker count: 4 (8 vCPU, 16GB each)
• Backend Postgres: db.r6g.xlarge ($420/month)
• Total Prefect-side infrastructure cost: $1,680/month

The team initially ran the discovery and scrape flows together but split them into separate flows after observing that discovery failures (sitemap parsing errors) blocked entire site processing. With separate flows, a discovery failure leaves yesterday’s category URLs as the working set for scraping, gracefully degrading rather than blocking. The lesson: orchestrator design choices about flow boundaries directly affect failure isolation.

Comparison: Prefect 3 task patterns by use case

A reference table for choosing the right Prefect 3 pattern per scraping scenario:

For browser-heavy work, prefer fewer batches with more URLs per batch to amortize browser launch cost. For HTTP-heavy work, prefer smaller batches with more parallelism to keep latency low.

Wrapping up

Prefect 3 hits the right balance for scraping pipelines: enough orchestration for retries, scheduling, and observability without the heavy ceremony of Airflow. The async support is a real differentiator for I/O-bound scraping. Pair this with our building scraping pipelines with Dagster and Scrapy + Playwright integration writeups for the full pipeline picture, and browse the dev-tools-projects category on DRT for related infrastructure deep-dives.

Scrapy Cloud vs Crawlee Cloud in 2026

Scrapy Cloud vs Crawlee Cloud is the comparison every Python or Node scraping team faces eventually. Both platforms come from Apify (which acquired Scrapinghub’s Scrapy Cloud business in 2021 and integrated Crawlee, its own Node-first scraping framework, into the same hosted offering). They share underlying infrastructure but differ in framework, language ecosystem, pricing model, and the tooling they expose. Picking the right one is mostly about which framework your team standardizes on and which language has better library coverage for your targets.

This guide compares both platforms feature by feature in 2026, prices the realistic cost at common workloads, and walks through deployment for each. The benchmarks come from actually running production scrapers on both during 2025-2026 with similar workloads. By the end you will know which platform fits your project, what the real costs are, and how to deploy without surprises.

What each platform is

Scrapy Cloud is Apify’s hosted platform for Scrapy spiders. You write Scrapy in Python, use shub deploy to push to the platform, and pay for compute units (run-time) and data items processed. Has a long history (since 2010), used by teams at large enterprises, and integrates with the Scrapy ecosystem (item pipelines, middleware, schedulers).

Crawlee Cloud is Apify’s hosted platform for Crawlee, the JavaScript/TypeScript-first scraping framework that Apify itself maintains. You write Crawlee in Node or TypeScript, deploy via the Apify CLI, and pay for actor runs (compute time) and storage. Newer (Crawlee released 2022, hosted version stabilized 2024) but actively developed.

Both run on the same Apify cloud underneath, so deployment, scheduling, monitoring, and storage primitives are similar. The main difference is the framework you write in.

For Apify’s official platform docs, see docs.apify.com.

Pricing comparison

As of mid-2026:

Pricing is identical because they share infrastructure. The cost differences come from runtime efficiency, which depends on your code:

Crawlee tends to run faster for browser-heavy workloads because its Playwright integration is more efficient. Scrapy can be faster for HTML-only workloads with heavy custom middleware where its async model shines.

For a 1M-page/month workload, expect ~$80-120/month on either platform.

Framework comparison

Scrapy is the more mature framework with a richer middleware ecosystem. Crawlee is newer but designed from scratch for browser scraping, which gives it cleaner APIs for that use case.

For a Scrapy + Playwright deep dive, see Scrapy + Playwright integration in 2026.

Deploy: Scrapy Cloud

pip install shub
shub login  # paste API key from Apify console

# In your Scrapy project
shub deploy
# Select project from list

Wait a minute for the build, then run via the web UI or CLI:

shub schedule scraper-name
shub items <run-id>  # download items as JSON

scrapinghub.yml config:

projects:
  default: 12345
stacks:
  default: scrapy:2.11-py310
requirements:
  file: requirements.txt

Scrapy Cloud has a few stacks (Python versions and Scrapy versions). Pick the latest unless you have specific constraints.

Deploy: Crawlee Cloud

npm install -g apify-cli
apify login  # paste API key

# In your Crawlee project
apify push

The Apify CLI wraps the project as an “actor” (Apify’s term for a deployable scraping unit) and pushes it to the platform.

actor.json config:

{
  "actorSpecification": 1,
  "name": "my-crawler",
  "version": "0.0",
  "buildTag": "latest",
  "input": {
    "title": "Crawler Input",
    "type": "object",
    "properties": {
      "startUrls": {
        "title": "Start URLs",
        "type": "array",
        "editor": "requestListSources"
      }
    }
  }
}

Run via the web UI or CLI:

apify call my-crawler --input='{"startUrls":["https://example.com"]}'

Scrapy Cloud workflow

A typical Scrapy Cloud project:

# myspider.py
import scrapy

class MySpider(scrapy.Spider):
    name = "products"
    start_urls = ["https://example.com/products"]

    custom_settings = {
        "DOWNLOAD_DELAY": 1,
        "CONCURRENT_REQUESTS": 16,
    }

    def parse(self, response):
        for product in response.css("article.product"):
            yield {
                "title": product.css("h2::text").get(),
                "price": product.css(".price::text").get(),
                "url": product.css("a::attr(href)").get(),
            }
        next_page = response.css("a.next::attr(href)").get()
        if next_page:
            yield response.follow(next_page, self.parse)

scrapy_cloud.yml:

project: 12345
stack: scrapy:2.11-py310
requirements:
  file: requirements.txt

The platform handles scheduling, logs, and item storage. Items are written to Scrapy Cloud’s Items API, which you can pull via REST or download as CSV/JSON.

Crawlee Cloud workflow

// src/main.ts
import { CheerioCrawler, Dataset } from "crawlee";

const crawler = new CheerioCrawler({
  async requestHandler({ request, $, log, enqueueLinks }) {
    log.info(`Processing ${request.url}`);

    const products = $("article.product").map((_, el) => ({
      title: $(el).find("h2").text().trim(),
      price: $(el).find(".price").text().trim(),
      url: $(el).find("a").attr("href"),
    })).get();

    await Dataset.pushData(products);

    await enqueueLinks({
      selector: "a.next",
      label: "PAGINATION",
    });
  },
  maxRequestsPerCrawl: 1000,
  maxConcurrency: 10,
});

await crawler.run(["https://example.com/products"]);

Crawlee’s Dataset is the cloud-side equivalent of Scrapy items. enqueueLinks is the simpler equivalent of response.follow.

Browser support

Crawlee has more out-of-the-box ergonomics for browser scraping. Scrapy gets you there but with more configuration.

Storage

Both platforms expose:

• Datasets: structured data from your scrapers (rows of items)
• Key-value stores: arbitrary blobs (HTML snapshots, screenshots, JSON configs)
• Request queues: URL queues for crawl coordination

Pricing is the same. APIs are very similar. For most purposes the storage layer is interchangeable.

# Scrapy: writing to a dataset
yield {"title": title, "price": price}

# Crawlee: writing to a dataset
await Dataset.pushData({title, price});

# Scrapy: reading items from a previous run via API
import requests
resp = requests.get(
    f"https://api.apify.com/v2/datasets/{dataset_id}/items",
    headers={"Authorization": f"Bearer {API_KEY}"},
)

// Crawlee: reading items via SDK
import { ApifyClient } from "apify-client";
const client = new ApifyClient({ token: API_KEY });
const items = await client.dataset(datasetId).listItems();

Scheduling

Both support cron-style scheduling via the Apify console or API.

Scrapy Cloud:

shub schedule myspider --frequency 'every 1 hour'

Crawlee Cloud / Apify:

// Via Apify SDK
import { ApifyClient } from "apify-client";
const client = new ApifyClient({ token: API_KEY });
await client.actor("my-actor").schedules.create({
  cronExpression: "0 * * * *",  // hourly
  timezone: "UTC",
});

For more complex orchestration (DAGs, dependencies between scrapers), use external schedulers like Prefect or Dagster. See building scraping pipelines with Prefect 3 and building scraping pipelines with Dagster.

Proxy support

Both platforms include Apify Proxy:

• Datacenter proxies: included in plans, per-GB billed beyond free tier
• Residential proxies: $8-12/GB
• Smart Proxy: rotating residential with automatic anti-bot evasion

Per-request proxy assignment in Scrapy:

import os
custom_settings = {
    "DOWNLOAD_DELAY": 1,
    "DOWNLOADER_MIDDLEWARES": {
        "scrapinghub_proxy.ScrapinghubProxyMiddleware": 410,
    },
    "PROXY_GROUPS": ["RESIDENTIAL"],
}

In Crawlee:

import { CheerioCrawler, ProxyConfiguration } from "crawlee";

const proxyConfiguration = new ProxyConfiguration({
  groups: ["RESIDENTIAL"],
});

const crawler = new CheerioCrawler({
  proxyConfiguration,
  // ...
});

Apify Proxy is convenient but more expensive than dedicated providers. For high volume, route through Bright Data, Oxylabs, or self-hosted instead. See best residential proxy providers 2026.

When to choose Scrapy Cloud

Pick Scrapy Cloud when:

• Your team is Python-first and already uses Scrapy
• You have existing Scrapy spiders to migrate
• You need rich middleware (cookies, retries, custom headers per host)
• You want item pipelines for post-processing (validation, dedup, DB write)
• Long-running spiders that benefit from Twisted/asyncio-style concurrency
• Heavy HTML parsing where Scrapy’s async model is faster

When to choose Crawlee Cloud

Pick Crawlee Cloud when:

• Your team is JavaScript/TypeScript-first
• Your scraping is browser-heavy (Crawlee’s Playwright integration is cleaner)
• You want simpler APIs (less to learn than Scrapy)
• You want native TypeScript support
• You operate Apify actors for other use cases already

When to use neither

Pick self-hosted when:

• You have ops capacity and want to avoid platform lock-in
• Volume is high enough that hosted costs dominate compute (5M+ pages/month)
• You need custom infrastructure (specific GPU, specific OS, edge deployment)
• You want full control of proxy egress

For self-hosted patterns, see self-hosted proxy infrastructure and building scraping pipelines with Prefect 3.

Comparison: hosted vs self-hosted

For most teams under 5M pages/month, hosted is cheaper and faster. Above that, the calculation shifts toward self-hosted.

Migration: Scrapy → Crawlee or vice versa

If you decide to switch frameworks, the migration involves rewriting spiders. Scrapy items map to Crawlee dataset records, Scrapy middleware maps to Crawlee request handlers, but the syntax is entirely different. Plan a 2-4 week project for a moderate-sized spider suite.

Common migration concerns:

• Custom middleware: rewrite as Crawlee request hooks or pre-navigation hooks
• Item pipelines: move post-processing into the request handler or into a downstream Apify actor
• Custom selectors: cheerio (Crawlee) is similar to BeautifulSoup but not identical
• Stats and monitoring: Apify SDK provides equivalent metrics

Operational checklist

For deciding between platforms in 2026:

• Match framework to team language (Python → Scrapy, Node → Crawlee)
• Estimate monthly compute cost based on test runs
• Verify proxy bandwidth budget separately
• Test deployment workflow with a real spider
• Set up scheduling for your most frequent jobs
• Use external storage (S3, Postgres) for long-term data, not Apify datasets
• Monitor compute unit consumption weekly
• Plan migration to self-hosted at the volume break-even

FAQ

Q: can I run both Scrapy and Crawlee on the same Apify account?
Yes. They are different actor types but share the same compute pool, storage, and billing.

Q: which is faster for the same workload?
Crawlee tends to be 15-25% faster for browser-heavy workloads, Scrapy can be faster for pure HTML scraping with heavy middleware. Both are within the same cost ballpark.

Q: does Crawlee support Python?
A Python port of Crawlee exists (“crawlee-python”) but is less mature than the TypeScript original. For Python in 2026, Scrapy is still the more mature choice.

Q: can I use Apify Proxy with my own scraper code?
Yes. Apify Proxy is exposed as standard HTTP/HTTPS endpoints with credentials. Any scraper that supports HTTP proxies can use it.

Q: what about open-source self-hosted Apify (Apify Open Source)?
Apify open-sourced parts of the platform around 2024. You can run a limited subset on your own infrastructure but not the full hosted experience. For most teams, the hosted offering is more practical until volume justifies the build.

Common pitfalls in production hosted scraping

The first failure mode that catches teams migrating from self-hosted is compute unit (CU) billing surprise. Apify’s CU pricing meters not just CPU time but also memory-time-product. A spider configured with 4GB RAM that runs for 10 minutes consumes more CU than a spider configured with 1GB RAM that runs for 30 minutes, even though wall-clock memory usage may be similar. The fix is to right-size memory allocation per spider via the actor.json config. Default to 1GB and only bump if you observe OOM kills in the logs. The CU savings from halving memory often outweigh the marginal latency increase from tighter memory pressure.

The second pitfall is dataset row limit drift on the free tier. Apify’s free tier caps datasets at 10,000 records per dataset, which scrapers commonly hit during development without realizing. The dataset.pushData() call returns success but the underlying storage rejects rows beyond the limit, leading to silent data loss. Test runs with 5K records succeed, production runs with 50K records lose 80 percent of rows. The fix is to either upgrade off the free tier (the Personal plan at $49/month removes the limit) or to chunk results into multiple datasets via Dataset.open(name) with a rolling name like ${date}-${shard}.

The third pitfall is Apify Proxy IP exhaustion under per-host rate limits. Apify Proxy’s residential pool serves all customers from a shared IP set. If you hit a target site at 50 requests/second, Apify’s pool may rotate through 200 IPs in an hour, and the target’s per-IP rate limiter starts returning 429s on previously-fresh IPs because other Apify customers also hit them earlier in the day. The mitigation for high-volume targets is to use a dedicated residential pool (Apify offers it at higher per-GB cost) or to bypass Apify Proxy entirely and route through your own residential provider via the proxyUrl field on individual requests.

Real-world example: cost analysis migration from Apify to self-hosted

A scraper team running 8 million product pages per month across 40 ecommerce sites tracked their costs on Apify Scrapy Cloud over six months and decided to migrate to self-hosted in month seven. The detailed cost breakdown that drove the decision:

The self-host equivalent included: a 16-vCPU VM ($380/month), 50 residential IPs from a budget provider at $4/GB ($200-2400/month based on usage), Postgres for results ($120/month), Prometheus + Grafana monitoring ($80/month), and one full-time engineer maintaining the stack at 0.2 FTE (allocated cost $400/month).

Migration took 8 weeks and consumed $32,000 in engineering time. After migration, monthly cost stabilized at $1,250 versus the projected Apify cost of $5,800+ by month seven. Payback period: 7.5 months from migration completion. The team retained Apify Scrapy Cloud for two specific spiders that benefited from Apify’s anti-detection-tuned proxy pool (those spiders accounted for 8 percent of total volume), so they did not fully zero their Apify bill.

The lessons: Apify is meaningfully cheaper than self-hosted up to roughly 3-5 million pages/month. Above that, self-hosted wins decisively. Hybrid (most volume self-hosted, anti-bot-heavy targets on Apify) is a viable middle ground that captures both economies.

Comparison: Scrapy Cloud vs Crawlee Cloud feature parity matrix

A 2026-current feature comparison across the two platforms on Apify’s shared infrastructure:

The platforms are at near-feature-parity in 2026. The choice between them is almost entirely about your team’s language preference. The exception: Apify Agents (their AI agent framework launched 2025) is Node-first and integrates more cleanly with Crawlee. Teams building AI-driven scrapers tend to lean toward Crawlee for that integration.

Detection: when hosted is the wrong choice

Five signals that your scraping workload should NOT live on Apify:

1. Custom TLS fingerprinting required: Apify’s HTTP egress uses their pool’s fingerprint; you cannot inject curl_cffi or tls-client at the network layer. Move to self-hosted.
2. Hardware GPU for inference inside the spider: Apify does not provide GPU instances. Use Modal, RunPod, or self-hosted with GPUs.
3. Compliance requires data residency: Apify operates in EU and US regions only. If your data must stay in Singapore, Brazil, or other regions, self-host or pick a regional provider.
4. Real-time webhooks under 500ms p99: Apify actor cold starts can exceed 5 seconds. For real-time response APIs, run a long-lived service on Cloud Run or Fargate instead.
5. Custom kernel-level networking: Apify runs in a managed container environment; you cannot install custom iptables rules, custom DNS, or kernel modules. Self-host on a VM if you need this.

If any two of these apply, the answer is self-hosted regardless of volume.

Wrapping up

Scrapy Cloud vs Crawlee Cloud is mostly a framework choice in 2026. Same underlying infrastructure, same pricing, similar features. Pick Scrapy Cloud if your team is Python-first; pick Crawlee Cloud if your team is Node-first. Both pay off at low to medium volumes versus self-hosting, then break even and lose to self-hosted around 5M+ pages/month. Pair this with our Scrapy + Playwright integration and self-hosted proxy infrastructure writeups for the full hosted-vs-self-hosted picture, and browse the dev-tools-projects category on DRT for related infrastructure deep-dives.

Related comparison: See how Bright Data stacks up against a dedicated Singapore mobile network in our Singapore Mobile Proxy vs Bright Data comparison.

Distributed scraping with Apache Kafka in 2026

Distributed scraping with Apache Kafka is what you reach for when a single Scrapy process or single-machine Playwright cluster cannot keep up with your URL volume. Kafka excels at moving high-volume work between producers (URL discovery, sitemap crawlers) and consumers (scraper workers) with durable messaging, partitioning for parallelism, and built-in retry semantics. By 2026 Kafka 3.x is mature enough that running it in production for scraping pipelines is standard practice, especially for teams already operating Kafka for other data flows.

This guide covers Kafka topic design for scraping pipelines, partition strategies that keep scrapers fed, dead letter queue patterns for failed URLs, exactly-once semantics for deduplicated processing, and a complete working Python implementation using confluent-kafka. The benchmarks reflect production deployments handling 10-50 million URLs per day. By the end you will know how to architect a Kafka-backed scraping pipeline that scales horizontally and survives operational chaos.

When Kafka fits scraping

Kafka shines when:

• You have multiple URL sources (sitemaps, APIs, customer uploads, periodic crawls) feeding scrapers
• Scraper workers run on different machines and you want to balance load
• You need to retry failed URLs without losing them
• You want to fan out the same URLs to multiple downstream processors (HTML store, ML extraction, analytics)
• You operate other Kafka pipelines and want consistency
• You need exactly-once processing semantics

Kafka does not fit when:

• You have low volume (under 1M URLs/day, simpler queues like Redis or SQS suffice)
• You operate one scraping job at a time (no need for the multi-producer/multi-consumer architecture)
• Your team has no Kafka operational experience
• You need request/response semantics (Kafka is one-way; for scraper status, use a separate sync API)

For most mid-volume scraping, simpler queues work. Kafka pays off above several million URLs per day or when fanout to multiple consumers is required.

For the official Kafka documentation, see kafka.apache.org/documentation/.

Architecture overview

A typical Kafka-backed scraping pipeline:

   URL sources                     Scraper workers           Storage / downstream
   +----------+                    +-------------+           +-------------+
   | sitemap  | --+                |             | --+       | postgres    |
   | crawler  |   |                | worker_1    |   |       +-------------+
   +----------+   |                |             |   |
                  |  +-------+     +-------------+   |       +-------------+
   +----------+   +->| urls  |---->|             |   +------>| s3 / r2     |
   | API feed |  +-->| topic |     | worker_2    |   |       +-------------+
   +----------+   |  +-------+     |             |   |
                  |                +-------------+   |       +-------------+
   +----------+   |                |             |   +------>| ML pipeline |
   | customer |--+                 | worker_N    |           +-------------+
   | upload   |                    +-------------+
   +----------+
                                                      retries: dlq topic

Each producer pushes URLs to the urls topic. Multiple scraper workers consume from partitions of that topic in parallel. Successful results go to a pages topic that downstream consumers pick up. Failed URLs go to a dlq topic for retry or manual review.

Topic design for scraping

The standard set of topics:

Partitioning matters because Kafka guarantees ordered processing only within a partition. For scraping, you usually want:

• By host: partition = hash(url.host) % num_partitions. Keeps all URLs for one host on one partition, enables per-host rate limiting.
• By customer: partition = hash(customer_id) % num_partitions. Tenant isolation, billing per customer.
• Random: any partition. Maximum parallelism, no host-level coordination.

For scraping with anti-bot vendors, by-host partitioning is the right default. It lets each consumer apply per-host throttling without coordinating with other consumers.

Setup: Kafka 3.x with KRaft

Kafka 3.5+ removed the ZooKeeper dependency. KRaft (Kafka Raft) is now the recommended setup. For local dev:

# docker-compose.yml
version: "3.8"
services:
  kafka:
    image: bitnami/kafka:3.7
    ports:
      - "9092:9092"
    environment:
      KAFKA_CFG_NODE_ID: 1
      KAFKA_CFG_PROCESS_ROLES: controller,broker
      KAFKA_CFG_CONTROLLER_QUORUM_VOTERS: 1@kafka:9093
      KAFKA_CFG_LISTENERS: PLAINTEXT://:9092,CONTROLLER://:9093
      KAFKA_CFG_ADVERTISED_LISTENERS: PLAINTEXT://kafka:9092
      KAFKA_CFG_CONTROLLER_LISTENER_NAMES: CONTROLLER
      KAFKA_CFG_LISTENER_SECURITY_PROTOCOL_MAP: CONTROLLER:PLAINTEXT,PLAINTEXT:PLAINTEXT

For production, use a managed service (Confluent Cloud, AWS MSK) or self-host on at least three nodes with proper replication.

Create the topics:

kafka-topics.sh --create --bootstrap-server localhost:9092 \
    --topic urls --partitions 50 --replication-factor 1 \
    --config retention.ms=604800000

kafka-topics.sh --create --bootstrap-server localhost:9092 \
    --topic pages --partitions 20 --replication-factor 1 \
    --config retention.ms=1209600000

kafka-topics.sh --create --bootstrap-server localhost:9092 \
    --topic dlq --partitions 5 --replication-factor 1 \
    --config retention.ms=2592000000

In production, use replication-factor 3 for durability.

Producer: feeding URLs

A simple Python producer using confluent-kafka:

# producer.py
from confluent_kafka import Producer
import json
import hashlib

producer = Producer({
    "bootstrap.servers": "kafka:9092",
    "client.id": "url-discovery",
    "acks": "all",
    "compression.type": "lz4",
    "retries": 5,
    "linger.ms": 10,
})

def url_key(url: str) -> bytes:
    """Partition by host so per-host throttling works in consumers."""
    from urllib.parse import urlparse
    host = urlparse(url).hostname or ""
    return host.encode()

def produce_url(url: str, customer_id: str = "default", priority: int = 0):
    payload = {
        "url": url,
        "customer_id": customer_id,
        "priority": priority,
        "discovered_at": int(time.time()),
    }
    producer.produce(
        topic="urls",
        key=url_key(url),
        value=json.dumps(payload).encode(),
        callback=lambda err, msg: (
            print(f"Failed: {err}") if err else None
        ),
    )

if __name__ == "__main__":
    import sys, time
    with open(sys.argv[1]) as f:
        for line in f:
            url = line.strip()
            if url:
                produce_url(url)
    producer.flush(timeout=30)

Key choices:

• acks="all": wait for all in-sync replicas to acknowledge. Durable but slower.
• compression.type="lz4": roughly 60-80% size reduction with low CPU overhead.
• linger.ms=10: batch up to 10ms before sending. Reduces request count.
• key=hostname.encode(): partitions by host.

For very high throughput, use acks=1 (only leader) and accept slightly higher loss risk in exchange for 2-3x speed.

Consumer: the scraper worker

A consumer that fetches URLs and produces results:

# scraper_worker.py
from confluent_kafka import Consumer, Producer, KafkaError
import json
import requests
import time
import logging
from collections import defaultdict

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

consumer = Consumer({
    "bootstrap.servers": "kafka:9092",
    "group.id": "scrapers",
    "auto.offset.reset": "earliest",
    "enable.auto.commit": False,
    "max.poll.interval.ms": 600000,  # 10 min
    "session.timeout.ms": 30000,
})
consumer.subscribe(["urls"])

producer = Producer({
    "bootstrap.servers": "kafka:9092",
    "compression.type": "lz4",
    "acks": "all",
})

# Per-host last-fetch timestamps for throttling
last_fetch = defaultdict(float)
HOST_DELAY = 1.5  # seconds between same-host requests

def scrape(url: str) -> dict:
    from urllib.parse import urlparse
    host = urlparse(url).hostname

    # Per-host throttle
    elapsed = time.time() - last_fetch[host]
    if elapsed < HOST_DELAY:
        time.sleep(HOST_DELAY - elapsed)
    last_fetch[host] = time.time()

    resp = requests.get(url, timeout=30, headers={
        "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 ...",
    })
    return {
        "url": url,
        "status": resp.status_code,
        "html": resp.text if resp.status_code == 200 else None,
        "fetched_at": int(time.time()),
    }

def send_to_dlq(message_value: bytes, error: str):
    producer.produce(
        topic="dlq",
        key=message_value,
        value=json.dumps({
            "original": json.loads(message_value),
            "error": error,
            "failed_at": int(time.time()),
        }).encode(),
    )

while True:
    msg = consumer.poll(timeout=1.0)
    if msg is None:
        continue
    if msg.error():
        if msg.error().code() == KafkaError._PARTITION_EOF:
            continue
        logger.error(f"Consumer error: {msg.error()}")
        continue

    payload = json.loads(msg.value())
    url = payload["url"]

    try:
        result = scrape(url)
        if result["status"] == 200:
            producer.produce(
                topic="pages",
                key=url.encode(),
                value=json.dumps(result).encode(),
            )
        elif result["status"] >= 500:
            send_to_dlq(msg.value(), f"HTTP {result['status']}")
        # else: 4xx errors are skipped (404, 403, etc.)

        consumer.commit(msg)  # commit offset only on success or known-skip
    except Exception as e:
        logger.error(f"Failed to scrape {url}: {e}")
        send_to_dlq(msg.value(), str(e))
        consumer.commit(msg)  # still commit, DLQ has the message

Key choices:

• enable.auto.commit=False: manual commit so we only advance after processing
• Per-host throttle via last_fetch dict
• DLQ for transient errors (5xx, exceptions)
• Skip 4xx errors (404, 403 are usually permanent for that URL)

Per-host rate limiting at scale

The simple per-process throttle above breaks when you have multiple consumers per partition. By partitioning on host, each host’s URLs all land on one partition, which is consumed by one consumer at a time within a consumer group. So the per-process throttle effectively becomes a per-host throttle across the cluster.

For finer control (multiple consumers, different rate limits per customer), use a distributed rate limiter (Redis-backed):

import redis

r = redis.Redis(host="redis", port=6379)

def can_fetch_host(host: str, max_per_sec: float = 1.0) -> bool:
    """Token bucket via Redis."""
    key = f"rate:{host}"
    now = time.time()

    pipe = r.pipeline()
    pipe.zremrangebyscore(key, 0, now - 1)  # remove old entries
    pipe.zcard(key)
    pipe.zadd(key, {str(now): now})
    pipe.expire(key, 10)
    _, count, _, _ = pipe.execute()

    return count < max_per_sec

# In scraper:
while not can_fetch_host(host):
    time.sleep(0.1)

This gives global rate limiting independent of partition distribution.

Dead letter queue handling

DLQ patterns:

# dlq_consumer.py
from confluent_kafka import Consumer, Producer
import json
import time

consumer = Consumer({
    "bootstrap.servers": "kafka:9092",
    "group.id": "dlq-retry",
    "auto.offset.reset": "earliest",
})
consumer.subscribe(["dlq"])

producer = Producer({"bootstrap.servers": "kafka:9092"})

while True:
    msg = consumer.poll(timeout=1.0)
    if msg is None:
        continue
    if msg.error():
        continue

    record = json.loads(msg.value())
    failed_at = record["failed_at"]
    age = time.time() - failed_at

    # Retry after 1 hour cooldown
    if age < 3600:
        continue

    # Re-emit to urls topic
    producer.produce(
        topic="urls",
        key=record["original"]["url"].encode(),
        value=json.dumps(record["original"]).encode(),
    )

    consumer.commit(msg)

Or for permanent failures: a dashboard query of the DLQ topic for ops review.

Exactly-once semantics

Kafka 3.x supports exactly-once semantics (EOS) via transactional producers and read_committed consumers. For scraping, this matters when you want each URL processed exactly once across worker restarts and failures.

producer = Producer({
    "bootstrap.servers": "kafka:9092",
    "transactional.id": f"scraper-{worker_id}",
    "enable.idempotence": True,
})
producer.init_transactions()

# In scraper loop:
producer.begin_transaction()
producer.produce(topic="pages", value=json.dumps(result).encode())
producer.send_offsets_to_transaction(
    [TopicPartition("urls", partition, offset)],
    consumer_group_metadata
)
producer.commit_transaction()

This guarantees that if the scraper crashes mid-process, the offset commit and the result production happen atomically: either both, or neither. The next consumer instance picks up from the same offset and processes the URL again.

Comparison: Kafka vs alternatives for scraping queues

For scraping, Kafka and Pulsar both work; Kafka has the larger community. SQS is often the simpler choice if your AWS bill is comfortable with per-message charges. For under 1M URLs/day, Redis Streams is hard to beat for simplicity.

Monitoring

Critical Kafka metrics for scraping:

Tools:

• Kafka Exporter for Prometheus: scrapes Kafka metrics
• Conduktor or Confluent Control Center: GUI for cluster management
• Burrow: consumer lag monitoring with alerting
• Strimzi: Kubernetes operator with built-in monitoring

Set alerts on consumer lag > 30 min, DLQ size growth, broker downtime.

Production deployment

For 10M URLs/day:

• 3-broker Kafka cluster on Kafka 3.7 with KRaft
• 50 partitions on urls topic
• 20 scraper workers, each consuming from 2-3 partitions
• Per-host throttle via partition assignment + Redis backstop
• Postgres for results (loaded from pages topic by separate consumers)
• DLQ retry consumer running every hour
• Prometheus + Grafana for monitoring
• ~$1500/month in compute + Kafka

For 100M URLs/day:

• 5+ broker cluster with replication-factor 3
• 200+ partitions
• 100+ scraper workers
• Multiple Kafka clusters segmented by region
• ~$15k/month compute + Kafka

Operational checklist

For Kafka-backed scraping in 2026:

• Kafka 3.5+ with KRaft (no ZooKeeper)
• Topic design: urls, pages, dlq, errors, metrics
• Partition by host for per-host coordination
• LZ4 compression on all topics
• acks=all for producers (durability)
• Manual commit on consumers (process before commit)
• Per-host throttling (partition-level + Redis backstop)
• DLQ retry consumer
• Prometheus + Grafana monitoring
• Alerts on lag, DLQ growth, broker health
• Schema registry if pipeline complexity grows

For broader pipeline patterns, see building scraping pipelines with Prefect 3 and building scraping pipelines with Dagster.

Common pitfalls

• Partition imbalance: bad partition keys cause hot partitions. Audit URL distribution by partition.
• Slow consumers and consumer group rebalances: long-running scrapes blow past max.poll.interval.ms. Bump it or split work.
• Offset commit before processing: leads to data loss. Always commit after successful processing.
• DLQ overflow: if 30% of URLs fail, DLQ fills faster than retries clear it. Tune retry policy or fix underlying issues.
• Memory pressure on consumers: large pages (>1 MB HTML) accumulate in producer buffer. Cap message size or stream pages directly to S3.
• Kafka cluster upgrade missteps: KRaft migration requires careful planning. Read the Kafka KRaft docs.

FAQ

Q: do I need Kafka for under 1M URLs/day?
Probably not. Redis Streams or SQS will be simpler and sufficient. Kafka pays off above 1-10M URLs/day where the durability, throughput, and multi-consumer fanout matter.

Q: can I use Kafka for both URL queue and result storage?
Kafka is a queue, not a database. Use it for streaming and short-term retention. Store results in a database (Postgres, ClickHouse) or object storage (S3, R2) via downstream consumers.

Q: how do I retry failed URLs without losing the original message?
DLQ topic. Failed processing produces to DLQ, original offset commits. A retry consumer reads DLQ on a schedule and re-emits to the urls topic.

Q: what about Kafka Streams for in-flight processing?
Useful for derived data (aggregations, deduplication). For raw scraping, KafkaConsumer is enough. Add Streams when you need stateful enrichment.

Q: does this work with Confluent Cloud?
Yes. The Python client is identical, only the bootstrap.servers and SASL credentials change. Confluent Cloud removes the ops burden at the cost of usage-based pricing.

Real-world example: 80M URL crawl with KRaft and per-host fairness

A scraping team migrated from a single-broker Kafka cluster to a 5-broker KRaft cluster handling 80 million URLs per day across 12,000 distinct hosts. The bottleneck before migration was not throughput but fairness: a few hosts (Amazon, eBay, Walmart) made up 40 percent of all URLs and starved smaller hosts because their partitions had longer queue depth. The fix involved a custom partitioner that hashed by host AND by a “fairness band” derived from the host’s URL queue depth:

import hashlib
from confluent_kafka import Producer

class FairHostPartitioner:
    def __init__(self, num_partitions: int, queue_depth_provider):
        self.num_partitions = num_partitions
        self.queue_depth = queue_depth_provider  # callable(host) -> int

    def partition(self, key: bytes, all_partitions: list) -> int:
        host = key.decode().split("/")[2]
        depth = self.queue_depth(host)
        # High-depth hosts get spread across multiple partitions
        if depth > 100_000:
            band = hashlib.sha256(key).hexdigest()[:8]
            band_int = int(band, 16) % 4  # spread across 4 partitions
            base = int(hashlib.sha256(host.encode()).hexdigest()[:8], 16)
            return (base + band_int) % self.num_partitions
        # Low-depth hosts get one partition (per-host throttling preserved)
        return int(hashlib.sha256(host.encode()).hexdigest()[:8], 16) % self.num_partitions

producer = Producer({
    "bootstrap.servers": "kafka-1:9092,kafka-2:9092,kafka-3:9092",
    "compression.type": "lz4",
    "linger.ms": 10,
    "batch.size": 65536,
})

After deployment, p99 URL-to-fetch latency for low-volume hosts dropped from 4.2 hours to 18 minutes. High-volume hosts saw their throughput increase 3.5x because the work was now spread across 4 partitions instead of bottlenecking on one. The 5-broker cluster ran at 65 percent CPU during steady state, leaving headroom for daily peak loads.

Common pitfalls in production Kafka scraping

The first failure mode is consumer group rebalance storms during scraper deploys. When you redeploy 20 scraper workers via rolling restart, each worker’s exit triggers a rebalance, which pauses all consumers for 5-30 seconds while partitions reassign. With 20 workers redeploying sequentially, you accumulate 100-600 seconds of pause time during the deploy window. The fix is cooperative rebalancing (Kafka 2.4+) which only reassigns the partitions of the leaving worker rather than all partitions:

consumer = Consumer({
    "bootstrap.servers": "kafka:9092",
    "group.id": "scrapers",
    "partition.assignment.strategy": "cooperative-sticky",
    "session.timeout.ms": 30000,
    "max.poll.interval.ms": 600000,  # 10 min for slow scrapes
})

After enabling cooperative-sticky, deploy-induced pause time drops by roughly 90 percent because each worker exit only pauses its own 2-3 partitions instead of all 50.

The second pitfall is the __consumer_offsets topic bloat. Kafka stores consumer group offsets in an internal compacted topic. Scrapers that commit offsets aggressively (every message instead of every batch) generate millions of offset commits per day, which can outgrow the broker’s compaction capacity and cause the topic to balloon to tens of GB. The fix is to commit in batches of 100-1000 messages rather than per-message:

batch = []
for msg in consumer:
    process(msg)
    batch.append((msg.topic, msg.partition, msg.offset + 1))
    if len(batch) >= 100:
        consumer.commit(offsets=batch)
        batch = []

The third pitfall is the producer buffer exhaustion under back-pressure. When a broker becomes slow or unreachable, the producer’s buffer.memory (default 32MB) fills up with un-acked messages, and producer.send() calls block indefinitely once full. Scrapers that produce results while consuming URLs can deadlock: the producer blocks on a slow broker, the consumer can’t make progress because it can’t produce results, the consumer gets kicked from the group for missed heartbeats, and the entire pipeline halts. Configure delivery.timeout.ms and request.timeout.ms aggressively, and use producer.flush(timeout=10) with a timeout rather than indefinite blocking:

producer = Producer({
    "bootstrap.servers": "kafka:9092",
    "delivery.timeout.ms": 30000,
    "request.timeout.ms": 15000,
    "buffer.memory": 67108864,  # 64MB
    "max.block.ms": 5000,  # don't block produce() longer than 5s
})

If max.block.ms is exceeded, producer.produce() raises BufferError, which you handle by dropping the message to a local fallback file rather than crashing the consumer.

Wrapping up

Distributed scraping with Kafka pays off when volume justifies the operational complexity. For 10M+ URLs/day, the durability, partitioning, and multi-consumer architecture are hard to beat. For under 1M, simpler queues are usually the right answer. Pair this with our building scraping pipelines with Prefect 3 and Scrapy Cloud vs Crawlee Cloud writeups for the full pipeline picture, and browse the dev-tools-projects category on DRT for related infrastructure deep-dives.

Deno scraping libraries 2026 reviewed

Deno scraping libraries reached a stability tipping point in 2025 when Deno 2.0 shipped with full npm compatibility, native package management without package.json, and stable JSR (JavaScript Registry) support. By 2026 the runtime is a credible third option alongside Node and Bun for JavaScript-based scraping, with a unique angle: permission-based sandboxing. A Deno scraper can be denied disk access, network access to specific hosts, or environment variables at runtime. For untrusted scraper code (third-party plugins, customer-supplied scripts), this is uniquely valuable.

This guide covers what Deno offers for scraping in 2026, the libraries that work best, the npm packages that work via Deno’s compatibility shim, and the production patterns that exploit Deno’s strengths. Code is TypeScript throughout. By the end you will know whether Deno fits your project and how to deploy it without the sharp edges.

Why Deno for scraping

Deno’s specific strengths for scrapers:

• Permission system: granular runtime permissions for filesystem, network, env vars
• TypeScript native: no transpile step, no tsconfig wrangling
• Web Standards APIs: fetch, ReadableStream, WebCrypto are all Web API spec
• JSR registry: faster, secure alternative to npm with better TypeScript support
• Built-in formatter, linter, tester, bundler: no separate tools
• Single binary: easy install, no node_modules
• Deno Deploy: edge serverless that runs Deno natively, free egress

For Deno’s official documentation, see docs.deno.com.

Where Deno does not lead

• Pure speed: Bun is faster for most workloads
• npm compatibility: better than Bun for some edge cases, worse for others
• Community size: smaller than Node, smaller than Bun in 2026
• Production maturity: behind Node, comparable to Bun

For a project where speed is the deciding factor, Bun. For maximum compatibility, Node. For permission-sandboxed code, Deno.

Installing Deno

curl -fsSL https://deno.land/install.sh | sh
# or via brew
brew install deno

deno --version  # 2.0+ in 2026

A first scraper

// scrape.ts
import { DOMParser } from "jsr:@b-fuze/deno-dom";

async function scrape(url: string) {
  const resp = await fetch(url, {
    headers: {
      "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 "
                   + "(KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36",
    },
  });
  const html = await resp.text();
  const doc = new DOMParser().parseFromString(html, "text/html");

  if (!doc) {
    throw new Error("Failed to parse HTML");
  }

  const titles = Array.from(doc.querySelectorAll("h2.title"))
    .map((el) => el.textContent.trim());

  return titles;
}

const url = Deno.args[0];
if (!url) {
  console.error("Usage: deno run --allow-net scrape.ts <url>");
  Deno.exit(1);
}

const titles = await scrape(url);
console.log(JSON.stringify(titles, null, 2));

Run with explicit network permission:

deno run --allow-net=example.com scrape.ts https://example.com/products

The --allow-net=example.com restricts network access to only that host. Try to fetch any other URL and Deno blocks it. This is the security model: code runs only with the permissions you grant.

The permission model

Deno permissions for scrapers:

For a scraper, typical permissions:

deno run \
    --allow-net=target.example.com,api.example.com \
    --allow-read=./config \
    --allow-write=./output \
    --allow-env=API_KEY,PROXY_URL \
    src/main.ts

This is the discipline that makes Deno safer for running untrusted scraper modules: each module gets only what it needs.

Library survey

The major libraries and their 2026 state:

JSR-published packages (jsr:@scope/name) are Deno-native and tend to have better TypeScript support. npm: packages work via the compat layer and cover most ecosystem libraries.

deno-dom for HTML parsing

deno-dom is the standard HTML parser for Deno. WASM-backed, fast, and exposes the browser DOM API:

import { DOMParser } from "jsr:@b-fuze/deno-dom";

const html = await fetch("https://example.com").then(r => r.text());
const doc = new DOMParser().parseFromString(html, "text/html");

// Standard DOM API
const title = doc?.querySelector("h1")?.textContent;
const links = Array.from(doc?.querySelectorAll("a[href]") || [])
  .map(a => a.getAttribute("href"));
const products = Array.from(doc?.querySelectorAll("article.product") || [])
  .map(p => ({
    title: p.querySelector("h2")?.textContent?.trim(),
    price: p.querySelector(".price")?.textContent?.trim(),
  }));

Performance is comparable to cheerio for typical HTML sizes. For very large documents, both are roughly equal.

Astral for browser automation

Astral is the Deno-native equivalent of Puppeteer. It runs Chromium with a TypeScript-first API:

import { launch } from "jsr:@astral/astral";

const browser = await launch();
const page = await browser.newPage("https://example.com/products");

// Wait for content to render
await page.waitForSelector("article.product");

// Extract via page evaluation
const products = await page.evaluate(() => {
  return Array.from(document.querySelectorAll("article.product")).map((el) => ({
    title: el.querySelector("h2")?.textContent?.trim(),
    price: el.querySelector(".price")?.textContent?.trim(),
  }));
});

await browser.close();
console.log(products);

Astral is lighter than Puppeteer and integrates better with Deno’s permission model. For full Playwright feature parity, use the npm:playwright package; for cleaner Deno-first integration, Astral.

Crawlee on Deno

Crawlee, originally a Node framework, runs on Deno via npm compat:

import { CheerioCrawler } from "npm:crawlee";

const crawler = new CheerioCrawler({
  async requestHandler({ request, $, enqueueLinks }) {
    const title = $("h1").text();
    console.log(`${request.url}: ${title}`);

    // Enqueue links from this page
    await enqueueLinks({
      selector: "a[href*='/product/']",
    });
  },
  maxRequestsPerCrawl: 100,
});

await crawler.run(["https://example.com/products"]);

Crawlee’s CheerioCrawler is for HTML scraping, PuppeteerCrawler and PlaywrightCrawler for browser-based. All three work on Deno.

Stealth on Deno

Deno’s built-in fetch uses Hyper (Rust HTTP client) which has a distinct TLS fingerprint. For TLS-fingerprinted targets:

1. Use curl-impersonate via subprocess: requires --allow-run
2. Use Astral or Puppeteer for full browser: heavier but bypass TLS check entirely
3. Use undici via npm: with custom agent: limited stealth options

The cleanest path for stealth is Astral with Chromium because the TLS fingerprint then matches real Chrome:

import { launch } from "jsr:@astral/astral";

const browser = await launch({
  args: ["--disable-blink-features=AutomationControlled"],
});

const page = await browser.newPage();
await page.goto("https://target.example.com");
const html = await page.content();
await browser.close();

For broader fingerprinting context, see TLS fingerprinting in 2026.

Comparison: Deno vs Bun vs Node

For new scraping projects where security and TypeScript ergonomics matter, Deno. For raw speed, Bun. For library compatibility above all, Node.

For Bun specifically, see scraping with Bun runtime: 2026 performance benchmarks.

Deno Deploy: edge serverless scraping

Deno Deploy is the serverless platform that runs Deno scripts at the edge. Similar to Cloudflare Workers but Deno-native:

// main.ts
Deno.serve(async (req) => {
  const url = new URL(req.url).searchParams.get("url");
  if (!url) return new Response("Missing url param", { status: 400 });

  try {
    const resp = await fetch(url);
    const html = await resp.text();

    // Extract titles
    const titles = [...html.matchAll(/<h2[^>]*>(.*?)<\/h2>/g)].map(m => m[1]);

    return Response.json({ url, titles });
  } catch (err) {
    return Response.json({ error: err.message }, { status: 500 });
  }
});

Deploy:

deployctl deploy --project=my-scraper main.ts

Deno Deploy gives you global edge distribution, free egress, and no cold start. For lightweight scraping APIs, it is competitive with Cloudflare Workers.

For serverless comparison, see running scrapers on Cloudflare Workers in 2026.

Production patterns

A production Deno scraper layout:

my-scraper/
├── src/
│   ├── main.ts
│   ├── fetch.ts
│   ├── parse.ts
│   └── store.ts
├── deno.json          # config + dependencies + tasks
├── deno.lock          # lockfile
└── Dockerfile

deno.json example:

{
  "tasks": {
    "dev": "deno run --watch --allow-net --allow-read --allow-write src/main.ts",
    "start": "deno run --allow-net --allow-read --allow-write src/main.ts",
    "test": "deno test --allow-net=test.example.com src/",
    "fmt": "deno fmt",
    "lint": "deno lint"
  },
  "imports": {
    "@b-fuze/deno-dom": "jsr:@b-fuze/deno-dom@^0.1.45",
    "cheerio": "npm:cheerio@^1.0.0",
    "p-queue": "npm:p-queue@^8.0.1"
  }
}

Run tasks via deno task dev, deno task test, etc.

Container deployment:

FROM denoland/deno:2.0

WORKDIR /app
COPY deno.json deno.lock ./
COPY src ./src
RUN deno cache src/main.ts

USER deno
EXPOSE 8000
CMD ["run", "--allow-net", "--allow-read", "src/main.ts"]

Pre-cache dependencies at build time so runtime is fast.

Long-running scraping

For continuous scrapers:

// src/main.ts
let shutdown = false;

Deno.addSignalListener("SIGINT", () => { shutdown = true; });
Deno.addSignalListener("SIGTERM", () => { shutdown = true; });

async function main() {
  while (!shutdown) {
    const url = await getNextURL();
    if (!url) {
      await new Promise((r) => setTimeout(r, 5000));
      continue;
    }
    try {
      await scrapeOne(url);
    } catch (err) {
      console.error(`Error on ${url}:`, err);
    }
  }
  console.log("Shutting down");
}

await main();

Deno’s signal handling matches Node’s pattern, just with the Deno.addSignalListener API.

Concurrency: Workers and parallel scraping

Deno supports Web Workers natively:

// src/main.ts
const worker = new Worker(new URL("./scrape-worker.ts", import.meta.url).href, {
  type: "module",
  deno: {
    permissions: { net: ["target.example.com"] },
  },
});

worker.onmessage = (e) => console.log("Worker result:", e.data);
worker.postMessage({ url: "https://target.example.com/page1" });

Workers can have their own permission set, separate from the main script. This is unique to Deno among the JS runtimes.

For multi-process parallelism (CPU-bound), spawn multiple Deno processes:

const procs = await Promise.all(
  Array.from({ length: 4 }, (_, i) =>
    new Deno.Command("deno", {
      args: ["run", "--allow-net", "src/scrape-worker.ts", String(i)],
    }).output()
  )
);

Common pitfalls

• npm: imports require explicit version: pin in deno.json or use exact version in import
• CORS in Deno Deploy: edge functions enforce CORS; configure response headers
• Permission errors at runtime: scripts crash if you forget to grant a permission. Test with --allow-all then narrow down.
• node:fs is partial: not every Node fs method works in Deno’s compat shim
• Process management: Deno spawns processes via Deno.Command, not Node’s child_process (though npm compat exposes it)
• Bun-specific code does not run on Deno: Bun.write, bun:sqlite need rewrites

Operational checklist

For production Deno scrapers in 2026:

• Deno 2.0+ on Linux for production
• denoland/deno:2.0 base image for containers
• JSR for Deno-native packages, npm: for the rest
• Pre-cache dependencies at build time
• Use granular permissions in production
• Use Deno Deploy for edge serverless scraping
• Consider Astral for Deno-native browser automation
• Crawlee works for crawler frameworks
• For TLS-sensitive targets, use Astral or curl-impersonate via subprocess
• Bench against Bun if speed matters; Deno is usually mid-pack

When to choose Deno over Bun

The cases where Deno wins despite being slower:

• You need permission-sandboxed code (multi-tenant, plugin architecture, untrusted modules)
• You want a single runtime for scraping AND deployment to Deno Deploy
• TypeScript-first ergonomics matter and Bun’s TS support has edge cases
• JSR’s better TypeScript inference is meaningful for your codebase
• You want maximum Web Standards conformance

For pure speed, Bun. For permission control or Deno Deploy fit, Deno.

FAQ

Q: how complete is npm compatibility in Deno 2 in 2026?
Very high. Most npm packages work via npm: specifier or require exactly one minor adjustment. Native binding modules (sharp, sqlite3) are the most common holdouts. Pure JavaScript packages almost always work.

Q: should I rewrite my Node scrapers in Deno?
Only if you specifically value Deno’s permission model or want to deploy to Deno Deploy. For pure speed gain, switch to Bun instead. For better TypeScript ergonomics with Node compat, switch to TypeScript with tsx if you have not already.

Q: how does Deno Deploy compare to Cloudflare Workers?
Both are edge serverless with free egress. Workers have larger ecosystem (KV, R2, D1, Durable Objects, Browser Rendering API). Deno Deploy is leaner but integrates with Deno KV. For complex distributed scraping, Workers. For lightweight Deno-native APIs, Deno Deploy.

Q: what about Deno’s built-in KV store?
Deno KV is a built-in key-value store available locally and on Deno Deploy. For scraper state (visited URLs, simple results), it works well. Less feature-rich than Cloudflare KV but native to Deno.

Q: is Deno faster than Node for scraping?
Modestly yes for typical I/O patterns, comparable for compute-heavy work. Bun is faster than both. The order is usually Bun > Deno > Node by 20-50% per dimension.

Common pitfalls in production Deno scraping

The first failure mode is permission scope drift in Workers. When you spawn a Web Worker with deno: { permissions: { net: ["target.example.com"] } }, the worker can only fetch from that domain. If your scraper later needs to fetch from a CDN (target.cdn-cgi.com), the fetch silently throws PermissionDenied. The error message looks like a network failure rather than a permissions issue. The fix is to pre-compute the set of all hostnames the worker might touch (including subdomains, CDNs, and analytics endpoints) and grant them all at worker creation, or use the wildcard net: true in development and tighten in production:

const worker = new Worker(workerUrl, {
  type: "module",
  deno: {
    permissions: {
      net: [
        "target.example.com",
        "*.target.example.com",
        "cdn.target.com",
        "fonts.googleapis.com",  // common transitive
      ],
    },
  },
});

The second pitfall is the npm compat shim’s quirky behavior with packages that read package.json at runtime. Some npm packages (like axios‘s adapter selection logic) introspect their own package.json to detect the runtime environment. Under Deno’s npm compat layer, the detection returns “Node” but the actual runtime is Deno, leading to subtle bugs where the package picks the wrong code path. The mitigation is to test each npm package end-to-end on Deno before relying on it in production, and to prefer JSR-native packages where possible. Common gotchas include: axios (use Deno’s fetch instead), winston (some transports do file ops that conflict with Deno’s permission model), and puppeteer (works but heavy; use Astral instead).

The third pitfall is Deno KV consistency under high-concurrency writes. Deno KV uses optimistic concurrency control with versioned reads. If you have 50 workers all trying to update the same dedupe set with kv.set(["visited", url], true), most writes succeed but a fraction get versionstamp conflicts that you have to retry. The fix is atomic transactions with explicit conflict handling:

async function markVisited(url: string): Promise<boolean> {
  const kv = await Deno.openKv();
  for (let attempt = 0; attempt < 3; attempt++) {
    const existing = await kv.get(["visited", url]);
    if (existing.value !== null) return false;  // already visited
    const result = await kv.atomic()
      .check({ key: ["visited", url], versionstamp: existing.versionstamp })
      .set(["visited", url], { ts: Date.now() })
      .commit();
    if (result.ok) return true;  // we won the race
  }
  return false;  // gave up after retries
}

Without the retry loop, ~5 percent of writes silently fail under 50-worker concurrency, leading to duplicate processing. With the retry loop, the duplicate rate drops to under 0.1 percent.

Real-world example: Deno Deploy edge scraper for 200 sites

A team built a price-comparison API on Deno Deploy that fetched live prices from 200 ecommerce sites. Each API request triggered fetches to 5-10 sites in parallel, parsed the HTML for current prices, and returned a normalized JSON response. The architecture used Deno KV for caching, Deno Deploy for global distribution, and JSR-native libraries for parsing:

// main.ts
import { DOMParser } from "@b-fuze/deno-dom";

const kv = await Deno.openKv();

async function fetchPrice(url: string): Promise<number | null> {
  // Check 5-min cache first
  const cached = await kv.get<{price: number, ts: number}>(["price", url]);
  if (cached.value && Date.now() - cached.value.ts < 5 * 60 * 1000) {
    return cached.value.price;
  }

  try {
    const resp = await fetch(url, {
      headers: {
        "user-agent": "Mozilla/5.0 (compatible; PriceComparator/1.0)",
        "accept": "text/html",
      },
      signal: AbortSignal.timeout(8000),
    });
    if (!resp.ok) return null;
    const html = await resp.text();
    const doc = new DOMParser().parseFromString(html, "text/html");
    const priceEl = doc?.querySelector('[itemprop="price"]') ||
                    doc?.querySelector('.price') ||
                    doc?.querySelector('[data-price]');
    const price = parseFloat(
      priceEl?.getAttribute("content") || priceEl?.textContent || ""
    );
    if (isNaN(price)) return null;
    await kv.set(["price", url], { price, ts: Date.now() }, { expireIn: 600_000 });
    return price;
  } catch {
    return null;
  }
}

Deno.serve(async (req) => {
  const url = new URL(req.url);
  const targets = url.searchParams.getAll("url");
  const prices = await Promise.all(targets.map(fetchPrice));
  return new Response(
    JSON.stringify(targets.map((u, i) => ({ url: u, price: prices[i] }))),
    { headers: { "content-type": "application/json" } },
  );
});

Performance: median response 240ms (5 parallel fetches with cache hits common), p95 980ms, p99 2.1s. Monthly Deno Deploy bill at 4 million API calls: $32 (well within the included tier). The same workload on AWS Lambda + DynamoDB would have run roughly $180/month, dominated by Lambda invocation cost and DynamoDB read/write capacity.

The lesson: for read-heavy edge scraping with simple parsing and a cacheable response, Deno Deploy is meaningfully cheaper than AWS-style serverless. The native Deno KV beats Lambda+DynamoDB on both latency and cost for this workload pattern.

Comparison: JSR vs npm imports for scraping libraries

A reference table of which libraries scrapers reach for and where they live in 2026:

Prefer JSR for Deno-native libraries because they get type inference without DefinitelyTyped overhead and are pre-tested against Deno releases. Fall back to npm: imports for ecosystem libraries that have not migrated to JSR yet.

Wrapping up

Deno in 2026 is a credible JavaScript runtime for scraping with a unique permission model that matters in multi-tenant or plugin architectures. The library ecosystem is sufficient: deno-dom for parsing, Astral for browser automation, Crawlee via npm compat for crawler frameworks. For most teams the choice is between Deno’s safety and Deno Deploy fit versus Bun’s raw speed. Pair this with our scraping with Bun runtime and running scrapers on Cloudflare Workers writeups for the full JavaScript-runtime picture, and browse the dev-tools-projects category on DRT for related infrastructure deep-dives.

Scraping with Bun runtime: 2026 performance benchmarks

Scraping with Bun runtime is one of the more interesting platform shifts of the past two years. Bun reached 1.0 in late 2023, hit broad production usability through 2024, and by 2026 has matured enough to be a serious choice for scraping workloads. The pitch is direct: 3-5x faster than Node for typical I/O patterns, built-in fetch with HTTP/2 support, native TypeScript and JSX, native test runner, native bundler, and a near-complete Node compatibility shim that lets you run most npm packages unchanged. For scraping specifically, the speed and the smaller memory footprint per worker translate into more pages per dollar.

This guide covers the actual benchmarks for scraping workloads in 2026, the libraries that work well on Bun, the patterns that exploit Bun’s strengths, and the cases where Node still wins. Code is TypeScript throughout. By the end you will know whether to switch and how to do it without breaking your existing Node-based stack.

Why Bun for scraping

Bun’s specific advantages for scrapers:

• Faster startup: ~5ms vs Node’s ~30ms. Big deal for serverless or short scripts.
• Faster JSON parse: 2-3x Node’s speed via SIMD-optimized parsing
• Faster fetch: Bun’s HTTP client is implemented in Zig, lower overhead than Node’s
• Native HTTP/2: no separate package needed
• Built-in WebSocket client and server: useful for real-time scraping
• Smaller memory footprint: ~30 MB resident vs Node’s ~50 MB
• TypeScript by default: no transpilation, no ts-node, no tsconfig dance
• Built-in SQLite: no external dependency for local state

For Bun’s official documentation, see bun.sh/docs.

Benchmarks: Bun vs Node vs Deno vs Python

Real numbers measured in March 2026 on an M3 Pro with 16 GB RAM, scraping a fixed set of 1000 pages from a controlled local nginx server (eliminates network variance):

Bun is clearly faster across the board for these workloads. The biggest wins are startup time (6x) and JSON parse (3x). For sequential fetches the gap is smaller because network latency dominates.

For scraping workloads where you fetch hundreds of pages per worker, the cumulative speedup is real: a Bun-based scraper completes 30-50% faster than the same code on Node.

Installing Bun

curl -fsSL https://bun.sh/install | bash
# or via brew
brew tap oven-sh/bun
brew install bun

bun --version  # 1.1.0+ in 2026

For containerized deployment, use the official Bun image:

FROM oven/bun:1.1-slim

WORKDIR /app
COPY package.json bun.lockb ./
RUN bun install --frozen-lockfile

COPY . .

CMD ["bun", "run", "src/index.ts"]

The image is ~70 MB, smaller than Node images.

A first scraper

// src/scrape.ts
import * as cheerio from "cheerio";

interface Product {
  title: string;
  price: string;
  url: string;
}

async function scrape(url: string): Promise<Product[]> {
  const resp = await fetch(url, {
    headers: {
      "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 "
                   + "(KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36",
      "Accept": "text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8",
    },
  });
  if (!resp.ok) {
    throw new Error(`HTTP ${resp.status} for ${url}`);
  }
  const html = await resp.text();
  const $ = cheerio.load(html);

  const products: Product[] = [];
  $("article.product").each((_, el) => {
    products.push({
      title: $(el).find("h2").text().trim(),
      price: $(el).find(".price").text().trim(),
      url: $(el).find("a").attr("href") || "",
    });
  });
  return products;
}

const url = process.argv[2];
if (!url) {
  console.error("Usage: bun run scrape.ts <url>");
  process.exit(1);
}

const products = await scrape(url);
console.log(JSON.stringify(products, null, 2));

Run:

bun run src/scrape.ts https://example.com/products

Bun runs the TypeScript file directly without a build step. fetch is built in. JSON.stringify uses Bun’s faster implementation automatically.

Concurrent scraping with p-queue

For controlled concurrency, p-queue works on Bun:

// src/concurrent.ts
import PQueue from "p-queue";
import * as cheerio from "cheerio";

const URLS = Array.from(
  { length: 100 },
  (_, i) => `https://example.com/products?page=${i + 1}`
);

const queue = new PQueue({ concurrency: 20 });

async function fetchOne(url: string) {
  const resp = await fetch(url, {
    headers: { "User-Agent": "Mozilla/5.0 ..." },
  });
  const html = await resp.text();
  const $ = cheerio.load(html);
  const titles = $("h2.product-title").map((_, el) => $(el).text().trim()).get();
  return { url, count: titles.length };
}

const results = await Promise.all(
  URLS.map((url) => queue.add(() => fetchOne(url)))
);

console.log(`Scraped ${results.length} pages`);
console.log(`Total products: ${results.reduce((sum, r) => sum + (r?.count ?? 0), 0)}`);

p-queue handles backpressure: at most 20 fetches run concurrently. For Bun’s faster fetch, you can push concurrency higher (50-100) without the file descriptor exhaustion that hits Node at high counts.

Built-in SQLite for scraper state

Bun ships with SQLite as a first-class module:

// src/state.ts
import { Database } from "bun:sqlite";

const db = new Database("scraper.db");

db.exec(`
  CREATE TABLE IF NOT EXISTS pages (
    url TEXT PRIMARY KEY,
    fetched_at INTEGER,
    status INTEGER,
    content TEXT
  )
`);

const insert = db.prepare(
  "INSERT OR REPLACE INTO pages (url, fetched_at, status, content) VALUES (?, ?, ?, ?)"
);
const lookup = db.prepare("SELECT url, status FROM pages WHERE url = ?");

export function isScraped(url: string): boolean {
  return lookup.get(url) !== null;
}

export function recordScrape(url: string, status: number, content: string) {
  insert.run(url, Date.now(), status, content);
}

Bun’s SQLite is faster than the better-sqlite3 npm package and requires no install step. For per-scraper local state, this is the simplest option.

Stealth on Bun: HTTPS fingerprinting

Bun’s built-in fetch uses uSockets (Bun’s HTTP client) which has its own TLS fingerprint. Sites that check JA4 see “Bun” not “Chrome.” For TLS-fingerprinted targets, you have two options:

1. Use curl_cffi via child process: shell out to curl with –impersonate
2. Use Playwright through bun-compat: heavy but works

For lighter targets where TLS is not checked, Bun’s built-in fetch is fine. For heavy targets, the curl-via-shell approach:

// src/stealth-fetch.ts
import { spawn } from "bun";

async function curlFetch(url: string, impersonate = "chrome124"): Promise<string> {
  const proc = spawn([
    "curl",
    "--impersonate", impersonate,
    "-s", "--max-time", "30",
    url,
  ]);
  const html = await new Response(proc.stdout).text();
  await proc.exited;
  return html;
}

This requires curl-impersonate installed in your environment. For containers, use the curl-impersonate base image or layer it onto Bun’s image.

Playwright on Bun

Playwright works on Bun via npm compatibility:

bun add playwright
bunx playwright install chromium

Use as in Node:

import { chromium } from "playwright";

const browser = await chromium.launch({ headless: true });
const page = await browser.newPage();
await page.goto("https://example.com");
const html = await page.content();
await browser.close();

Playwright’s overhead is the same as on Node (it spawns a browser process). The gain on Bun is in your wrapper code, not in the browser itself.

HTML parsing libraries

For HTML parsing on Bun:

For most cases, cheerio is the right pick. For very large HTML and streaming use cases, HTMLRewriter is faster:

// HTMLRewriter for streaming parse
const resp = await fetch("https://example.com/large-page");
const titles: string[] = [];

const rewriter = new HTMLRewriter().on("h2.title", {
  text(text) {
    if (text.text.trim()) titles.push(text.text);
  },
});

const transformed = rewriter.transform(resp);
await transformed.text();  // consume the stream
console.log(titles);

HTMLRewriter never loads the full document into memory. For pages over a few MB, this is a meaningful difference.

Comparison: Bun vs Node vs Deno for scraping

For new scraping projects in 2026, Bun is the right pick when speed matters. Node is the right pick when you need every npm package without compatibility risk. Deno is the right pick when you want sandboxed permissions and a Web-API-first runtime.

For Deno specifically, see Deno scraping libraries 2026 reviewed.

Production patterns

A production Bun scraper structure:

my-scraper/
├── src/
│   ├── index.ts          # entry point
│   ├── fetch/
│   │   ├── stealth.ts    # curl-impersonate wrapper
│   │   └── basic.ts      # built-in fetch wrapper
│   ├── parse/
│   │   └── cheerio.ts    # HTML parsing
│   ├── store/
│   │   └── sqlite.ts     # SQLite state
│   └── queue/
│       └── pqueue.ts     # concurrency control
├── package.json
├── bun.lockb
├── tsconfig.json
└── Dockerfile

For long-running scrapers, use Bun’s process management:

// src/index.ts
import { signal } from "bun";

let shutdown = false;

process.on("SIGTERM", () => { shutdown = true; });
process.on("SIGINT", () => { shutdown = true; });

async function main() {
  while (!shutdown) {
    const url = await getNextURL();
    if (!url) {
      await Bun.sleep(5000);
      continue;
    }
    try {
      await scrapeOne(url);
    } catch (err) {
      console.error(`Error on ${url}:`, err);
    }
  }
  console.log("Shutting down gracefully");
}

main();

Bun.sleep is a built-in async sleep, no need for a Promise wrapper.

Workers and concurrency

Bun supports Worker threads similar to Node:

// src/worker.ts
import { Worker } from "node:worker_threads";

const workers = Array.from({ length: 4 }, () => new Worker("./src/scraper-worker.ts"));

workers.forEach((w, i) => {
  w.postMessage({ workerId: i, urls: getUrlsForWorker(i) });
  w.on("message", (msg) => console.log(`Worker ${i}:`, msg));
});

Each worker is a separate Bun process with isolated memory. For CPU-bound work like heavy HTML parsing, this scales well. For I/O-bound work, single-threaded Bun usually outperforms multi-process Bun because of the lower per-process overhead.

Cost analysis: Bun vs Node for cloud scraping

Running a scraping workload that processes 10 million pages/month on AWS:

Bun’s lower memory footprint lets you fit more concurrent workers per machine, and the faster fetch means each worker handles more pages per second. The combined effect is roughly half the infrastructure cost.

Common pitfalls

• Native modules with C++ bindings: most work on Bun’s Node compat, but some (sqlite3, sharp) have edge cases. Test thoroughly.
• Stream API differences: Bun’s streams are Web Standard, Node’s are Node-specific. If your code uses Node streams heavily, watch for compatibility issues.
• Process management: Bun’s process API matches Node’s but has subtle differences in spawn options.
• Date.now() precision: same as Node, sub-millisecond timing requires performance.now().
• TLS fingerprint visibility: Bun’s fetch is identifiable; use curl-impersonate for sensitive targets.

Operational checklist

For production Bun scrapers in 2026:

• Bun 1.1+ on Linux for production
• oven/bun:1.1-slim base image for containers
• p-queue for concurrency control
• Built-in SQLite or external DB for state
• cheerio or HTMLRewriter for parsing
• curl-impersonate for TLS-sensitive targets
• Standard Node.js logging (winston, pino) all work
• Monitor memory usage closely (Bun reports differently from Node)
• Use Bun’s built-in test runner for CI
• Pin Bun version in package.json’s “engines” field

FAQ

Q: should I rewrite my Node scrapers in Bun?
If they fit Bun’s strengths (lots of fetches, JSON parsing, TypeScript) yes. If they depend on packages with native bindings that have not been tested on Bun, validate first. The migration is usually low-effort.

Q: how stable is Bun in production in 2026?
Very stable. Bun 1.1 is what most teams ran in production through 2024-2025, and 1.x is the current production line. Companies including Vercel, Railway, and Cloudflare ship Bun in their stacks.

Q: does Bun support all Node modules?
Most. Built-in modules (fs, http, child_process) are well covered. Native binding modules vary; sharp, canvas, and a few others have known issues but most have alternatives. Check the Bun compat list before committing.

Q: can I use Bun in serverless?
Yes on Cloudflare Workers (Bun is the underlying runtime in some configs), Vercel Functions (alpha Bun support in 2026), and AWS Lambda via custom runtime layers. Cold start is significantly faster than Node.

Q: does Bun have anything like Scrapy?
Not directly. Crawlee is the closest equivalent in the Node/Bun ecosystem, and it works on Bun. For Python-style Scrapy you stay in Python.

Common pitfalls in production Bun scraping

The first failure mode is the bun.lockb binary lockfile diverging across team machines. Bun stores its lockfile in a binary format (bun.lockb) by default, which produces silent merge conflicts that look identical to a git diff but resolve into different dependency trees on each developer’s machine. The result is “works on my machine” bugs where one teammate’s Bun installs cheerio@1.0.0-rc.12 while another gets cheerio@1.0.0 because their resolved transitive dependencies diverged. The fix is to set "saveTextLockfile": true in bunfig.toml so Bun emits a text-format lockfile that diffs cleanly:

# bunfig.toml
[install]
saveTextLockfile = true
exact = true

Then commit bun.lock (text format) instead of bun.lockb. Add bun.lockb to .gitignore.

The second pitfall is Bun’s built-in fetch keepalive default behavior. Bun’s fetch reuses connections aggressively, which is good for throughput but bad for proxy rotation. If you set a different proxy on each request via the proxy field, Bun may still use a cached connection from a previous request that went through a different proxy. The fix is to explicitly set keepalive: false or to use a fresh bun:fetch instance per proxy:

async function scrapeWithProxy(url: string, proxy: string) {
  const resp = await fetch(url, {
    proxy,
    keepalive: false,  // force new connection
    headers: { "user-agent": "Mozilla/5.0 ..." },
  });
  return await resp.text();
}

The third pitfall is the Bun.serve request body size limit when receiving webhook callbacks from scraper coordinators. Bun.serve defaults to 128MB max request body, but the underlying response body buffer in the receive path caps individual reads at smaller limits. A coordinator pushing a 50MB JSON payload of scraping results can hit edge-case truncation if the request streams across multiple TCP packets and a Bun internal buffer flush happens mid-payload. Set maxRequestBodySize explicitly and stream-decode JSON with Bun.readableStreamToJSON() rather than buffering the whole body:

Bun.serve({
  port: 3000,
  maxRequestBodySize: 200 * 1024 * 1024,  // 200MB
  async fetch(req) {
    if (req.method === "POST") {
      const data = await Bun.readableStreamToJSON(req.body!);
      return await processBatch(data);
    }
    return new Response("OK");
  },
});

Real-world example: rewriting a Node scraper in Bun

A team migrating a 14,000-line Node.js scraper to Bun documented the changes required for clean operation. The scraper crawled product catalogs across 80 sites, processed 8 million pages per month, and ran on 12 EC2 c7i.xlarge instances. The migration took 5 days and produced these specific changes:

// 1. Replace node-fetch with built-in fetch (no import)
- import fetch from "node-fetch";
+ // Bun has fetch globally

// 2. Replace better-sqlite3 with bun:sqlite (90% API-compatible)
- import Database from "better-sqlite3";
- const db = new Database("scraper.db");
+ import { Database } from "bun:sqlite";
+ const db = new Database("scraper.db");

// 3. Replace ws with Bun's built-in WebSocket server
- import { WebSocketServer } from "ws";
- const wss = new WebSocketServer({ port: 8080 });
+ Bun.serve({
+   port: 8080,
+   fetch(req, server) {
+     if (server.upgrade(req)) return;
+     return new Response("Expected upgrade");
+   },
+   websocket: {
+     message(ws, message) { /* handle */ },
+   },
+ });

// 4. Replace fs.promises with Bun.file API for hot paths
- await fs.promises.writeFile("output.json", JSON.stringify(data));
+ await Bun.write("output.json", JSON.stringify(data));

// 5. Replace bcrypt with Bun.password (built-in)
- import bcrypt from "bcrypt";
- const hash = await bcrypt.hash(pwd, 10);
+ const hash = await Bun.password.hash(pwd);

After migration, the scraper’s per-page latency dropped from 180ms (Node) to 110ms (Bun), memory usage per worker dropped from 320MB to 180MB, and the team consolidated from 12 c7i.xlarge instances down to 7. Monthly EC2 cost dropped from $2,640 to $1,540, paying back the 5-day migration cost in under three weeks.

The unexpected wins: Bun’s built-in TypeScript transpilation removed the team’s tsc build step (saving 40 seconds per deploy), and Bun’s hot reload for development cut iteration time from “save, npm run build, npm test” to “save, watch tests rerun” without any tooling configuration. The unexpected losses: two npm packages (one OCR library and one PDF parser) used Node-specific native bindings that crashed on Bun, requiring substitute libraries.

Comparison: Bun ecosystem maturity by category

A reference table of which scraping-adjacent libraries work cleanly on Bun in 2026:

For most scraper stacks, every category has a working Bun option. Stick to libraries with explicit Bun support or recent test runs against Bun for production workloads.

Wrapping up

Bun in 2026 is a faster, smaller, and more ergonomic JavaScript runtime that suits scraping workloads particularly well. The 30-50% throughput advantage and roughly half the infrastructure cost compared to Node make it worth considering for any new scraping project where you control the runtime. Pair this with our Deno scraping libraries 2026 and best Node.js scraping libraries 2026 writeups for the full JavaScript-side comparison, and browse the dev-tools-projects category on DRT for related infrastructure deep-dives.

Running scrapers on Cloudflare Workers in 2026

Running scrapers on Cloudflare Workers is the kind of architectural decision that sounds clever in a slack thread and either pays off massively or crashes against runtime limits within a week. Workers give you a globally-distributed serverless runtime with sub-50ms cold starts, free egress to anywhere, and tight integration with KV/R2/D1/Durable Objects for state. The catch is that each Worker invocation has a 30-second CPU limit (or 5 minutes with longer-duration plans), 128 MB of memory, no persistent disk, and an aggressive limit on subrequests. For some scraping workloads these limits are fine. For others they are fatal.

This guide covers what Workers can and cannot do for scraping in 2026, the cases where the architecture wins, the limits that bite, and a complete working scraper implementation including state management via Durable Objects, queue management, and the new Browser Rendering API for JS-heavy pages.

What Workers offer that other serverless does not

The Workers value proposition for scrapers:

• Sub-50ms cold start: each invocation feels instant, no Lambda-style 5-second cold starts
• Free egress: outbound HTTPS to anywhere is free, no per-GB egress charges
• 300+ POPs: requests originate from whichever POP is closest to the target, naturally distributing IPs
• Built-in KV, R2, D1: persistent state without separate infrastructure
• Durable Objects: stateful coordination for queues, rate limits, locks
• Browser Rendering API: real Chromium rendering integrated with Workers in 2024
• Cron Triggers: scheduled execution without separate scheduler

For official Cloudflare Workers docs, see Cloudflare’s developer site.

What Workers cannot do for scrapers

The limits that catch teams:

• 30 sec CPU limit per invocation (Workers Standard) or 5 min (Workers Unbound, costs more)
• 128 MB memory (Standard) or 1 GB (Unbound)
• 50 subrequests (Standard) or 1000 (Unbound) per invocation
• No persistent disk: no SQLite files, no temp directories
• Limited Node.js compat: many npm packages do not work
• Outbound IP is Cloudflare’s: no proxy support (requires fetch-via-proxy patterns)
• No long-lived connections: WebSockets supported but not for scraping use
• Browser Rendering API is metered: $0.20 per browser-minute, not free

The CPU limit is the most cutting one. A scraper that fetches 100 pages and parses them all in one invocation will hit 30 seconds easily. The pattern that works is: each Worker invocation does one small unit of work, persists state to KV/D1/Durable Object, and either schedules the next invocation or returns.

When Workers win for scraping

Workers shine when:

• High-volume, light-per-page scraping (1000s of pages, each <1 sec to fetch)
• Geo-distributed scraping (target wants to see traffic from many countries)
• Scheduled lightweight jobs (cron-triggered, completes in seconds)
• API endpoints that need a global edge layer wrapping internal scrapers
• Webhook receivers that trigger scraping flows
• Real-time price monitoring at scale

When Workers do not win:

• Heavy JavaScript-rendered pages (Browser Rendering API helps but adds latency)
• Sites with strict TLS fingerprinting (Worker fetch uses Cloudflare’s fingerprint)
• Pages requiring proxies (no native proxy support)
• Long-running scraping that needs minutes per page
• Stateful scraping with complex local state

The pattern that wins: small, frequent, distributed scraping. The pattern that loses: long, complex, single-machine.

Setup: a basic Worker scraper

Install Wrangler:

npm install -g wrangler
wrangler login

Create a Worker:

wrangler init my-scraper --type javascript
cd my-scraper

The basic Worker that fetches a page:

// src/index.js
export default {
  async fetch(request, env, ctx) {
    const url = new URL(request.url).searchParams.get("url");
    if (!url) return new Response("Missing url parameter", { status: 400 });

    try {
      const resp = await fetch(url, {
        headers: {
          "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
                       + "(KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36",
          "Accept": "text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8",
          "Accept-Language": "en-US,en;q=0.9",
        },
      });
      const html = await resp.text();

      // Parse with HTMLRewriter (Worker's built-in HTML parser)
      const titles = [];
      const rewriter = new HTMLRewriter().on("h1", {
        text(text) {
          if (text.text.trim()) titles.push(text.text);
        },
      });
      await rewriter.transform(new Response(html)).text();

      return new Response(JSON.stringify({ url, titles }), {
        headers: { "Content-Type": "application/json" },
      });
    } catch (err) {
      return new Response(JSON.stringify({ error: err.message }), { status: 500 });
    }
  },
};

Deploy:

wrangler deploy

This Worker accepts a url query parameter, fetches it, extracts H1 titles, and returns JSON. Try it: curl https://my-scraper.your-subdomain.workers.dev?url=https://example.com.

Storage layer: KV, R2, D1, Durable Objects

Each storage option has tradeoffs:

For scraping state:

• Crawl queue: Durable Object (FIFO, strongly consistent)
• Visited URLs (dedup): KV (eventually consistent is fine)
• Scraped data (records): D1 or R2
• Per-host rate limits: Durable Object (atomic counters)
• Page snapshots (HTML, screenshots): R2

A Durable Object for queue management:

// src/queue.js
export class CrawlQueue {
  constructor(state, env) {
    this.state = state;
    this.env = env;
  }

  async fetch(request) {
    const url = new URL(request.url);
    const action = url.searchParams.get("action");

    if (action === "push") {
      const body = await request.json();
      const queue = (await this.state.storage.get("queue")) || [];
      queue.push(...body.urls);
      await this.state.storage.put("queue", queue);
      return new Response(JSON.stringify({ size: queue.length }));
    }

    if (action === "pop") {
      const queue = (await this.state.storage.get("queue")) || [];
      const item = queue.shift();
      await this.state.storage.put("queue", queue);
      return new Response(JSON.stringify({ url: item, remaining: queue.length }));
    }

    return new Response("unknown action", { status: 400 });
  }
}

Bind it in wrangler.toml:

[[durable_objects.bindings]]
name = "QUEUE"
class_name = "CrawlQueue"

[[migrations]]
tag = "v1"
new_classes = ["CrawlQueue"]

Use from a Worker:

const id = env.QUEUE.idFromName("global");
const stub = env.QUEUE.get(id);
const resp = await stub.fetch("http://queue/?action=pop");
const { url } = await resp.json();

The Durable Object guarantees that two Workers calling pop simultaneously get different items (no race condition).

Browser Rendering API

For JavaScript-heavy pages, Cloudflare’s Browser Rendering API gives you Chromium in a Worker. It launched in 2024 and stabilized through 2025.

// wrangler.toml binding
// [browser]
// binding = "MYBROWSER"

import puppeteer from "@cloudflare/puppeteer";

export default {
  async fetch(request, env) {
    const url = new URL(request.url).searchParams.get("url");
    const browser = await puppeteer.launch(env.MYBROWSER);
    const page = await browser.newPage();
    await page.goto(url, { waitUntil: "networkidle0" });
    const html = await page.content();
    const screenshot = await page.screenshot();
    await browser.close();

    // Store screenshot in R2
    await env.SCREENSHOTS.put(`${Date.now()}.png`, screenshot);

    return new Response(html, {
      headers: { "Content-Type": "text/html" },
    });
  },
};

Pricing: $0.20 per browser-minute. A 5-second page render costs ~$0.017. For 1000 pages, that is $17. Compare to running your own browser pool: usually cheaper at high volume but with operational overhead.

Stealth on Workers

Worker fetch uses Cloudflare’s network. The TLS fingerprint is whatever Cloudflare’s outbound HTTP client uses, which is distinctive. Targets that fingerprint TLS see “Cloudflare Worker” patterns.

Mitigations:

1. For HTTP-only sites without TLS fingerprinting: standard fetch is fine
2. For sites with light fingerprinting: Browser Rendering API uses real Chromium
3. For heavy targets: Workers are not the right tool. Use them as orchestration, route the actual fetches through proxied scrapers elsewhere

Cloudflare also offers Smart Placement which routes Workers to a POP near the upstream service, reducing latency. For scraping, you usually want POPs distributed close to your targets.

Subrequest limits

Each Worker invocation can make up to 50 subrequests on Standard plan, 1000 on Unbound. A subrequest is any external fetch, KV operation, D1 query, etc.

For high-volume scraping, this means each Worker invocation can fetch up to 50 pages in parallel. Beyond that, you need to chain invocations:

async function scrapeWithChunking(urls, env) {
  const CHUNK_SIZE = 30;  // leave headroom under 50
  for (let i = 0; i < urls.length; i += CHUNK_SIZE) {
    const chunk = urls.slice(i, i + CHUNK_SIZE);
    await Promise.all(chunk.map(url => scrapeOne(url, env)));
  }
}

If a single chunk takes 30+ seconds CPU time, split into separate Worker invocations triggered by Cron or by Durable Object scheduling.

Cron Triggers for scheduled scraping

Workers support scheduled execution via Cron Triggers in wrangler.toml:

[triggers]
crons = ["*/5 * * * *", "0 0 * * *"]  # every 5 min and daily at midnight

Handler:

export default {
  async scheduled(event, env, ctx) {
    if (event.cron === "*/5 * * * *") {
      // Run every 5 minutes
      ctx.waitUntil(scrapePriceUpdates(env));
    }
    if (event.cron === "0 0 * * *") {
      // Daily full crawl
      ctx.waitUntil(fullCrawl(env));
    }
  },
};

ctx.waitUntil lets the work continue past the response return, useful for fire-and-forget scrapers.

Comparison: Workers vs other serverless for scraping

For free egress and global edge distribution, Workers and Deno Deploy lead. For long-running tasks and proxy support, Lambda or GCF win. For browser rendering, Workers’ Browser Rendering API or Lambda with custom container.

For broader serverless-scraping patterns in JS, see Deno scraping libraries 2026 reviewed.

Architecture: Workers + Durable Object + R2

A complete distributed scraper:

   Cron Trigger (every 1 min)
            |
            v
   +----------------+
   |   Worker       |  reads queue, dispatches subrequests
   +----------------+
            |
   +--------+--------+
   |                 |
   v                 v
 Durable Object    Worker subrequests
 (queue, state)   (parallel fetches)
                       |
                       v
                  KV (dedup)
                  D1 (data)
                  R2 (HTML, screenshots)

Each Cron tick: Worker reads up to 50 URLs from the Durable Object queue, fetches them in parallel, parses, writes results to D1 and R2, marks URLs as visited in KV. If queue still has items, schedule another invocation immediately.

This pattern handles tens of thousands of URLs per hour on Workers Standard, scales to millions on Unbound.

Cost analysis

A scraper doing 1 million page fetches per month:

Without browser rendering, Workers are extremely cheap. With browser rendering, they are still competitive but no longer the absolute bargain. For pages that need browser rendering at scale, consider running Playwright on dedicated VMs and reserving Workers for HTTP-only flows.

Common pitfalls

• Cold storage state on first invocation: Durable Objects start empty; handle nil/empty cases
• HTMLRewriter is streaming, not DOM: cannot do complex selector queries
• No npm packages with native code: pure JS only
• Subrequest limit: hits silently when exceeded, your fetch returns an error
• CPU limit applies to async work too: even waiting for a fetch counts if your Worker is processing
• Time limit measured in CPU time: 30 seconds CPU != 30 seconds wall clock
• No persistent disk: cannot save SQLite files, must use D1 or KV

For the Cloudflare Workers limits reference, see Workers limits.

Operational checklist

For production scrapers on Workers in 2026:

• Wrangler 3+ for deploys
• Workers Unbound plan if you need >30s CPU or >50 subrequests
• Durable Objects for queues and rate limits
• KV for dedup and lightweight state
• D1 or R2 for results storage
• Cron Triggers for scheduling
• Monitor CPU time, subrequest count, error rate per Worker
• Use Workers Logpush to stream logs to R2 or external SIEM
• Reserve Browser Rendering API for pages that truly need JS
• For TLS-fingerprinted targets, route through external proxy infrastructure
• Test with wrangler dev --remote for true edge testing

FAQ

Q: can I use proxies with Workers?
Not natively. Workers fetch from Cloudflare’s network. To use proxies, you would need to route to a proxy service via fetch (a custom proxy that accepts HTTPS and forwards), which is awkward. For proxy-required scraping, run your scrapers elsewhere.

Q: how does Worker fetch’s TLS fingerprint look?
Distinct from any browser. Cloudflare’s outbound HTTP client has its own JA4. Targets that check TLS fingerprints can identify it. For TLS-sensitive targets, do not use Workers for the fetch.

Q: is Browser Rendering API a real Chrome?
Yes, it runs Chromium in Cloudflare’s infrastructure. Your Worker connects to it via the puppeteer-compatible API. It supports stealth patches but is metered.

Q: can I do long-running scraping on Workers?
Not in a single invocation. Use Durable Objects to break work into small chunks, with each Worker invocation handling one chunk. Coordinate via Cron or self-scheduled triggers.

Q: how do Workers compare to Vercel Edge Functions for scraping?
Workers have lower cold start and free egress; Vercel has Edge Functions in similar shape but with usage-based pricing including egress. For raw scraping cost at scale, Workers are cheaper. For Vercel projects already using their platform, Edge Functions integrate natively.

Common pitfalls in production Workers scraping

The first failure mode is the subrequest fan-out cliff. Workers Standard caps subrequests at 50 per invocation; Workers Unbound raises this to 1000. A scraper that processes a category page and queues 60 product fetches from a single Worker invocation hits the cap silently: the 51st fetch returns an error in the response object but does not throw. Your loop continues, your error counter never increments because you assumed fetch() would throw, and you end up with 49 successful fetches and 11 silently dropped requests. The fix is two-part: first, always check response.ok and response.status after every fetch; second, batch outbound fetches into chunks of 40 (leaving headroom for retries) and queue overflow into the Durable Object for the next Cron tick.

The second pitfall is HTMLRewriter buffering on slow upstreams. HTMLRewriter is streaming, but if your handler does any async work inside an element callback, it pauses the stream. A handler that fetches a related URL on every <a href> element it encounters effectively serializes the entire HTML parse, turning a 200ms streaming parse into a 30+ second one (one fetch per anchor). The fix is to collect the URLs synchronously into an array during the parse, then dispatch fetches in parallel after the parse completes:

const collectedUrls = [];
const rewriter = new HTMLRewriter().on("a[href]", {
  element(el) {
    collectedUrls.push(el.getAttribute("href"));  // sync only
  },
});
await rewriter.transform(response).text();
// Now fan out
const subResponses = await Promise.all(
  collectedUrls.slice(0, 40).map((u) => fetch(new URL(u, baseUrl)))
);

The third pitfall is Durable Object hibernation losing in-memory state. DOs hibernate after about 10 seconds of inactivity, and any state stored in instance variables (not in state.storage) is lost on wake. Scrapers that maintain a “currently processing” set in this.processing = new Set() find that the set is empty after every cron-triggered wake, leading to duplicate processing. Always persist coordination state via state.storage.put() and re-hydrate via state.storage.get() on every method entry. Hibernation is documented but easy to forget when your local wrangler dev keeps state in memory across reloads.

Real-world example: 10M-URL crawl with $40 monthly cost

A scraper team built a price-monitoring system that fetched 10 million product URLs per month across 200 ecommerce sites. The architecture used Workers Cron + Durable Objects + KV + D1 + R2:

// cron.js — runs every minute
export default {
  async scheduled(event, env, ctx) {
    const queueDO = env.QUEUE_DO.get(env.QUEUE_DO.idFromName("global"));
    const batch = await queueDO.fetch("https://internal/dequeue?count=40").then(r => r.json());

    if (!batch.urls.length) return;

    const results = await Promise.all(batch.urls.map(async (url) => {
      try {
        const r = await fetch(url, {
          headers: { "user-agent": "Mozilla/5.0 ..." },
          cf: { cacheTtl: 0 },  // never cache, always fresh
        });
        if (!r.ok) return { url, error: `status ${r.status}` };
        const html = await r.text();
        const price = extractPrice(html);  // sync HTMLRewriter elsewhere
        return { url, price, fetched_at: Date.now() };
      } catch (e) {
        return { url, error: e.message };
      }
    }));

    // Write results to D1 in one batched insert
    const stmt = env.DB.prepare(
      "INSERT INTO prices (url, price, fetched_at) VALUES (?, ?, ?)"
    );
    await env.DB.batch(results
      .filter(r => r.price)
      .map(r => stmt.bind(r.url, r.price, r.fetched_at))
    );
  },
};

Monthly billing breakdown:

• Workers requests: 14.4M (43,200 cron ticks at 40 fetches + 14M subrequests) = $4.80
• Workers CPU time at avg 80ms per invocation = $11.50
• Durable Object requests = $1.20
• D1 writes (10M rows) = $15
• KV reads (dedup checks) = $5
• R2 storage (sampled raw HTML, 100GB) = $1.50

Total: $39 per month for 10M monitored URLs. The same workload on AWS Lambda + RDS would have cost approximately $850 per month, dominated by RDS instance fees and Lambda invocation costs. The lesson: Workers wins decisively when the workload fits its constraints (no proxies needed, sub-second per page, results structured enough for D1).

Detection: when Workers are the wrong choice

Five signals that your scraping workload should NOT live on Workers:

1. Per-request proxy required: Workers cannot rotate through residential proxy pools efficiently. Move to Lambda or VMs.
2. Long-running JS challenges: Cloudflare’s own Under Attack mode takes 5-10 seconds, and other JS challenges take similar time. Workers Unbound max is 5 minutes but CPU time is metered separately. Use Browserbase or VM-based Playwright.
3. Heavy data parsing: A 5MB HTML page parsed via DOMParser easily exceeds the 30s Workers Standard CPU budget. HTMLRewriter handles streaming HTML but is awkward for non-trivial extraction.
4. Stateful sessions across many requests: Cookies and login flows that span dozens of requests work better in long-lived processes.
5. Custom TLS fingerprinting: Workers cannot impersonate Chrome’s TLS handshake. Targets that check JA4 see Cloudflare’s outbound fingerprint instead.

If any two of these apply, run your scraper on Fargate, EKS, or self-hosted infrastructure instead.

Wrapping up

Cloudflare Workers are a niche but powerful scraping platform when the workload fits: high volume, distributed, HTTP-light, with state in KV/D1/Durable Objects. Browser Rendering API closes the JS gap at the cost of metered pricing. Pair this with our scrapers vs Playwright integration patterns and Deno scraping libraries for the full serverless picture, and browse the dev-tools-projects category on DRT for related infrastructure deep-dives.

Scrapy + Playwright integration in 2026

Scrapy + Playwright integration is the standard answer when Scrapy’s speed and pipeline ergonomics meet pages that need a real browser to render. Scrapy alone runs at thousands of pages per second through asynchronous Twisted callbacks but cannot execute JavaScript. Playwright can drive a real browser through any page but is a single-page-at-a-time tool that does not give you Scrapy’s middleware, item pipelines, or deduplication. The scrapy-playwright plugin bridges them: Scrapy still handles the orchestration, scheduling, and pipelines, but pages route through Playwright when needed.

This guide covers scrapy-playwright 0.0.40+ on Scrapy 2.11+ in 2026, the setup steps, the page coroutines pattern, proxy and stealth integration, and production patterns that scale. Code is Python 3.12 throughout. By the end you will have a Scrapy project that can mix HTML-only requests (fast) and Playwright-rendered requests (slow but JS-capable) in one spider, with shared middleware and pipelines.

Why this combination wins

Scrapy alone is unbeatable for HTML-only scraping at scale. Its async model, request prioritization, deduplication, and pipeline architecture are mature. The weakness is that any JavaScript-rendered page returns empty HTML.

Playwright alone handles JS perfectly but has no concept of crawl orchestration. You write your own request queue, deduplication, retry logic, and item processing.

scrapy-playwright lets each page in a Scrapy spider opt into Playwright via a metadata flag. Pages without the flag stay as fast HTML requests. Pages with the flag route through a Playwright pool, render fully, and return rendered HTML to your callback as if Scrapy fetched it directly.

For Scrapy’s official docs, see Scrapy 2.11+ documentation. For Playwright’s, see Playwright Python documentation.

Installation

pip install scrapy scrapy-playwright playwright
playwright install chromium

Verify versions:

scrapy version       # 2.11.0 or later
python -c "import scrapy_playwright; print(scrapy_playwright.__version__)"
# 0.0.40 or later

Project setup

Create a Scrapy project:

scrapy startproject dynamic_scraper
cd dynamic_scraper

Edit dynamic_scraper/settings.py to enable scrapy-playwright:

# settings.py
DOWNLOAD_HANDLERS = {
    "http": "scrapy_playwright.handler.ScrapyPlaywrightDownloadHandler",
    "https": "scrapy_playwright.handler.ScrapyPlaywrightDownloadHandler",
}

TWISTED_REACTOR = "twisted.internet.asyncioreactor.AsyncioSelectorReactor"

PLAYWRIGHT_BROWSER_TYPE = "chromium"
PLAYWRIGHT_LAUNCH_OPTIONS = {
    "headless": True,
    "args": [
        "--disable-blink-features=AutomationControlled",
        "--disable-features=IsolateOrigins",
    ],
}

# How many concurrent Playwright contexts to maintain
PLAYWRIGHT_MAX_CONTEXTS = 8
PLAYWRIGHT_DEFAULT_NAVIGATION_TIMEOUT = 30000

# Reasonable defaults for mixed crawls
CONCURRENT_REQUESTS = 16
CONCURRENT_REQUESTS_PER_DOMAIN = 4
DOWNLOAD_DELAY = 0
RETRY_TIMES = 3

# Standard middleware for header rotation, etc
DOWNLOADER_MIDDLEWARES = {
    "scrapy.downloadermiddlewares.useragent.UserAgentMiddleware": None,
    "scrapy_user_agents.middlewares.RandomUserAgentMiddleware": 400,
    "dynamic_scraper.middlewares.ProxyMiddleware": 410,
}

# Item pipeline (whatever you need)
ITEM_PIPELINES = {
    "dynamic_scraper.pipelines.JsonExportPipeline": 300,
}

The download handlers route every HTTP request through scrapy-playwright. Pages without the playwright meta flag are still fetched via regular HTTP (scrapy-playwright detects this automatically), so you do not pay the browser cost for pages that do not need it.

A first spider mixing HTML and Playwright requests

Most real spiders mix request types. The product listing might be static HTML, but the product detail might require JavaScript. scrapy-playwright handles this cleanly:

# spiders/products.py
import scrapy
from scrapy_playwright.page import PageMethod


class ProductSpider(scrapy.Spider):
    name = "products"
    allowed_domains = ["example.com"]
    start_urls = ["https://example.com/products"]

    def parse(self, response):
        # Listing page is plain HTML, no Playwright needed
        for product_url in response.css("a.product-link::attr(href)").getall():
            yield response.follow(
                product_url,
                callback=self.parse_product,
                meta={
                    "playwright": True,
                    "playwright_include_page": False,
                    "playwright_page_methods": [
                        PageMethod("wait_for_selector", "div.product-detail"),
                    ],
                },
            )

        next_page = response.css("a.next::attr(href)").get()
        if next_page:
            yield response.follow(next_page, callback=self.parse)

    def parse_product(self, response):
        yield {
            "title": response.css("h1.product-title::text").get(),
            "price": response.css("span.price-current::text").get(),
            "stock": response.css("div.stock-info::text").get(),
            "description": response.css("div.product-description").get(),
        }

Key points:

• The listing page is fetched normally (no playwright: True in meta)
• Product detail pages are fetched via Playwright (playwright: True)
• playwright_page_methods runs on the Playwright page before returning the HTML
• wait_for_selector ensures the JavaScript-rendered detail is present before scrapy-playwright captures the HTML

This pattern keeps the spider’s overall structure conventional while opting specific pages into the heavyweight rendering path.

Interacting with the page

Some scraping requires interaction (click a “Show more” button, scroll to load infinite content, fill a search form). For this, request the actual Playwright Page object back via playwright_include_page=True:

import scrapy
from scrapy_playwright.page import PageMethod


class InfiniteScrollSpider(scrapy.Spider):
    name = "infinite"
    start_urls = ["https://example.com/feed"]

    def start_requests(self):
        for url in self.start_urls:
            yield scrapy.Request(
                url,
                meta={
                    "playwright": True,
                    "playwright_include_page": True,
                    "playwright_page_methods": [
                        PageMethod("wait_for_selector", "article.feed-item"),
                    ],
                },
                callback=self.parse_feed,
                errback=self.errback_close_page,
            )

    async def parse_feed(self, response):
        page = response.meta["playwright_page"]
        # Scroll to load more items
        for _ in range(5):
            await page.evaluate("window.scrollBy(0, document.body.scrollHeight)")
            await page.wait_for_timeout(2000)
        html = await page.content()
        await page.close()

        # Parse the now-loaded HTML with Scrapy selectors
        from scrapy.http import HtmlResponse
        new_response = HtmlResponse(url=response.url, body=html, encoding="utf-8")
        for article in new_response.css("article.feed-item"):
            yield {
                "title": article.css("h2::text").get(),
                "url": article.css("a::attr(href)").get(),
                "summary": article.css("p.summary::text").get(),
            }

    async def errback_close_page(self, failure):
        page = failure.request.meta.get("playwright_page")
        if page:
            await page.close()

The playwright_include_page=True means the Page object is attached to the response. Your callback can interact with it (scroll, click, wait) before extracting HTML. Always close the page in your callback or errback to avoid leaking browser contexts.

Proxy integration

Pass proxies to Playwright either at the launch level (one proxy for all contexts) or per-context (different proxy per request). Per-context is the more flexible pattern.

# settings.py addition
PLAYWRIGHT_CONTEXTS = {
    "default": {
        "proxy": {
            "server": "http://squid.internal:3128",
            "username": "scraper_user",
            "password": "secret",
        },
    },
}

For per-request proxies via a custom middleware:

# middlewares.py
import random


class ProxyRotationMiddleware:
    PROXIES = [
        "http://user1:pass1@proxy1.provider.com:8080",
        "http://user2:pass2@proxy2.provider.com:8080",
        "http://user3:pass3@proxy3.provider.com:8080",
    ]

    def process_request(self, request, spider):
        if request.meta.get("playwright"):
            # For Playwright requests, set the context with a specific proxy
            proxy = random.choice(self.PROXIES)
            user_pass, host_port = proxy.split("//")[1].split("@")
            user, password = user_pass.split(":")
            host, port = host_port.split(":")
            request.meta["playwright_context"] = f"proxy_{hash(proxy)}"
            request.meta["playwright_context_kwargs"] = {
                "proxy": {
                    "server": f"http://{host}:{port}",
                    "username": user,
                    "password": password,
                },
            }
        else:
            # For regular HTTP requests, set the standard proxy meta
            request.meta["proxy"] = random.choice(self.PROXIES)

scrapy-playwright caches contexts by name, so reusing the same proxy uses the same context. This avoids spawning a new browser context per request, which is expensive.

Stealth: integrating patchright

For sites that fingerprint, use patchright (Playwright stealth fork) instead of vanilla Playwright. patchright is a drop-in replacement, so you can swap it in via setting the playwright module:

# In your settings.py or as an environment variable
import patchright as playwright_module

PLAYWRIGHT_BROWSER_TYPE = "chromium"
# scrapy-playwright will use whatever playwright module is installed

Or install patchright as the playwright package (it shadows vanilla playwright):

pip install patchright
playwright install chromium

After install, scrapy-playwright automatically uses patchright’s launch logic, which applies stealth patches.

For fine-grained stealth on a per-request basis, inject scripts via init_script:

yield scrapy.Request(
    url,
    meta={
        "playwright": True,
        "playwright_context_kwargs": {
            "user_agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
                          "(KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36",
        },
        "playwright_page_init_script": """
            Object.defineProperty(navigator, 'webdriver', { get: () => undefined });
            Object.defineProperty(navigator, 'plugins', { get: () => [1, 2, 3] });
        """,
    },
)

For the full picture on browser fingerprinting, see canvas fingerprinting bypass and WebGL fingerprinting bypass.

Performance: how much does Playwright slow you down

Throughput comparison on a single worker:

The cost is real. Use Playwright only for pages that need it. The mixed pattern (HTML for listings, Playwright for details) is the right balance for most sites.

To scale Playwright throughput:

• Increase PLAYWRIGHT_MAX_CONTEXTS (each context costs ~100 MB RAM)
• Reuse contexts across requests by giving them stable names
• Run multiple Scrapy instances on different machines
• Keep navigation timeout tight (30 seconds is usually enough)

Comparison: scrapy-playwright vs alternatives

For most Python teams in 2026, scrapy-playwright is the right pick. For Node teams, Crawlee. For very simple JS sites, scrapy-splash still works but is on its way out.

For the alternative Crawlee path, see Scrapy Cloud vs Crawlee Cloud in 2026.

Item pipelines and Playwright

Item pipelines work identically whether the request was HTML-only or Playwright. The pipeline sees a Scrapy Item (or dict) and processes it. No changes needed:

# pipelines.py
import json
from datetime import datetime


class JsonExportPipeline:
    def open_spider(self, spider):
        self.file = open(f"output/{spider.name}_{datetime.now():%Y%m%d_%H%M%S}.jsonl", "w")

    def close_spider(self, spider):
        self.file.close()

    def process_item(self, item, spider):
        line = json.dumps(dict(item)) + "\n"
        self.file.write(line)
        return item

This is one of the main wins of scrapy-playwright over rolling your own: you get to keep all your existing Scrapy pipelines and middleware.

Common patterns

Wait for specific elements before extracting:

yield scrapy.Request(
    url,
    meta={
        "playwright": True,
        "playwright_page_methods": [
            PageMethod("wait_for_selector", "div.product-loaded"),
            PageMethod("wait_for_load_state", "networkidle"),
        ],
    },
)

Click a button to load content:

yield scrapy.Request(
    url,
    meta={
        "playwright": True,
        "playwright_page_methods": [
            PageMethod("click", "button.show-more"),
            PageMethod("wait_for_selector", "div.expanded-content"),
        ],
    },
)

Take a screenshot for debugging:

yield scrapy.Request(
    url,
    meta={
        "playwright": True,
        "playwright_page_methods": [
            PageMethod("screenshot", path="debug.png", full_page=True),
        ],
    },
)

Intercept network requests:

async def parse(self, response):
    page = response.meta["playwright_page"]
    # Capture all XHR responses
    captured_data = []
    page.on("response", lambda r: captured_data.append(r.url))
    await page.wait_for_load_state("networkidle")
    await page.close()
    yield {"all_xhr_urls": captured_data}

Production deployment

A production scrapy-playwright deployment looks like:

• Compute: container with Chromium installed (use playwright-base images)
• Concurrency: tune CONCURRENT_REQUESTS based on RAM (each Playwright context is ~100 MB)
• Proxy: routed through Squid or commercial pool
• Stealth: patchright as drop-in
• Monitoring: Scrapy stats + Playwright context counter
• Persistence: items to a database or queue, not just files
• Scheduling: scrapyd, scrapy-cluster, or Airflow

A reasonable container Dockerfile:

FROM mcr.microsoft.com/playwright/python:v1.45.0-jammy

WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

COPY . .

ENV PLAYWRIGHT_BROWSERS_PATH=/ms-playwright

CMD ["scrapy", "crawl", "products"]

Use the playwright-python base image to skip installing Chromium and dependencies separately.

Operational checklist

For production scrapy-playwright deployments in 2026:

• Scrapy 2.11+, scrapy-playwright 0.0.40+
• Use playwright-python base image for containers
• patchright for stealth
• Per-request opt-in to Playwright via meta flag
• Always close pages in callbacks and errbacks
• Use named contexts to reuse across requests
• Tune CONCURRENT_REQUESTS based on RAM
• Set navigation timeouts (30s default is reasonable)
• Log per-spider stats including Playwright vs HTML request count
• Pair with proxy rotation middleware
• Monitor RAM usage closely (Playwright contexts leak if not closed)

FAQ

Q: can I use Playwright with Scrapy without scrapy-playwright?
You can, but it is painful. You would have to write your own download handler, manage the browser lifecycle, and integrate with Scrapy’s async model. scrapy-playwright handles all this for you.

Q: what about scrapy-splash?
Splash is essentially unmaintained as of 2024-2025. The JS engine is older and many modern SPAs do not render correctly. Use scrapy-playwright unless you have legacy reasons to stay on Splash.

Q: how do I handle Playwright authentication and cookies?
Pass cookies via the request meta or set them on the context:

meta={"playwright": True, "playwright_context_kwargs": {"storage_state": "auth.json"}}

auth.json is generated by saving Playwright’s context.storage_state() after login.

Q: can I run scrapy-playwright on AWS Lambda?
Technically yes via Playwright’s Lambda layers, but the cold start is brutal (10+ seconds per invocation just to launch Chromium). For Lambda, prefer dedicated browser services like Browserbase or run scrapy-playwright on Fargate or EKS instead.

Q: does scrapy-playwright support Firefox?
Yes. Set PLAYWRIGHT_BROWSER_TYPE = "firefox" and run playwright install firefox. Most patches work on Firefox too, though some stealth libraries are Chromium-focused.

Common pitfalls in production scrapy-playwright

The first failure mode that catches teams off guard is the page-object leak. scrapy-playwright passes the open Page to your callback via response.meta["playwright_page"], but if your parse function raises an exception or yields requests without closing the page, the browser context retains a reference. After 50-100 leaked pages, Chromium hits its per-context limit (typically 30 tabs depending on RAM) and new requests stall waiting for tab slots. The fix is a try/finally pattern in every parse method that touches playwright_page:

async def parse_product(self, response):
    page = response.meta.get("playwright_page")
    try:
        await page.wait_for_selector(".product-loaded", timeout=10000)
        title = await page.text_content("h1")
        yield {"title": title}
    except Exception as e:
        self.logger.error(f"parse failed: {e}")
        raise
    finally:
        if page:
            await page.close()

Add an errback that also closes the page on download failure, otherwise pages leak from network errors that never reach your parse callback.

The second pitfall is the headers-vs-Playwright disconnect. Setting headers on a scrapy.Request does not automatically pass them to Playwright when playwright=True. Scrapy treats those headers as the eventual download request, but Playwright bypasses Scrapy’s downloader entirely and uses its own. The result: your User-Agent rotation middleware appears to set a Chrome 124 UA but Playwright launches with whatever UA the browser instance has. The fix is to set headers via playwright_context_kwargs on each request:

yield scrapy.Request(
    url,
    meta={
        "playwright": True,
        "playwright_context_kwargs": {
            "user_agent": rotated_ua,
            "extra_http_headers": rotated_headers,
        },
    },
)

Each new context spawn now gets the rotated UA and headers. Beware: spawning a new context per request defeats Playwright’s connection reuse, so pair this with named contexts to amortize the cost across same-UA requests.

The third pitfall is silent JavaScript navigation that bypasses Scrapy’s URL tracking. When a page does window.location.href = "/next-page" mid-load, Playwright follows it and ends up at a URL different from what you originally requested. Scrapy’s response.url reflects the post-navigation URL, but your duplicate-detection middleware only saw the original URL. You can end up scraping the same final page repeatedly through different entry-point URLs. The fix is to track final URLs in your DupeFilter:

from scrapy.dupefilters import RFPDupeFilter
import hashlib

class FinalUrlDupeFilter(RFPDupeFilter):
    def request_seen(self, request):
        # If response has been received, dedupe on response.url too
        final_url = request.meta.get("playwright_final_url")
        if final_url:
            fp = hashlib.sha256(final_url.encode()).hexdigest()
            if fp in self.fingerprints:
                return True
            self.fingerprints.add(fp)
        return super().request_seen(request)

Set playwright_final_url in your parse callback after capturing page.url, then re-yield the request through the dedupe filter.

Real-world example: 50,000-product crawl with mixed HTML and Playwright

A scrapy-playwright project crawled an electronics retailer with 50,000 products across 800 category pages. Initial implementation used Playwright for everything and took 18 hours per full crawl with 90 percent CPU usage on a 16-core machine. The optimization that cut this to 2.5 hours was a tiered request strategy that classified URL patterns and used the cheapest method per pattern:

class TieredSpider(scrapy.Spider):
    name = "products_tiered"

    def start_requests(self):
        # Category pages: pure HTML, server-rendered
        for cat_url in self.category_urls:
            yield scrapy.Request(cat_url, callback=self.parse_category)

    def parse_category(self, response):
        # Extract product URLs from server-rendered HTML
        for url in response.css(".product-link::attr(href)").getall():
            full_url = response.urljoin(url)
            # Tier 1: standard product page (HTML works)
            if "/p/" in full_url:
                yield scrapy.Request(full_url, callback=self.parse_product_html)
            # Tier 2: SPA product detail (Playwright needed)
            elif "/spa-product/" in full_url:
                yield scrapy.Request(
                    full_url,
                    meta={
                        "playwright": True,
                        "playwright_page_methods": [
                            PageMethod("wait_for_selector", ".price-loaded"),
                        ],
                    },
                    callback=self.parse_product_playwright,
                )

    def parse_product_html(self, response):
        yield {
            "url": response.url,
            "title": response.css("h1::text").get(),
            "price": response.css(".price::text").get(),
        }

    async def parse_product_playwright(self, response):
        page = response.meta["playwright_page"]
        try:
            title = await page.text_content("h1")
            price = await page.text_content(".price")
            yield {"url": response.url, "title": title, "price": price}
        finally:
            await page.close()

Of the 50,000 products, only 8,000 needed Playwright (the SPA product pages). The remaining 42,000 went through pure Scrapy HTML at 200+ requests/second. Playwright handled the 8,000 SPA pages at 5 requests/second. Total throughput: 5.5 hours for the HTML tier and 27 minutes for the SPA tier, running concurrently in 2.5 hours total. The lesson: classify your URLs by render type and route accordingly. Default-to-Playwright is the most expensive choice you can make.

Comparison: scrapy-playwright settings tuning by site type

Different site types need different tuning profiles. A reference table from a 2026 production deployment:

For listing pages with infinite scroll, use a manual scroll loop:

async def scroll_until_done(page, max_scrolls=20):
    for i in range(max_scrolls):
        prev_height = await page.evaluate("document.body.scrollHeight")
        await page.evaluate("window.scrollTo(0, document.body.scrollHeight)")
        await page.wait_for_timeout(1500)
        new_height = await page.evaluate("document.body.scrollHeight")
        if new_height == prev_height:
            break  # no more content loaded

Without the scroll-until-done pattern, you get only the first viewport’s worth of items even when the listing has 500+. Pair with a per-spider MAX_SCROLL_ITERATIONS setting so a runaway page does not loop forever.

Memory profiling: detecting Playwright leaks

Production scrapy-playwright deployments suffer from RAM growth that takes hours to manifest. The cause is almost always context reuse without bounded lifetime. Each long-lived context accumulates DOM nodes, IndexedDB entries, and service worker state. After 4-6 hours of constant scraping through one context, RAM usage doubles. The detection pattern:

import psutil
import os

class MemoryMonitorMiddleware:
    def __init__(self):
        self.process = psutil.Process(os.getpid())
        self.baseline_mb = self.process.memory_info().rss / 1024 / 1024
        self.requests_processed = 0

    def process_response(self, request, response, spider):
        self.requests_processed += 1
        if self.requests_processed % 100 == 0:
            current_mb = self.process.memory_info().rss / 1024 / 1024
            growth = current_mb - self.baseline_mb
            spider.logger.info(
                f"requests={self.requests_processed} "
                f"rss={current_mb:.0f}MB growth={growth:.0f}MB"
            )
            if growth > 2000:  # 2GB growth, recycle contexts
                spider.logger.warning("memory bloat detected, recycling contexts")
                # Trigger context recycling in your context manager
        return response

Recycle contexts every 1000-2000 requests by closing them and letting scrapy-playwright spawn fresh ones. This caps RAM growth at the cost of one cold-context launch per recycle (about 800ms). For long-running spiders, this is a worthwhile trade.

Wrapping up

Scrapy + Playwright is the right combination for any Python scraping project that mixes simple HTML pages with JavaScript-heavy pages. scrapy-playwright glues them cleanly enough that you keep all of Scrapy’s pipeline ergonomics while getting Playwright’s rendering capability where you need it. Pair this guide with our Scrapy Cloud vs Crawlee Cloud and best Python scraping libraries 2026 writeups, and browse the framework-tutorials category on DRT for related tutorials.

Self-hosted proxy infrastructure: complete 2026 guide

Self-hosted proxy infrastructure is a category that ranges from “I run Squid in front of five purchased upstream proxies” to “I operate a 200-modem mobile proxy farm with ISP business contracts and four-9s uptime.” This guide is about the latter end of the spectrum: the build that actually saves money compared to buying from Bright Data or Oxylabs at scale, the operational reality of running it, and the break-even point where self-hosting starts paying off.

The numbers in this guide come from operating real proxy infrastructure on the SMP (Singapore Mobile Proxy) and similar farms in the Singapore-Malaysia region. The pricing, hardware choices, and ISP dynamics are 2026 reality. If you are scraping under 1 million pages per month, self-hosting is probably not worth it. Above 10 million, the unit economics are compelling. In between is the gray zone where it depends on your engineering capacity.

Why self-host

The case for self-hosted proxies in 2026:

• Cost: at scale, self-hosted mobile proxies cost $0.30-$1 per GB versus $5-15/GB from premium vendors
• Control: you choose the ISPs, IP ranges, geo distribution, and rotation behavior
• Reliability: no upstream vendor outages affecting your business
• Privacy: your traffic does not pass through a third-party gateway that logs everything
• Customer billing: you can resell capacity if you have spare throughput

The case against:

• Capex: hardware, racks, network equipment cost upfront ($50k+ for a serious build)
• Opex: data plans, electricity, internet uplink, maintenance hours
• ISP relationships: requires business contracts, sometimes legal entity in the country
• 24/7 ops: things break at 3am, you fix them
• Capacity inflexibility: cannot quickly add 500 more IPs the way Bright Data can

Most teams should buy proxies. Self-hosting is for the small number of teams whose volume justifies the operational burden.

Architecture: the SMP-style mobile proxy farm

A typical mid-scale mobile proxy farm:

              + customers (scrapers, ad-verification, OSINT) +
                              |
                              | HTTPS to public IP
                              v
                  +-------------------------+
                  |   Edge proxy (Squid +   |
                  |   ProxySmart controller) |
                  +-------------------------+
                              |
                              | private 10.x routing
                              v
              +-------+-------+-------+-------+
              |       |       |       |       |
           server1 server2 server3 server4 ...
              |       |       |       |       |
              | USB  | USB  | USB  | USB  |
              v       v       v       v       v
            modem1  modem2  modem3  modem4  modem5  ...
            (LTE)  (LTE)  (LTE)  (LTE)  (LTE)
              |       |       |       |       |
              v       v       v       v       v
            ISP A    ISP B    ISP A    ISP C    ISP B   (carrier APNs)

Each modem produces one mobile IP via PPP-over-LTE through the carrier’s APN. The server hosts a process that exposes that modem as an HTTP/SOCKS5 port to the controller. The controller assigns ports to customers, handles rotation, and meters usage.

This pattern scales modularly: add a server with 50 modems, expose 50 more ports, customers see them as additional capacity. Each modem is one IP with full carrier-grade NAT obfuscation, which is the reason mobile IPs are valuable for scraping (they look like millions of real users sharing the carrier’s NAT pool).

Hardware: per-server build

Per-server hardware for a 50-modem build:

Cost per server: roughly $4,000-6,000 in hardware plus $300-500/month in SIM data plans (varies by country and plan size).

For the full multi-server build:

• 4-server cluster: ~$20,000 hardware + $1,500/month opex
• Yields 200 mobile IPs
• Throughput: ~10 Mbit/s per modem aggregate, so 2 Gbit/s peak across the farm
• Bandwidth budget: usually 50-200 GB per modem per month (carrier-dependent)

Modems are the failure-prone component. Plan for 5-10% modem replacement per year. Keep spare modems and SIMs on hand.

Software stack

Each server runs:

• Linux: Ubuntu 22.04 LTS or 24.04 LTS (most modem driver compatibility)
• PPP daemon: pppd per modem with carrier-specific APN config
• ProxySmart: per-modem HTTP/SOCKS proxy daemon exposing the PPP IP as a port
• Controller agent: communicates with the central controller for port assignment, rotation, telemetry

ProxySmart specifically wraps each modem’s PPP interface and exposes it as an HTTP CONNECT proxy. Customers connect to a port (e.g., server2.example.com:8123) and traffic egresses through the modem’s mobile IP.

For DNS within the proxy farm, run a local resolver (unbound or dnsmasq) so that DNS queries do not leak to ISP resolvers, which can throttle or log them.

Network: ISP business contracts

The hardest part of self-hosting mobile proxies is not the technology, it is the SIM card supply. Personal-line carriers throttle or terminate SIMs that are clearly used for proxying. You need:

• Business SIM contracts: a corporate account with 50+ lines, possibly negotiated bulk pricing
• Data plans: typically 100-200 GB/month per line, billed as a corporate package
• APN access: business APN (sometimes different from consumer APN)
• Rotating contract terms: some carriers prohibit “automation” in their AUP; you need terms that allow this use

Country-specific dynamics in 2026:

Without a business relationship, the SIMs get shut down within weeks of heavy use. With a relationship, they last months or years.

Carrier rotation: rebooting modems for fresh IPs

Mobile IPs rotate on PPP reconnection. To get a new IP:

1. Send AT command to modem to reset
2. Wait for PPP to drop and reestablish
3. New IP is assigned by the carrier

Time to rotate: 10-30 seconds per modem. Customers see this as a “rotate” API call that returns a new IP within 30 seconds.

# Sketch of modem rotation script
import subprocess
import time

def rotate_modem(modem_id: int) -> str:
    # Send AT command via the modem's serial interface
    subprocess.run(
        ["echo", "AT+CFUN=4", ">", f"/dev/ttyUSB{modem_id}_at"],
        shell=True,
    )
    time.sleep(2)
    subprocess.run(
        ["echo", "AT+CFUN=1", ">", f"/dev/ttyUSB{modem_id}_at"],
        shell=True,
    )
    # Wait for PPP to come back up
    for _ in range(30):
        time.sleep(1)
        ip = subprocess.check_output(
            ["ip", "addr", "show", f"ppp{modem_id}"]
        ).decode()
        if "inet " in ip:
            new_ip = ip.split("inet ")[1].split("/")[0]
            return new_ip
    raise TimeoutError(f"modem {modem_id} did not reconnect")

In production, ProxySmart handles this via a REST API exposed per modem. Customers call POST /rotate/<port_token> and get a fresh IP within 30 seconds.

Customer-facing API

A typical customer interaction:

# Customer's scraper uses the proxy port
curl -x http://user:pass@proxy.example.com:8123 https://target.example.com

# When they need rotation
curl -X POST -H "Authorization: Bearer api_key" \
    https://proxy.example.com/api/v1/rotate/port_8123

# Check current IP
curl -x http://user:pass@proxy.example.com:8123 https://api.ipify.org

The controller exposes a REST API for customer self-service:

• POST /api/v1/rotate/<port>: trigger rotation
• GET /api/v1/status/<port>: current IP, data used, plan limit
• POST /api/v1/swap/<port>: swap to a different modem (different ISP/IP range)
• GET /api/v1/usage: per-customer billing info

For an example of this kind of platform, see Singapore Mobile Proxy which exposes mobile proxies via this pattern.

Costs and break-even

Real numbers for a 200-modem mobile farm in Singapore:

Capacity: 200 IPs * 100 GB/month = 20,000 GB/month total bandwidth. At a customer price of $1.50/GB, revenue is $30,000/month at full utilization. Margin: 65% gross.

For comparison, buying 20,000 GB of mobile proxy from Bright Data at $8/GB costs $160,000/month. Self-hosting saves $150k/month at this scale.

Break-even: roughly 30 modems / 3,000 GB/month is the point where self-hosting beats buying. Below that, the engineering and ops overhead is not worth the savings.

Monitoring

Critical metrics:

• Per-modem PPP uptime: alert when any modem is down >5 minutes
• Per-modem data usage: track against carrier plan limits
• Per-port latency: median and p95 from edge proxy
• Per-port throughput: bytes/sec by direction
• Carrier IP reputation: periodically check IPs against known-bot lists
• Modem hardware health: temperature, USB hub status

Tooling:

• Prometheus + node_exporter on each server
• Custom exporter for modem state (PPP uptime, IP, data usage)
• Grafana dashboards for operator view
• AlertManager routing to PagerDuty or Telegram

Sample dashboard panels:

Failure modes

Real failure modes from running this kind of farm:

• USB hub disconnect: a 16-port hub stops responding, takes down 16 modems. Fix: replace hub or USB-reset.
• Modem firmware crash: single modem stops responding. Fix: USB-reset or physical reboot.
• PPP daemon hang: ppp0 interface up but no traffic. Fix: kill pppd, restart.
• Carrier IP block: ISP blocks one of your IP ranges. Fix: rotate to different APN or wait.
• SIM data plan exhausted: SIM stops working. Fix: switch to backup SIM or wait for plan reset.
• DNS poisoning by carrier: ISP injects DNS responses. Fix: use local resolver via DoT/DoH to bypass ISP.
• Server crash: takes down all modems on that server. Fix: redundancy, multiple servers.
• Network uplink flap: brief outages drop customer connections. Fix: dual uplinks with LACP.

The annual maintenance burden is roughly 100-200 engineering hours per 200-modem farm.

Comparison: self-host vs commercial

Self-host when: high volume in one geo, regulatory needs to keep traffic local, custom rotation requirements, or building a product.

Commercial when: low volume, multi-geo needed, no engineering capacity for ops, want zero burden.

For broader proxy strategy, see best mobile proxy providers 2026 and building a custom rotating proxy pool with Squid.

Legal and compliance

Operating proxy infrastructure has compliance touchpoints:

• Local telecoms law: in some countries, providing internet access requires a license
• Customer KYC: depending on jurisdiction, you may need to verify customer identity
• Logging requirements: some countries require ISPs to retain connection logs
• AUP enforcement: customer abuse (CSAM, fraud) can implicate you if not addressed
• Tax: B2B services have VAT/GST in many jurisdictions

For OWASP guidance on related security topics, see OWASP Top 10. For the specifics of running a regulated telecom service in your country, consult local counsel.

Operational checklist

For running a self-hosted mobile proxy farm in 2026:

• Hardware sized for 50-200 modems per server
• Business SIM contracts with carriers
• Local DNS resolver (unbound)
• Per-modem PPP daemons + ProxySmart
• Central controller for port assignment
• Customer REST API
• Prometheus + Grafana monitoring
• 24/7 alerting on PPP uptime and data usage
• Spare modems on hand (10% of fleet)
• Regular IP reputation checks against deny lists
• Customer billing system (Stripe, BMS, custom)
• Legal review of AUP and terms

Smaller-scale alternative: residential at home

If you cannot justify mobile farm capex, a smaller residential setup at scale:

1. Get 5-10 residential ISP lines at home or in a colo with multiple ISPs
2. Run Squid or HAProxy on a single machine routing across the lines
3. Each line gives you one residential IP
4. Total cost: $200-500/month for the lines plus colo

Capacity is small (maybe 1-10 GB/day per line) but the IPs are real residential, very high quality. Useful for low-volume high-value scraping (account creation, captcha solving) where IP quality matters more than capacity.

FAQ

Q: how long do mobile SIMs last before getting flagged or shut down?
With a business contract: months to years. Without one: weeks. The relationship with the carrier is the durability factor, not the technology.

Q: can I run this on AWS or GCP?
Mobile proxies require physical USB-connected modems, so no. You can run the controller and edge proxy in cloud, but the modems themselves need physical hosting (your office, a colo, or a residential location).

Q: what about IPv6?
Most carriers in 2026 are dual-stack. ProxySmart supports both. For target sites that prefer IPv4 (most do), force IPv4 routing. For target sites with IPv6, expose the v6 address.

Q: is self-hosting worth it for residential proxies (not mobile)?
Less so. Residential ISP lines are slower to provision, harder to multi-tenant, and have hostile AUPs. Mobile is the better fit for self-hosting because of carrier-grade NAT and the established business-line market.

Q: how do I price for customers?
Common patterns: per-GB ($1-3/GB), per-port-month ($30-100 per dedicated port), or hybrid. Mobile proxies command higher prices than datacenter or residential. Match competitor pricing in your geo and undercut by 20% to get traction.

Common pitfalls in production self-hosted proxy infrastructure

The first failure mode is USB hub contention. Mobile proxy farms typically connect 16-32 USB modems per host through powered USB 3.0 hubs. Cheap hubs (under $40) commonly share one USB controller across all 16 ports, which means total throughput across the hub caps at roughly 480 Mbit/s shared. With 16 LTE modems each capable of 50-100 Mbit/s, the hub bandwidth becomes the bottleneck before the modems do. The fix is per-port-controlled hubs (Yepkit YKUSH3 or VIA-chipset hubs from D-Link DUB-H7 generation) where each port has its own USB controller chain. Budget $80-120 per 7-port hub. Verify with lsusb -t that each modem reports under a separate Bus rather than a shared one.

The second pitfall is carrier-side IP rotation policy mismatch. Different carriers rotate the LTE-assigned IP at different intervals: Singtel rotates every 4 hours by default, M1 rotates every 24 hours, StarHub rotates only on PDP context teardown (so it can stay sticky for days). If your scraper expects a stable IP per session and you provision a Singtel modem, your customer’s session token bound to the original IP becomes invalid after 4 hours. Match carrier choice to use case: use StarHub for sticky sessions, use Singtel for high-rotation requirements. Document each modem’s carrier in your inventory and route customer traffic accordingly.

The third pitfall is modem firmware silent regression. Quectel and Fibocom modems push firmware updates over the air for some operators. A firmware bump from EM05CEFAR06A07M4G to EM05CEFAR07A02M4G can change the AT command set, the PPP negotiation timing, or the TLS-handshake-passthrough behavior. Your provisioning scripts that worked yesterday produce silent failures today. The mitigation is firmware pinning where the modem allows it (AT+QFOTADL=0 to disable Quectel OTA on Quectel EM05/EM06/EC25 modems), or a weekly automated firmware-version inventory that flags any modem whose AT+CGMR output changed since last week.

Real-world example: month-end IP exhaustion recovery

A self-hosted mobile farm running 60 modems across three Singapore carriers hit a customer-facing outage on month-end of February 2026. Sixteen Vivifi modems exhausted their 100GB monthly data caps within the first 12 hours of February 28 because three large customers had over-rotated through the same proxy pool. Carrier behavior on cap exhaustion: the modem still attaches, PPP comes up, but every TCP connection times out or returns 503 from a carrier portal page. The pool’s automated health checks reported the modems “online” because PPP was up, but real customer traffic was 100 percent failing.

The recovery had three steps:

# Step 1: detect data exhaustion via active probe (not just PPP state)
async def is_modem_alive(port: int) -> bool:
    try:
        resp = await fetch_via_proxy(
            f"http://192.168.1.1:{port}",
            target_url="https://1.1.1.1/cdn-cgi/trace",
            timeout=5,
        )
        return resp.status == 200 and "ip=" in resp.text
    except Exception:
        return False

# Step 2: auto-rotate exhausted modems out of the active pool
exhausted = [m for m in MODEMS if not await is_modem_alive(m.port)]
for modem in exhausted:
    await mark_modem_unavailable(modem)

# Step 3: rotate to LTESpace reserve SIMs from the spare inventory
reserve = await get_reserve_sims(carrier="vivifi", min_data_gb=80)
for spare_sim, dead_modem in zip(reserve, exhausted):
    await swap_sim(dead_modem.imei, spare_sim.iccid)

The recovery took 47 minutes from first alert to full pool restoration. Lessons: PPP-state monitoring is necessary but not sufficient (must also probe an active TCP destination), reserve SIM inventory needs to be maintained at 20 percent of fleet size to absorb month-end exhaustion spikes, and customer billing should track per-port data usage daily so heavy users can be throttled before they exhaust shared pool capacity.

Wrapping up

Self-hosted proxy infrastructure pays off above ~3,000 GB/month if you have engineering capacity for ops. Below that threshold, buying from commercial vendors is the right call. For teams that build into self-hosting, the unit economics are very good: 65% gross margins, durable competitive advantage, customer trust through ownership of the stack. Pair this with our building a custom rotating proxy pool with Squid and best mobile proxy providers 2026 guides for both the build-vs-buy decision and the technical implementation, and browse the dev-tools-projects category on DRT for related infrastructure deep-dives.

Related comparison: See how Bright Data stacks up against a dedicated Singapore mobile network in our Singapore Mobile Proxy vs Bright Data comparison.