Author

Agent: 通信SDK马

Summary

Round-2/4 review ① per 通信龙 dispatch 2026-06-28 — highest-severity unowned item. server/src/push.ts had no per-client cap on the SSE ReadableStream queue. A half-open consumer (TCP drop, client crash without FIN) keeps its slot in clients AND keeps receiving controller.enqueue() calls from both the keepalive timer and every pushEvent. Bytes pile up in the stream's internal queue forever — clear OOM vector once exposed to the public internet (any commhub-server hosted at <hub-domain> reachable over WAN).

v1 → v2 (post-review fix)

v1 had a unit-mismatch bug caught by 通信牛: the default web-streams queuing strategy counts chunks, not bytes. So desiredSize < -1_000_000 meant "1 million chunks queued", not "1 MB queued" — a 1 MiB single enqueue only nudges desiredSize by -1 under the default strategy. The byte-cap was effectively defeated.

v2 fix: construct the SSE stream with an explicit byte-counting strategy so the byte-cap actually fires:

new ReadableStream<Uint8Array>(src, {
  highWaterMark: 0,
  size: (chunk) => chunk.byteLength,
})

通信牛's Bun probe (verified locally in the v2 test suite): default-strategy desiredSize goes 1 → 0 after a 1 MiB chunk and -1 after 2 MiB; byte-strategy goes 0 → -1048576 after 2 MiB.

What changed

server/src/push.ts

• Byte-counting queuing strategy on the SSE ReadableStream (the critical v2 fix).
• tryEnqueueBytes(client, bytes) — every enqueue path goes through this guard. Two thresholds:
  • Hard ceiling: desiredSize < -MAX_QUEUE_BACKPRESSURE_BYTES (1 MB default) → close stream immediately. One huge event can't blow up a stuck client.
  • Stuck timeout: desiredSize < 0 continuously for STUCK_CLOSE_MS (60s default) → close. Catches slow-trickle leaks.
• controller.desiredSize === null treated as no headroom → close. Defends against runtime errored stream.
• sweepLiveness() runs every LIVENESS_SWEEP_MS (15s default) so half-open detection doesn't require event traffic.
• closeClient(client, reason) idempotent, clears keepalive timer, best-effort close. All log lines include reason for ops triage.
• Initial connected frame goes through tryEnqueueBytes too.
• All thresholds env-overridable: ANET_SSE_MAX_QUEUE_BYTES, ANET_SSE_STUCK_CLOSE_MS, ANET_SSE_KEEPALIVE_MS, ANET_SSE_LIVENESS_SWEEP_MS.
• Liveness timer uses .unref() so it doesn't hold the event loop open.

Tests

$ bun test src/push.test.ts
 26 pass · 0 fail · 65 expect() calls

$ COMMHUB_DB=/tmp/anet-sse-test.db NODE_ENV=test bun test
 142 pass · 1 fail · 462 expect() calls

The 1 fail is the pre-existing missing @modelcontextprotocol/sdk dev dep — unrelated to this change, tracked in #261 / round-4 dependency hygiene.

26 tests:

• 21 fake-controller branches: tryEnqueueBytes (ceiling, stuck-set, stuck-recover, null-desired, enqueue-throw, already-closed), sweepLiveness (null/over/stuck/first-stuck/healthy/mixed), pushEvent integration (half-open/mixed/over-ceiling/unknown), getSSEStats (closed excluded).
• 3 real-ReadableStream (NEW in v2 — the assertion class that catches v1's unit-mismatch bug):
  • createSSEStream uses byte-counting strategy: enqueue 1 KB → desiredSize drops by ≥1024 (not by 1).
  • non-reading consumer past MAX bytes triggers hard ceiling close: pushes that overshoot 1 MB total bytes close the client. Under v1's strategy this test would NOT close (each push = 1 chunk).
  • reading consumer drains queue → no spurious close: drain promptly, no false-positive close.
• 2 pre-existing rekey tests untouched.

Constraint compliance (per 通信龙 dispatch)

• Open PR — not pushed to main directly
• 通信牛 review gate (no self-merge); v1 → v2 reflects review feedback
• Docker / isolated test (COMMHUB_DB=/tmp/anet-sse-test.db, NODE_ENV=test)
• No production deploy — code only
• Did not touch production hub at <hub-domain> or any prod DB
• All tunables env-overridable so ops can dial defaults without code change

Out of scope (queued — separate PRs)

• ② retention sweep + VACUUM + server-health index split (task [bug][P0] anet ls/import/run/server — async commands missing await, fa08eb4 race (preview.6 regression) #139)
• ③ /api/status read-path write-amp → background staleness timer (task [release][P0] v0.10.0 — Direct Runtime + Observability Foundations 规划 #140)

Test plan

• bun test src/push.test.ts → 26/26 green (includes 3 real-stream byte-cap probes)
• bun test baseline 142/143 (1 pre-existing fail unrelated, see 系统 review round 1 — 可靠性/安全/Runtime/测试·CI·文档 findings #261)
• Manual smoke (post-merge): spin a SSE consumer that reads slowly + one that disconnects without FIN; observe close-reason log lines + memory steady-state