Worker limit lab

HTTP Workers have no fixed wall-time limit while connected. This timer waits without burning CPU.

1 s

A quick baseline.

30 / 31 s

At and beyond the post-response waitUntil() window.

125 / 126 s

At and beyond the default origin proxy read timeout.

5 min

Paid Workers’ maximum CPU time.

15 min

Cron, queue consumer and alarm duration limit.

20 min

Grace period for in-flight requests during runtime updates.

These are comparison points. This request uses a timer in an HTTP handler; it does not test CPU, an origin server, or background work. The 1-hour cap is set by this app.

Search for primes up to the chosen number using trial division. Larger numbers require more CPU work, with little memory.

CPU-heavy jobs include sorting large datasets, parsing and transforming JSON, compression, image processing, and cryptographic calculations.

Workers Free allows 10 ms of CPU per request. Paid defaults to 30 s, configurable up to 5 min. Occasional overruns have some flexibility.

The timer shows elapsed wall time. Exact CPU time and exceededCpu outcomes are recorded in Cloudflare invocation logs.

Allocate buffers, touch each memory page, and retain them for one second. The value is the buffer allocation, not total memory used.

RAM-heavy jobs include buffering file downloads, loading whole CSV/JSON datasets, decoding images, and building large in-memory indexes.

Workers have 128 MB of memory per isolate, shared across concurrent requests, including the JavaScript heap and WebAssembly. Runtime overhead leaves less room for data.

128 MiB and above deliberately challenge that limit. Check exceededMemory in Cloudflare logs to confirm a memory failure.

Download a JPEG → optionally decode pixels → base64 JSON upload → wait for a mock vision API response.

Decode + send expands the JPEG into RGBA pixels inside the Worker and keeps them in memory while the API responds. The uploaded image is the original JPEG.

Base64 + send sends the compressed JPEG without expanding pixels. This models an application forwarding an image to a vision model.

A 4,096 × 4,096 image needs 64 MiB for its RGBA pixels alone. Decoder buffers, the JPEG, base64, and JSON add more memory.

The receiver is a separate HTTP endpoint on this app. It verifies the image bytes and simulates a response delay; it performs no model inference. Fixtures are reproducible test patterns.