cloudflare

Choose by consistency requirement first: read-modify-write, counters, locks, and ordering require a Durable Object because KV and R2 allow only 1 write per second per key with no atomicity. Then eliminate on quota: KV values cap at 25 MiB, D1 at 10 GB, and R2 objects at 5 TiB.

Use the declarative exports map (Wrangler 4.107+) with a state of renamed and renamed_to pointing at the new class. Deploy in three steps: alias the old name to the new class, apply the rename while the alias is live, then remove the alias. Delete-plus-create destroys all stored data permanently.

Isolates are reused across requests from different users, so any module-scope variable assigned inside a handler leaks one user's data to the next. Move all request-scoped state inside the fetch handler; this bug is invisible in single-user local development.

No. Local KV under wrangler dev is read-after-write consistent and local D1 and Durable Objects are single-process, so eventual-consistency, contention, and quota failures cannot reproduce locally. Diagnose production-only failures from Workers Logs, invocation outcomes, and wrangler tail.

Workers Free is fixed at 10 ms CPU; Workers Paid defaults to 30,000 ms and can be raised to 300,000 ms with limits.cpu_ms. Memory is 128 MB per isolate, not per request, so buffering large bodies with response.text() causes out-of-memory failures.

Replace pages_build_output_dir with assets.directory, add a compatibility_date, and set not_found_handling explicitly since Workers infers nothing from index.html. Set run_worker_first if middleware must run before assets, and re-apply _headers and _redirects inside the Worker for Worker-generated responses.