Benchmark

The same Python source on both runtimes — log regex extract, markdown rendering with markdown_it_py + jinja2, and HTML scraping with BeautifulSoup4. Three pymode workers (one per workload, each AOT-built) under *.pymode.teamchong.net, one Pyodide worker handling all three endpoints at pyodide-bench.chong1222.workers.dev.

Same Python source on both runtimes. p50/p99 come from N back-to-back requests sent by your browser when this page loaded (first request discarded as a warmup, so the numbers reflect steady-state warm behaviour).

Cold is the "Worker Startup Time" Cloudflare reports immediately after wrangler deploy — it's the cost of bringing an isolate up before any snapshot system kicks in. You pay it on the very first request to a fresh deploy, or the next request after CF evicts an idle isolate (typically a few minutes of no traffic). You do not pay it on every request — once an isolate is warm in a PoP, subsequent requests in that PoP skip the startup entirely. In production both runtimes also use memory snapshots that reduce this number further than the deploy-time signal shows.

Iterations Run again

Auto-running benchmark…

Regex log extract

POST /api/logs

Stdlib re — both runtimes can run it. Real CloudWatch log-scrubbing pattern.

	Cold	p50	p99	OK
pymode	43 ms	—	—	—
Pyodide (CF)	1827 ms	—	—	—

Markdown → HTML (markdown_it + jinja2)

POST /api/markdown

CommonMark renderer + Python template engine. Both runtimes can run the same packages from pip via uv.

	Cold	p50	p99	OK
pymode	61 ms	—	—	—
Pyodide (CF)	1827 ms	—	—	—

HTML scrape with BeautifulSoup

POST /api/scrape

Real-world page → structured title/headings/links via BeautifulSoup4. Both runtimes support bs4 via pip.

	Cold	p50	p99	OK
pymode	55 ms	—	—	—
Pyodide (CF)	1827 ms	—	—	—

KV fan-out: 20 reads per request

POST /api/kv-fanout

Read 20 keys from CF KV in a loop, aggregate to JSON. Exercises the Python ↔ JS boundary 20× per request — host-import overhead shows up here.

	Cold	p50	p99	OK
pymode	60 ms	—	—	—
Pyodide (CF)	545 ms	—	—	—

numpy: array mean + sum

POST /api/numpy

np.array([1..5]).mean() + sum(). Real numpy import + native array ops. pymode loads numpy via the python-numpy.wasm variant (linalg stubbed); Pyodide loads its pre-built numpy wheel.

	Cold	p50	p99	OK
pymode	34 ms	—	—	—
Pyodide (CF)	1559 ms	—	—	—

Each pymode workload runs on its own deployed worker under *.pymode.teamchong.net with tailored preimports — that way the Worker Startup Time CF reports at deploy varies per workload. The Pyodide comparison runs at pyodide-bench.chong1222.workers.dev. First request of each measurement is a discarded warmup so the reported p50/p99 reflect steady-state behaviour. Latencies include network round-trip from your browser.

What the columns mean

Cold is the Worker Startup Time Cloudflare reports right after wrangler deploy. You pay it on the first request to a fresh deploy, and on the next request after an idle isolate is evicted from a PoP. You do not pay it on every request. In production, Cloudflare’s Python Workers also use memory snapshots, so an actual cold-isolate spin-up in production is faster than this deploy-time signal — especially for Pyodide. The pymode number is closer to the real cold-isolate cost because pymode’s snapshot is baked into the wasm at deploy via wizer, so every isolate boot reads from the same pre-initialised state.

p50 / p99 are warm-request latencies measured live by your browser. First request of each measurement is discarded as a warmup. Numbers include network RTT from your machine to the closest Cloudflare PoP.

Honest differences between the two runtimes

	pymode	Pyodide on CF
Python	Upstream CPython 3.13 from source	Pyodide’s port (currently lagging CPython by ~1 minor version)
Deploy flow	pymode deploy — AOT-compiles a wasm tailored to the app’s imports (~10 min)	pywrangler deploy — Pyodide loads packages at deploy time (~seconds)
Pure-Python pip packages	Any (uv pip install at deploy)	Any compatible with Pyodide
Compiled C extensions	Compile via zig cc to wasm32-wasi (manual)	Curated set Cloudflare ships (numpy, pandas, Pillow, lxml, etc.)
Host imports (KV, R2, D1, fetch)	AOT-specialised: hot-path call sites patched into direct JS stubs	JsProxy/PyProxy translation per call
DurableObject classes	yes	yes (via workers-py SDK)
Per-deploy startup	tens of ms	~1-2 seconds
Vendored deploy artifact	one wasm per workload (~10 MB raw)	~1.5 MB modules + Pyodide runtime loaded by CF

Where pymode is meaningfully different

• Upstream CPython. No Pyodide-specific patches around signal handling, threading, random, or filesystem semantics. If you care about exact CPython behaviour, pymode is git-CPython compiled via zig cc. If you want 3.13 features today, you get them; Pyodide’s CPython port lags the upstream release cadence.
• AOT-specialised host imports. pymode’s deploy CLI generates a wasm tailored to the app’s imports and patches I/O hot paths (KV, R2, fetch, etc.) into direct JS stubs. The trade-off is a much longer deploy in exchange for skipping the Python ↔ JS FFI translation that Pyodide pays per host call. The page doesn’t currently include a workload that exercises this; it shows up in I/O-bound handlers that do many small KV / R2 / fetch operations per request.
• Compile-anything C extensions. Pyodide ships a curated set of compiled extensions. pymode lets you build any C extension that cross-compiles to wasm32-wasi via zig cc. The long tail of niche scientific libs or internal C modules is where this matters.

Where Pyodide is the easier choice

• Deploy iteration. pymode’s AOT build is minutes; Pyodide’s is seconds.
• Numpy / pandas / Pillow / lxml today. Pyodide ships them; pymode has the python-numpy.wasm variant but hasn’t wired numpy into the AOT deploy path yet, so a non-stdlib numpy app deploys cleanly on Pyodide and doesn’t on pymode.
• The official integration. Cloudflare maintains workers-py; pymode is a community project against the same platform.

This page exists because pymode is a personal experiment and the official Python Workers are the obvious baseline to compare against. The numbers are what they are.