feat(observability): BackendT2Degraded alert for silent tier-2 offload degradation#125
Open
EdHasNoLife wants to merge 1 commit into
Open
feat(observability): BackendT2Degraded alert for silent tier-2 offload degradation#125EdHasNoLife wants to merge 1 commit into
EdHasNoLife wants to merge 1 commit into
Conversation
Codex reviewFiles reviewed
FindingsBlockingNone. Should-fixdocs/observability/alerts.md:779 — The runbook says the Nitdocs/observability/alerts.md:732 — The section opens with “A managed CacheBackend’s tier-2...” even though the alert is documented as broader and applicable to Per-category coverage
Verdictchanges-requested |
…d degradation Per-CacheBackend Prometheus alert that fires when a backend is actively serving tier-2 (external offload) queries yet recalling zero reloads — the silent degradation the 32B/TP=2 dig surfaced (PYTHONHASHSEED, server OOM, version skew). Per-CacheBackend companion to the per-pod LMCacheT2NoHits (no scrape-scoping caveat) and the Prometheus view of the advisory T2Degraded CR condition. - BackendT2Degraded in both alerting-rules.yaml (flat) and the PrometheusRule CR; warning, for: 5m. Activity-gated: rate(inferencecache_backend_t2_query_tokens_total[10m]) > 1000 tokens/sec AND max by (namespace, backend) (inferencecache_backend_t2_hit_rate) == 0 - promtool tests: fires (+ pending-before-dwell), not-on-serving, not-on-idle, below-floor, per-install isolation - runbook section in docs/observability/alerts.md; enablement docs note the controller scrape both controller-side alerts need - correct the now-stale LMCacheT2NoHits note (the engine-metrics scraper reads vLLM's external_prefix_cache_* counters now) Re-lands the alert onto main; the original PR merged into the stacked base branch rather than main, so it never reached main.
EdHasNoLife
force-pushed
the
cac-192-backendt2-alert
branch
from
June 29, 2026 02:43
71ecf68 to
5bff981
Compare
Codex reviewFiles reviewed
FindingsBlockingNone. Should-fixNone. NitNone. Per-category coverage
Verdictapprove |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Per-CacheBackend Prometheus alert for silently-degraded tier-2 (external offload). Fires when a backend is actively serving tier-2 queries (
rate(inferencecache_backend_t2_query_tokens_total[10m]) > 1000tokens/sec) yet recalling zero reloads (inferencecache_backend_t2_hit_rate == 0). The query-rate clause is the activity gate — an idle/cold backend has no query growth (rate ≈ 0), so it doesn't page. Per-backend companion to the per-podLMCacheT2NoHits; the Prometheus view of the advisoryT2Degradedcondition (broader — also covers External backends, whose condition the controller doesn't manage).BackendT2Degradedin both the flatalerting-rules.yamland thePrometheusRuleCR (drift-gated) — warning,for: 5mdocs/observability/alerts.mdLMCacheT2NoHitsnote (the engine-metrics scraper reads vLLM'sexternal_prefix_cache_*now)make verify-prometheusgreen (drift-synced, kustomize clean). Consumesinferencecache_backend_t2_hit_rate+inferencecache_backend_t2_query_tokens_total, both already onmain.Note: re-lands the alert onto
main— #120 merged into the stacked base branch (cac-192-t2-alerting) rather than main, so the alert never reached main. Content matches #120 plus the alert-scope reword + lifetime-cumulative caveat.