ContextQMD
Back to Omniroute

Admission lanes — two lane systems, what gates each, where each reports

docs/architecture/admission-lanes.md

3.8.505.3 KB
Original Source

Admission lanes (#9654) — two lane systems, what gates each, where each reports

OmniRoute has two process-local lane systems with different scopes. They are complementary; operators should know which one they are looking at.

1. Byte-level per-connection lanes (chatBodyAdmission.ts)

  • Scope: the buffered-body/heap path for POST /v1/chat/completions. Guards against heap amplification from large coding-agent bodies (#4380).
  • Gate: always on. Each distinct API key (hashed) — or anonymous — gets its own lane with CHAT_MAX_HEAVY_IN_FLIGHT capacity, so one session's burst cannot starve another session's heavyweight slot.
  • Tuning:
    • OMNIROUTE_CHAT_VIRTUAL_TTL_MS — idle-lane eviction (default 60000)
    • OMNIROUTE_CHAT_VIRTUAL_MAX_SESSIONS — lane count cap (default 64)
    • OMNIROUTE_CHAT_ADMISSION_QUEUE_MS — queue-wait before 503 (default 2000)
    • OMNIROUTE_CHAT_ADMISSION_MAX_QUEUED_BYTES — queued-bytes heap valve (default 4 MB)
  • Reports: not in GET /api/monitoring/health today; observable via PerConnectionAdmissionController.snapshot() (sessionId hash, activeHeavy, idleMs).

2. Adaptive runtime virtual lanes (open-sse/services/admission)

  • Scope: tenant-key admission for provider dispatch — queue cost, latency-guided limit adaptation, lane queueing, and lane metrics.
  • Gate: opt-in. Disabled unless OMNIROUTE_CHAT_VIRTUAL_LANES=true. Without it, the adaptive controller keeps the shared queue behavior (criterion 1 of #9654 only holds once an operator enables lanes).
  • Tuning: OMNIROUTE_CHAT_VIRTUAL_LANES + adaptive config (maxQueueCount, maxQueueCost, defaultMaxWaitMs, …).
  • Reports: GET /api/monitoring/health → adaptiveAdmission → laneCount, laneQueuedCount, laneQueuedCost, laneTenants (opaque lane IDs, never raw keys), and virtualLanes — the authoritative "lanes are on" flag in the snapshot.

3. Fan-out probes — per-target admission for combo/fusion (#9654 Wave 2)

Combo (priority / round-robin) and fusion fan out N model targets under one parent request. Since #9654 Wave 2, each fan-out target is gated before dispatch by a per-target probe (PerTargetAdmissionHook, built by createPerTargetAdmissionHook) against the parent's tenant lane.

  • Scope: every fan-out target dispatched by combo, fusion, and the chaos engine. System 1 (byte-level) is unaffected — it never probes fan-out targets.
  • Gate: opt-in with system 2. A no-op when OMNIROUTE_CHAT_VIRTUAL_LANES is unset — the parent request already holds the shared-queue lease in that mode, so probing would double-count and reject combo targets.
  • Semantics:
    • Strictly non-blocking — skip, never queue. maxWaitMs 0: a full lane skips the target and the combo's fallback machinery (or fusion's survivor panel) serves instead. This is deliberate: a fan-out target is redundant work, and queueing it piles more load onto the exact congestion lanes exist to stop. defaultMaxWaitMs therefore applies to the parent request only; fan-out probes never wait, and there is intentionally no knob to make them wait (issue history shows wait knobs produced the mass-502/504 class #9654 prevents — revisit only if an operator reports skipped fan-out targets hurting response quality).
    • Release-on-admit. An admitted probe releases its lease immediately: it is a capacity gate, not a hold. The parent's lease covers the fan-out; holding N more would inflate shared active cost and reject other tenants. Best-effort, not a reservation: the lane can refill between probe and dispatch, so under heavy contention the gate may admit into a lane that is full again by the time the target dispatches.
    • Priced from the real fan-out body. The probe estimates cost from the target's actual body — including the request class derived from its stream flag, exactly like the parent path — so fusion panel members (stream: false) are priced at the non-streaming class they will truly occupy, and priority/RR targets at whatever the user requested.
  • Reports: a probe skip after the first target bumps combo's per-request fallbackCount (mirroring the existing fallback semantics; visible in combo logs); fusion returns 503 when every panel member is skipped. There is no aggregate counter (e.g. virtualFanoutSkipped) on the snapshot today — if an operator reports they cannot tell how often the lane gate skips fan-out targets, that is the trigger to add one.

Which one is showing in a dashboard

  • adaptiveAdmission.laneCount / laneTenants → adaptive virtual lanes (system 2).
  • adaptiveAdmission.virtualLanes === true → the fan-out probes of section 3 are also active. A payload with virtualLanes missing or false means OMNIROUTE_CHAT_VIRTUAL_LANES is unset — the byte-level lanes (system 1) are still active, but nothing under adaptiveAdmission (and no fan-out gating) is in effect until it is enabled.

Why both exist

The byte-level lanes bound the memory-heavy parse/compress path; the adaptive lanes bound dispatch cost per tenant. #9654's criterion 1 ("one session's burst does not 503 another") is enforced by system 1 unconditionally and by system 2 once opt-in is enabled.