ovstorage-operator-debug-broker

Use when triaging a misbehaving ovstorage-broker - covers startup failures, authz denies, listener authn mismatches, streaming write aborts, reload misfires, and drain hangs.

SKILL.md

GitHub

Debug the Broker

Goal

Triage a misbehaving broker using the typed errors, metric labels, and tracing fields the daemon emits. The broker fails closed by design; most live-incident classes surface as a typed error before any backend dispatch.

Recipe

  1. Read docs/public/broker-operator/README.md § Debug runbook for the structured walkthrough.
  2. For startup failures, run cargo run -p ovstorage-cli -- list-routes --config /etc/ovstorage/broker.toml against the broker's config. The same TOML validator the daemon runs at startup will report bad routes, unresolvable secret refs, and invalid listener authn config without binding any sockets.
  3. For authz denies, look at the tracing output for broker_authz_decisions_total{outcome="deny"}. The deny path threads the plugin's reason and explanation into the gRPC PermissionDenied message; the explanation is typically a stable rule id from the built-in policy evaluator.
  4. For PolicyEpochStale errors, the request was stamped with a pre-reload epoch. Nothing in the SDK refreshes and retries it: it reaches the caller as FailedPrecondition (HTTP 409), which the caller re-issues itself against the reloaded policy.
  5. For listener authn failures (Unauthenticated), confirm the mode matches what your callers send:
    • jwt_verify mode: bearer JWT against the configured issuer / audience / jwks_url. JWKS is cached 5 minutes; an unknown kid triggers one refetch before Unauthenticated.
    • peer_cred mode: UDS / npipe only. Invalid on TCP listeners.
    • trusted_* modes: only on trusted_proxy = true listeners with trusted_peers CIDR allow-list.
    • mtls: reserved; startup fails with Unsupported if configured.
  6. For "broker is up but no routes are visible," check [[connections]] blocks: backends register at startup; a missing or misnamed plugin manifest fails startup. Confirm OVSTORAGE_PLUGIN_DIR is set and readable. Remember that list-routes shows only roots marked visible: the broker carries a hidden root to clients marked hidden, and the client keeps it routable without listing it. A suppressed root is absent from the wire and a direct caller receives NoRoute, as does an address whose scheme is not in anonymous_fallback_schemes when no connection serves it. A granted scheme reaches clients as a hidden s3:/// root, and a brokered client honours it only when its own [ovstorage.layers.broker] table grants the scheme too.
  7. For streaming-write issues (slow, corrupted, aborted), distinguish:
    • Graceful EOF (Ok(None) from the inbound stream) — broker commits with the chunks received.
    • RST_STREAM(CANCEL) (Err(status) with status.code() == Cancelled) — broker aborts with Status::cancelled. Look for broker.write spans ending in Cancelled to identify these.
    • Inline upload is bounded by WRITE_BODY_BYTE_CAP (64 MiB) after authz. Larger writes need the redirect branch (S3 multipart, GCS resumable, etc.); the plugin must mint redirects via WriteStep::Redirects.
  8. For lifecycle issues:
    • broker_lifecycle_events_total{event="reload_failed"} increments on validation failure; old broker stays live.
    • broker_policy_epoch_advances_total increments on successful reload.
    • Drain hangs: tune drain_timeout (default DEFAULT_DRAIN_TIMEOUT). A wedged plugin / upstream pushes the broker past the deadline and forces truncated in-flight RPCs.

Common live-incident classes

  • Cert expired. TCP+TLS handshake fails; rotate via SIGHUP (Unix) or process restart (Windows). Cert hot-reload is not implemented.
  • JWKS unreachable. Unauthenticated for every jwt_verify request. Check the IdP's reachability from the broker host.
  • Authorization latency. The built-in policy evaluator is local and synchronous; rising RPC latency points to listener authentication, storage dispatch, or an outer timeout rather than a remote policy plugin. Watch histogram_quantile(0.99, rate(broker_rpc_seconds_bucket[5m])) trending up — though note the RPC-latency histogram is dormant today (registered but not observed), so this query may return empty.
  • State root I/O error. Policy-epoch persistence fails; broker_lifecycle_events_total{event="reload_failed"} increments and the broker keeps the previous epoch.
  • Watch fan-out exhausted. The 257th concurrent watcher for one principal (summed across every watch key on the connection) receives ResourceExhausted (the per-principal cap is 256).

Pre-deploy validation

Always run cargo run -p ovstorage-cli -- list-routes --config <broker-config> against a config change before SIGHUP or restart. The same validator the daemon runs at startup will surface typed errors without binding any socket.

What's not debuggable today

  • No explain-decision <audit-id> tool. Trace the audit_id through your log aggregator.
  • No durable audit sink; rely on tracing output.
  • route.id and backend.id are not stamped on per-RPC spans yet — closing that gap is a tracked work item.

References