mirror of
https://github.com/coder/coder.git
synced 2026-09-21 20:51:01 +08:00
Michael Suchacz
fc24c27dfd
## Context Follow-up fix from live UAT of the merged chat lifecycle hooks stack (#27430). Its companion UAT fix (#27655) has merged, so this targets `main` directly. ## Why? UAT measured a burst of 1,500 concurrent chat creations against a consumer with 1.2s latency. 255 were admitted and 1,245 got `502 hook_dispatch_failed (over_capacity)`, which is correct fail-closed behavior. The collateral wasn't: the same burst failed 24 `stop` dispatches, parking chats that had already been admitted and had already executed tools. One 256-slot semaphore served every event, so new-work admission could take every slot and kill turns in flight. Callers now classify each dispatch as admission or generation, and admission draws from a 192-slot gate held *before* the shared pool. At least 64 shared slots stay reachable only by dispatches for work a chat already admitted. The dispatcher is per `coderd` replica, so these limits are per replica, not deployment-wide, and the docs say so. **The caller classifies, not the event type.** Event type isn't a reliable proxy in either direction: a subagent spawn dispatches `user_prompt_submit` from inside a running turn, and the edit path dispatches `session_start` at admission time. `CapacityClassUnset` is rejected in `Dispatch`, so a new call site fails closed rather than silently inheriting a share. **Acquisition order is load-bearing.** Admission takes its own gate first. Taking a shared slot first would let admissions queued on the gate occupy the very capacity the reserve protects. `acquireCapacity` is the only path that takes either pool, so the order can't be bypassed. ## What this does not guarantee Nothing bounds how many turns generate concurrently, so the 192/64 split is a judgement call, not a derived ceiling. This stops an *admission* burst from consuming every slot; it does not make the remainder sufficient. A large enough generation load can still exhaust the reserve and error a running chat. The docs say so explicitly rather than promising a guarantee the code doesn't deliver. Generation can now take all 256 slots, so generation traffic starves admission harder than before. That's the intended priority: rejecting a new prompt is recoverable, ending a turn that already ran tools is not. ## Testing Red-green proved both new tests. Removing the release-on-failure path fails `RefusedSharedAcquireReleasesAdmission` deterministically; removing the expired-deadline check fails `ExpiredDeadlineRefusesFreeSlot` in 18/30 runs. That deadline check fixes a real race found in review. `acquire` previously shared one `time.Timer` across both acquires. Because `select` picks a ready case at random, an admission dispatch could take a slot after its capacity deadline had passed. Measured over 300 trials: 135 late acquisitions, worst overshoot 2.1ms. `acquire` now takes an absolute deadline and refuses an expired one before selecting, which measures 0/300. Go: `coderd/x/agenthooks/...` and `coderd/x/chatd/...`, plus `-race -count=3` on the dispatcher. > Mux opened this PR on Mike's behalf.