Daily benchmark/6 languages × 8 tasks
Cross-language latency benchmarks for Cloudflare Workers: JavaScript, TypeScript, Python, Rust, Go (TinyGo), and C++ (Emscripten), measured daily across eight task categories.
01 — Latency heatmap
Mean end-to-end latency in milliseconds for each language × category combination. Green is faster, red is slower, on a log scale so the fast cells stay distinguishable from Python. Toggle between warm requests and each run's cold first hit, or hide Python to rescale the colours across the remaining languages.
02 — Language leaderboard
Mean warm latency across all categories (square), with the observed min–max range across the category matrix (bar). Shorter is faster.
03 — Cold-start penalty
How much slower each run's first request to a worker is than the warm requests that follow it. By language shows the median across the eight categories; switch to see every worker.
04 — By category
Per-category small multiples: warm latency per language, sorted fastest first. The dashed orange line is the per-category median across languages.
05 — Methodology
Once a day a cron trigger on the driver worker calls each benchmark worker in turn with
fetch(). Latencies are the driver's view of those subrequests, measured in
whichever data center ran the cron (shown above), so they exclude the public internet hop.
Smaller is faster.
The first request a run sends to each worker. With about 24 hours since the previous run, that request usually has to load the script in that data center and start a fresh isolate. It is not a post-deploy measurement, and a real visitor in the same location could occasionally have warmed the worker first.
The mean of the requests that follow (five per run), which reuse the isolate the first hit started.
Workers freeze performance.now() until the next I/O as a Spectre mitigation, so
in-worker timers read 0 for CPU-only work. Cloudflare's Server-Timing: cfReqDur
is only sent to external clients, not to Worker subrequests, so neither appears here.
Sample sizes are small by design, and the cron's data center changes from day to day. Use the range bars to gauge noise, not just the means.
Python runs on Pyodide, and Rust, Go, and C++ run as Wasm inside the JavaScript runtime. Raw "hello" is the cleanest view of runtime overhead.