[santigimeno]

Workers can refill the send queue faster than the StatsD thread drains it. Draining until empty can monopolize the agent loop and prevent endpoint reconfiguration and shutdown from running.

Process at most 1024 queued sends per callback and signal another pass when work remains. Keep the existing shutdown lock and FIFO ordering.

The worker reconfiguration stress test is unchanged. Native syntax compilation and whitespace checks pass; the runtime regression still needs a rebuilt binary.

Assisted-by: Codex

[ns-control-tower]

Note

Reviewing the latest changes on this PR — this may take a few minutes.
This comment will be updated in place with the findings; no other top-level comments will be posted.

[ns-control-tower]

Walkthrough

agents/statsd/src/statsd_agent.cc: send_stats_msg_cb_ used to drain send_stats_msg_q_ until empty inside a single async callback. Under a worker flood — threads calling nsolid.statsd.sendRaw() / the native binding — one pass could run unbounded and monopolize the agent loop, delaying endpoint reconfiguration and shutdown. This change caps each pass at 1024 dequeues and, while the queue is still non-empty, re-arms send_stats_msg_ (uv_async_send) so the remainder is drained on a later loop turn. FIFO order and the stop_lock_ discipline are unchanged (+6/−1, one file).

This head (0d96c33c) is a force-push of the same commit: it wraps the previously over-long comment and was rebased onto the current base tip (node-v24.x-nsolid-v6.x @ 42202e45). The base move (otel changes landed by other PRs) is not part of this PR's diff. The PR was merged at 2026-10-28T10:14:05Z with this head.

Assessment

• ✔ Resolved — lint blocker: the comment at agents/statsd/src/statsd_agent.cc:769 is now wrapped across two lines (73/13 columns) and every changed line is ≤ 80 columns; the fresh Linters / lint-cpp run on this head passes (job). Local tools/cpplint.py + tools/checkimports.py on the file are also clean.
• ✅ No lost wakeup: send() — the only producer — always enqueues then signals under the read lock (agents/statsd/src/statsd_agent.cc:779-786); the vendored libuv clears the async pending flag before invoking the callback, so any signal issued during or after a pass schedules another pass; a drained queue stops cleanly (no busy-loop; at most one coalescing-safe uv_async_send per pass).
• ✅ Race-safe, order preserved: TSQueue::empty() and dequeue() are spinlock-guarded (src/nsolid/thread_safe.h:31-78), serialized with enqueue; a single drainer keeps FIFO order.
• ✅ Shutdown path intact: the re-arm runs while holding stop_lock_ (read); do_stop() closes send_stats_msg_ only under the write lock (agents/statsd/src/statsd_agent.cc:564-578), so the new send() cannot reach a closing handle.
• ✅ ASSERT_EQ (deps/asserts-cpp) is evaluated in release builds by design, so the re-signal is never compiled out; vendored uv_async_send returns 0 on all paths, so the assert cannot abort.
• ℹ️ CI: on this head lint-cpp passed (it was the blocker on the previous head); the other Linters jobs passed (lint-js-and-md was still running — no JS/MD is changed), and the heavy runtime jobs (Test Linux/Test macOS/Coverage */Build from tarball) were still in progress. On the previous head their failures came from unrelated agent-test suites (gRPC profiling / ZMQ heap-profile), not StatsD.
• (Re-review of force-pushed head 0d96c33c; the PR was merged before this run completed, so no separate review object is posted — this comment carries the verdict.)

Verdict: APPROVE — the only prior blocker (lint-cpp) is verifiably fixed on this head; the queue-bounding logic is unchanged from the full pass and remains correct.

[brycebaril]

LGTM though cpp-lint says this line is too long:

agents/statsd/src/statsd_agent.cc:769