API delivery hardening
Broader caching for hot API routes — High-traffic read endpoints got longer cache lifetimes, including package lists, metadata, versions, docs/source/search, downloads, diffs, and global stats. The goal was to cut backend load while preserving cache busting around writes.
Docs and diff crawling pressure reduced — The frontend started marking per-symbol docs and diff pages as noindex, sharing more cacheable responses across callers, and applying short negative caching for 404s. Tracing and request plumbing were also improved so the real caller shows up in API logs.
Observability cleanup — API traces now export logs to Grafana via OTLP, and 4xx responses are reported as ok rather than error, making operational alerts more accurate.
Publish reliability fixes
Stranded publish tasks now requeue themselves — The publish system now detects stale tasks stuck in processing or processed and requeues them, recovering from interrupted publishes that could leave metadata incomplete or versions hidden from the resolver.
Publish status and metadata cache bugs fixed — Dynamic auth and publish-status endpoints were switched to no-store, and the CDN cache key was adjusted so purges target the public URL. That should eliminate hung publish polling and make freshly published package metadata appear immediately.
API cutover attempt and rollback
New API Worker front-door was tried, then reverted — The repo briefly switched api.jsr.io to a new workers-rs api Worker with DB client-certificate support and deployment wiring, but the traffic swap was rolled back the same day. The database auth groundwork stayed in place, but the new front-door path wasn’t ready.
Other misc changes
- CI, Terraform, and load-balancer wiring updates for the API split
- Cache write-path and middleware/proxy plumbing fixes
- Minor docs, architecture wording, and instrumentation attribute cleanups