Stack for rivet-dev/rivet

• Add observation for serde_bare #5321 oskoutmw · 2026-06-23 20:23:10
• Add observation and ser/de metrics for serde calls #5320 ztnuwvts · 2026-06-23 20:23:09
• [SLOP(claude-opus-4-8-medium)] chore: add profiling cargo profile for heaptrack leak investigation #5314 xsoysnqs · 2026-06-22 22:42:01 👈
• [SLOP(claude-opus-4-8)] fix(kitchen-sink): serve built frontend from server #5306 rymmurrk · 2026-06-19 01:04:14
• main

Get stack: forklift get 5314
Push local edits: forklift submit
Merge when ready: forklift merge 5314

This PR does two things: adds a `[profile.profiling]` Cargo profile for heaptrack-based leak investigation, and refactors `run_event_loop` in `napi_actor_events.rs` to interleave `JoinSet` reaping with event dispatch so completed background tasks don't accumulate RSS over an actor's lifetime. The change is small and well-motivated. Four issues worth addressing:

`drain_tasks` inconsistency with cancelled-task error suppression
rivetkit-typescript/packages/rivetkit-napi/src/napi_actor_events.rs:195 — The new run_event_loop join arm correctly suppresses cancelled-task errors (if !error.is_cancelled()), but drain_tasks at ~line 854 still error-logs all join failures including cancellations. Tasks cancelled via abort.cancel() before the event loop exits will be silently passed in the loop drain, then if somehow reaped in drain_tasks they'd spuriously error-log. At minimum the two paths should be consistent — drain_tasks should also apply !error.is_cancelled() to match.

Profiling profile comment overstates inlining suppression
Cargo.toml:660 — The comment says "keep frames un-inlined," but opt-level = 3 is inherited from release and not overridden. codegen-units = 16 + lto = "thin" suppresses cross-codegenunit inlining, but opt-level = 3 still aggressively inlines within each codegen unit. For heaptrack allocation backtraces, intra-crate allocation sites inside hot functions will still collapse into their callers. If the goal is fully un-inlined frames, add opt-level = 1. If opt-level = 3 is intentional (to keep runtime performance representative), the comment should say "limits cross-crate inlining" rather than implying frames are generally un-inlined.

select! without biased — probabilistic event delay under task burst
rivetkit-typescript/packages/rivetkit-napi/src/napi_actor_events.rs:187 — When many background tasks complete simultaneously, select! randomly picks between join_next and events.recv() each iteration, so N completed tasks in the backlog means event processing waits an expected O(N) extra iterations before the action queue drains. This is bounded and exponentially unlikely to cause real starvation, but adding biased; to the select! (placing events.recv() first) would guarantee event processing is never delayed by a task-reap backlog and more closely matches the original code's event-priority semantics.

Missing actor_id on the new join-failure error log
rivetkit-typescript/packages/rivetkit-napi/src/napi_actor_events.rs:195 — run_event_loop receives ctx: &ActorContext which exposes actor_id. The existing error logs throughout this file include it as a structured field, but the new log does not: tracing::error!(?error, "napi background task failed to join"). Should be tracing::error!(?error, actor_id = %ctx.inner().actor_id(), "napi background task failed to join") (or however actor_id is accessed on ctx) so join failures are attributable in production.

🚅 Deployed to the rivet-pr-5314 environment in rivet-frontend

Service	Status	Web	Updated (UTC)
frontend-cloud	😴 Sleeping (View Logs)	Web	Jun 24, 2026 at 4:15 pm
website	😴 Sleeping (View Logs)	Web	Jun 24, 2026 at 5:09 am
kitchen-sink	😴 Sleeping (View Logs)	Web	Jun 23, 2026 at 8:38 pm
frontend-inspector	😴 Sleeping (View Logs)	Web	Jun 23, 2026 at 8:32 pm
ladle	✅ Success (View Logs)	Web	Jun 22, 2026 at 10:43 pm
mcp-hub	✅ Success (View Logs)	Web	Jun 22, 2026 at 10:42 pm