Cloudflare Python Workers

Status: Beta (requires python_workers compatibility flag) Runtime: Pyodide (Python 3.12+ compiled to WebAssembly) Package Versions: workers-py@1.7.0, workers-runtime-sdk@0.3.1, wrangler@4.58.0 Last Verified: 2026-01-21

Quick Start (5 Minutes)

1. Prerequisites

Ensure you have installed:

• uv - Python package manager
• Node.js - Required for Wrangler

2. Initialize Project

bash
1# Create project directory
2mkdir my-python-worker && cd my-python-worker
3
4# Initialize Python project
5uv init
6
7# Install pywrangler
8uv tool install workers-py
9
10# Initialize Worker configuration
11uv run pywrangler init

3. Create Entry Point

Create src/entry.py:

python
1from workers import WorkerEntrypoint, Response
2
3class Default(WorkerEntrypoint):
4    async def fetch(self, request):
5        return Response("Hello from Python Worker!")

4. Configure wrangler.jsonc

jsonc
1{
2  "name": "my-python-worker",
3  "main": "src/entry.py",
4  "compatibility_date": "2025-12-01",
5  "compatibility_flags": ["python_workers"]
6}

5. Run Locally

bash
1uv run pywrangler dev
2# Visit http://localhost:8787

6. Deploy

bash
1uv run pywrangler deploy

Migration from Pre-December 2025 Workers

If you created a Python Worker before December 2025, you were limited to built-in packages. With pywrangler (Dec 2025), you can now deploy with external packages.

Old Approach (no longer needed):

python
1# Limited to built-in packages only
2# Could only use httpx, aiohttp, beautifulsoup4, etc.
3# Error: "You cannot yet deploy Python Workers that depend on
4# packages defined in requirements.txt [code: 10021]"

New Approach (pywrangler):

toml
1# pyproject.toml
2[project]
3dependencies = ["fastapi", "any-pyodide-compatible-package"]

bash
1uv tool install workers-py
2uv run pywrangler deploy  # Now works!

Historical Timeline:

• April 2024 - Dec 2025: Package deployment completely blocked
• Dec 8, 2025: Pywrangler released, enabling package deployment
• Jan 2026: Open beta with full package support

See: Package deployment issue history

Core Concepts

WorkerEntrypoint Class Pattern

As of August 2025, Python Workers use a class-based pattern (not global handlers):

python
1from workers import WorkerEntrypoint, Response
2
3class Default(WorkerEntrypoint):
4    async def fetch(self, request):
5        # Access bindings via self.env
6        value = await self.env.MY_KV.get("key")
7
8        # Parse request
9        url = request.url
10        method = request.method
11
12        return Response(f"Method: {method}, URL: {url}")

Accessing Bindings

All Cloudflare bindings are accessed via self.env:

python
1class Default(WorkerEntrypoint):
2    async def fetch(self, request):
3        # D1 Database
4        result = await self.env.DB.prepare("SELECT * FROM users").all()
5
6        # KV Storage
7        value = await self.env.MY_KV.get("key")
8        await self.env.MY_KV.put("key", "value")
9
10        # R2 Object Storage
11        obj = await self.env.MY_BUCKET.get("file.txt")
12
13        # Workers AI
14        response = await self.env.AI.run("@cf/meta/llama-2-7b-chat-int8", {
15            "prompt": "Hello!"
16        })
17
18        return Response("OK")

Supported Bindings:

• D1 (SQL database)
• KV (key-value storage)
• R2 (object storage)
• Workers AI
• Vectorize
• Durable Objects
• Queues
• Analytics Engine

See Cloudflare Bindings Documentation for details.

Request/Response Handling

python
1from workers import WorkerEntrypoint, Response
2import json
3
4class Default(WorkerEntrypoint):
5    async def fetch(self, request):
6        # Parse JSON body
7        if request.method == "POST":
8            body = await request.json()
9            return Response(
10                json.dumps({"received": body}),
11                headers={"Content-Type": "application/json"}
12            )
13
14        # Query parameters
15        url = URL(request.url)
16        name = url.searchParams.get("name", "World")
17
18        return Response(f"Hello, {name}!")

Scheduled Handlers (Cron)

python
1from workers import handler
2
3@handler
4async def on_scheduled(event, env, ctx):
5    # Run on cron schedule
6    print(f"Cron triggered at {event.scheduledTime}")
7
8    # Do work...
9    await env.MY_KV.put("last_run", str(event.scheduledTime))

Configure in wrangler.jsonc:

jsonc
1{
2  "triggers": {
3    "crons": ["*/5 * * * *"]  // Every 5 minutes
4  }
5}

Python Workflows

Python Workflows enable durable, multi-step automation with automatic retries and state persistence.

Why Decorator Pattern?

Python Workflows use the @step.do() decorator pattern because Python does not easily support anonymous callbacks (unlike JavaScript/TypeScript which allows inline arrow functions). This is a fundamental language difference, not a limitation of Cloudflare's implementation.

JavaScript Pattern (doesn't translate):

javascript
1await step.do("my step", async () => {
2  // Inline callback
3  return result;
4});

Python Pattern (required):

python
1@step.do("my step")
2async def my_step():
3    # Named function with decorator
4    return result
5
6result = await my_step()

Source: Python Workflows Blog

Concurrency with asyncio.gather

Pyodide captures JavaScript promises (thenables) and proxies them as Python awaitables. This enables Promise.all-equivalent behavior using standard Python async patterns:

python
1import asyncio
2
3@step.do("step_a")
4async def step_a():
5    return "A"
6
7@step.do("step_b")
8async def step_b():
9    return "B"
10
11# Concurrent execution (like Promise.all)
12results = await asyncio.gather(step_a(), step_b())
13# results = ["A", "B"]

Why This Works: JavaScript promises from workflow steps are proxied as Python awaitables, allowing standard asyncio concurrency primitives.

Source: Python Workflows Blog

Basic Workflow

python
1from workers import WorkflowEntrypoint, WorkerEntrypoint, Response
2
3class MyWorkflow(WorkflowEntrypoint):
4    async def run(self, event, step):
5        # Step 1
6        @step.do("fetch data")
7        async def fetch_data():
8            response = await fetch("https://api.example.com/data")
9            return await response.json()
10
11        data = await fetch_data()
12
13        # Step 2: Sleep
14        await step.sleep("wait", "10 seconds")
15
16        # Step 3: Process
17        @step.do("process data")
18        async def process_data():
19            return {"processed": True, "count": len(data)}
20
21        result = await process_data()
22        return result
23
24
25class Default(WorkerEntrypoint):
26    async def fetch(self, request):
27        # Create workflow instance
28        instance = await self.env.MY_WORKFLOW.create()
29        return Response(f"Workflow started: {instance.id}")

DAG Dependencies

Define step dependencies for parallel execution:

python
1class MyWorkflow(WorkflowEntrypoint):
2    async def run(self, event, step):
3        @step.do("step_a")
4        async def step_a():
5            return "A done"
6
7        @step.do("step_b")
8        async def step_b():
9            return "B done"
10
11        # step_c waits for both step_a and step_b
12        @step.do("step_c", depends=[step_a, step_b], concurrent=True)
13        async def step_c(result_a, result_b):
14            return f"C received: {result_a}, {result_b}"
15
16        return await step_c()

Workflow Configuration

jsonc
1{
2  "compatibility_flags": ["python_workers", "python_workflows"],
3  "compatibility_date": "2025-12-01",
4  "workflows": [
5    {
6      "name": "my-workflow",
7      "binding": "MY_WORKFLOW",
8      "class_name": "MyWorkflow"
9    }
10  ]
11}

Package Management

pyproject.toml Configuration

toml
1[project]
2name = "my-python-worker"
3version = "0.1.0"
4requires-python = ">=3.12"
5dependencies = [
6    "beautifulsoup4",
7    "httpx"
8]
9
10[dependency-groups]
11dev = [
12    "workers-py",
13    "workers-runtime-sdk"
14]

Supported Packages

Python Workers support:

• Pure Python packages from PyPI
• Pyodide packages (pre-built for WebAssembly)

See Pyodide packages list.

HTTP Clients

Only async HTTP libraries work:

python
1# ✅ WORKS - httpx (async)
2import httpx
3
4async with httpx.AsyncClient() as client:
5    response = await client.get("https://api.example.com")
6
7# ✅ WORKS - aiohttp
8import aiohttp
9
10async with aiohttp.ClientSession() as session:
11    async with session.get("https://api.example.com") as response:
12        data = await response.json()
13
14# ❌ DOES NOT WORK - requests (sync)
15import requests  # Will fail!

Requesting New Packages

Request support for new packages at: https://github.com/cloudflare/workerd/discussions/categories/python-packages

FFI (Foreign Function Interface)

Access JavaScript APIs from Python via Pyodide's FFI:

JavaScript Globals

python
1from js import fetch, console, Response as JSResponse
2
3class Default(WorkerEntrypoint):
4    async def fetch(self, request):
5        # Use JavaScript fetch
6        response = await fetch("https://api.example.com")
7        data = await response.json()
8
9        # Console logging
10        console.log("Fetched data:", data)
11
12        # Return JavaScript Response
13        return JSResponse.new("Hello!")

Type Conversions

Important: to_py() is a METHOD on JavaScript objects, not a standalone function. Only to_js() is a function.

python
1from js import Object
2from pyodide.ffi import to_js
3
4# ❌ WRONG - ImportError!
5from pyodide.ffi import to_py
6python_data = to_py(js_data)
7
8# ✅ CORRECT - to_py() is a method
9async def fetch(self, request):
10    data = await request.json()  # Returns JS object
11    python_data = data.to_py()   # Convert to Python dict
12
13# Convert Python dict to JavaScript object
14python_dict = {"name": "test", "count": 42}
15js_object = to_js(python_dict, dict_converter=Object.fromEntries)
16
17# Use in Response
18return Response(to_js({"status": "ok"}))

Source: GitHub Issue #3322 (Pyodide maintainer clarification)

Known Issues Prevention

This skill prevents 11 documented issues:

Issue #1: Legacy Handler Pattern

Error: TypeError: on_fetch is not defined

Why: Handler pattern changed in August 2025.

python
1# ❌ OLD (deprecated)
2@handler
3async def on_fetch(request):
4    return Response("Hello")
5
6# ✅ NEW (current)
7class Default(WorkerEntrypoint):
8    async def fetch(self, request):
9        return Response("Hello")

Issue #2: Sync HTTP Libraries

Error: RuntimeError: cannot use blocking call in async context

Why: Python Workers run async-only. Sync libraries block the event loop.

python
1# ❌ FAILS
2import requests
3response = requests.get("https://api.example.com")
4
5# ✅ WORKS
6import httpx
7async with httpx.AsyncClient() as client:
8    response = await client.get("https://api.example.com")

Issue #3: Native/Compiled Packages

Error: ModuleNotFoundError: No module named 'numpy' (or similar)

Why: Only pure Python packages work. Native C extensions are not supported.

Solution: Use Pyodide-compatible alternatives or check Pyodide packages.

Issue #4: Missing Compatibility Flags

Error: Error: Python Workers require the python_workers compatibility flag

Fix: Add to wrangler.jsonc:

jsonc
1{
2  "compatibility_flags": ["python_workers"]
3}

For Workflows, also add "python_workflows".

Issue #5: I/O Outside Workflow Steps

Error: Workflow state not persisted correctly

Why: All I/O must happen inside @step.do for durability.

python
1# ❌ BAD - fetch outside step
2response = await fetch("https://api.example.com")
3@step.do("use data")
4async def use_data():
5    return await response.json()  # response may be stale on retry
6
7# ✅ GOOD - fetch inside step
8@step.do("fetch and use")
9async def fetch_and_use():
10    response = await fetch("https://api.example.com")
11    return await response.json()

Issue #6: Type Serialization Errors

Error: TypeError: Object of type X is not JSON serializable

Why: Workflow step return values must be JSON-serializable.

Fix: Convert complex objects before returning:

python
1@step.do("process")
2async def process():
3    # Convert datetime to string
4    return {"timestamp": datetime.now().isoformat()}

Issue #7: Cold Start Performance

Note: Python Workers have higher cold starts than JavaScript. With Wasm memory snapshots (Dec 2025), heavy packages like FastAPI and Pydantic now load in ~1 second (down from ~10 seconds previously), but this is still ~2x slower than JavaScript Workers (~50ms).

Performance Numbers (as of Dec 2025):

• Before snapshots: ~10 seconds for FastAPI/Pydantic
• After snapshots: ~1 second (10x improvement)
• JavaScript equivalent: ~50ms

Mitigation:

• Minimize top-level imports
• Use lazy loading for heavy packages
• Consider JavaScript Workers for latency-critical paths
• Wasm snapshots automatically improve cold starts (no config needed)

Source: Python Workers Redux Blog | InfoQ Coverage

Issue #8: Package Installation Failures

Error: Failed to install package X

Causes:

• Package has native dependencies
• Package not in Pyodide ecosystem
• Network issues during bundling

Fix: Check package compatibility, use alternatives, or request support.

Issue #9: Dev Registry Breaks JS-to-Python RPC

Error: Network connection lost when calling Python Worker from JavaScript Worker Source: GitHub Issue #11438

Why It Happens: Dev registry doesn't properly route RPC calls between separately-run Workers in different terminals.

Prevention:

bash
1# ❌ Doesn't work - separate terminals
2# Terminal 1: npx wrangler dev (JS worker)
3# Terminal 2: npx wrangler dev (Python worker)
4# Result: Network connection lost error
5
6# ✅ Works - single wrangler instance
7npx wrangler dev -c ts/wrangler.jsonc -c py/wrangler.jsonc

Run both workers in a single wrangler instance to enable proper RPC communication.

Issue #10: HTMLRewriter Memory Limit with Data URLs

Error: TypeError: Parser error: The memory limit has been exceeded Source: GitHub Issue #10814

Why It Happens: Large inline data: URLs (>10MB) in HTML trigger parser memory limits. This is NOT about response size—10MB plain text works fine, but 10MB HTML with embedded data URLs fails. Common with Python Jupyter Notebooks that use inline images for plots.

Prevention:

python
1# ❌ FAILS - HTMLRewriter triggered on notebook HTML with data: URLs
2response = await fetch("https://origin.example.com/notebook.html")
3return response  # Crashes if HTML contains large data: URLs
4
5# ✅ WORKS - Stream directly or use text/plain
6response = await fetch("https://origin.example.com/notebook.html")
7headers = {"Content-Type": "text/plain"}  # Bypass parser
8return Response(await response.text(), headers=headers)

Workarounds:

• Avoid HTMLRewriter on notebook content (stream directly)
• Pre-process notebooks to extract data URLs to external files
• Use text/plain content-type to bypass parser

Issue #11: PRNG Cannot Be Seeded During Initialization

Error: Deployment fails with user error Source: Python Workers Redux Blog

Why It Happens: Wasm snapshots don't support PRNG initialization before request handlers. If you call pseudorandom number generator APIs (like random.seed()) during module initialization, deployment FAILS.

Prevention:

python
1import random
2
3# ❌ FAILS deployment - module-level PRNG call
4random.seed(42)
5
6class Default(WorkerEntrypoint):
7    async def fetch(self, request):
8        return Response(str(random.randint(1, 100)))
9
10# ✅ WORKS - PRNG calls inside handlers
11class Default(WorkerEntrypoint):
12    async def fetch(self, request):
13        random.seed(42)  # Initialize inside handler
14        return Response(str(random.randint(1, 100)))

Only call PRNG functions inside request handlers, not at module level.

Best Practices

Always Do

• Use WorkerEntrypoint class pattern
• Use async HTTP clients (httpx, aiohttp)
• Put all I/O inside workflow steps
• Add python_workers compatibility flag
• Use self.env for all bindings
• Return JSON-serializable data from workflow steps

Never Do

• Use sync HTTP libraries (requests)
• Use native/compiled packages
• Perform I/O outside workflow steps
• Use legacy @handler decorator for fetch
• Expect JavaScript-level cold start times

Framework Note: FastAPI

FastAPI can work with Python Workers but with limitations:

python
1from fastapi import FastAPI
2from workers import WorkerEntrypoint
3
4app = FastAPI()
5
6@app.get("/")
7async def root():
8    return {"message": "Hello from FastAPI"}
9
10class Default(WorkerEntrypoint):
11    async def fetch(self, request):
12        # Route through FastAPI
13        return await app(request)

Limitations:

• Async-only (no sync endpoints)
• No WSGI middleware
• Beta stability

See Cloudflare FastAPI example for details.

Official Documentation

• Python Workers Overview
• Python Workers Basics
• How Python Workers Work
• Python Packages
• FFI (Foreign Function Interface)
• Python Workflows
• Pywrangler CLI
• Pyodide Package List

Dependencies

json
1{
2  "workers-py": "1.7.0",
3  "workers-runtime-sdk": "0.3.1",
4  "wrangler": "4.58.0"
5}

Note: Always pin versions for reproducible builds. Check PyPI workers-py for latest releases.

Production Validation

• Cloudflare changelog: Dec 8, 2025 (Pywrangler + cold start improvements)
• workers-py 1.7.0: Latest stable (Jan 2026)
• Python Workflows beta: Aug 22, 2025
• Handler pattern change: Aug 14, 2025

Compatibility Date Guidance:

• Use 2025-12-01 for new projects (latest features including pywrangler improvements)
• Use 2025-08-01 only if you need to match older production Workers