Metrics reference

Engine server processes expose /metrics on observability.metrics_listen (:9550 by default) when metrics are enabled. The admin console exposes its own metrics on its HTTP listener. The tables below describe the application families; engine registries also include Go and process collectors.

Labels use closed enumerations; index names, query text, document identities, object keys, and node addresses are never labels. A registry that would break that rule is rejected at construction.

Two rules are enforced in code before a process serves anything.

Every label must be on the allowlist. The complete set:

role  lane  service  method  code  outcome  cache_tier  object_operation
exactness  cache  state  storage_mode  segment_format  strategy  event
source  operation  publication_operation  catalog_operation  gc_reason
lease_operation  cache_state  requested_mode  policy_reason  phase  reason
tier  ack_status  seal_trigger  mode  similarity  candidate_mode  tail_ordering

Certain substrings are forbidden outright, so that a refusal says why:

query  document  segment_id  tenant  generation  object_key  error_message
url  model  fingerprint  text_digest  field  index  node  address  endpoint
peer  worker  aggregator  unit

Each of those would publish one series per thing that exists rather than per value in an enum.

Every series carries role as a constant label. Its value is the role mix the process serves, not a single role: node --roles builder,publisher reports role="builder-publisher". Under the Kubernetes operator each Deployment runs one role, so values are plain role names there.

instance and pod are not emitted by the process and must come from the scrape target. See observability for the consequences.

These are the closed enumerations the tables below refer to.

The shared outcome label uses four values with a common mapping. A canceled context or a CANCELED/DEADLINE_EXCEEDED status is canceled; an explicit server rejection — INVALID_ARGUMENT, NOT_FOUND, FAILED_PRECONDITION, RESOURCE_EXHAUSTED, OUT_OF_RANGE, PERMISSION_DENIED, UNAUTHENTICATED, UNIMPLEMENTED — is rejected; anything else is error. What an operator branches on is whether the caller gave up, whether the server refused, or whether something broke. The error's own text says the rest, and is exactly what may not become a label.

Cache operations use their own outcome vocabulary rather than the shared one: hit, miss, admitted, evicted, quarantined, error, and — for the block tier — bypassed. The byte counters add restored for the full tier.

The custom collectors below expose object I/O totals and WavesDB cache resources.

These are constant histograms. Their sample count is the request total and their sample sum is transferred bytes — they have no meaningful bucket distribution. Read them as:

rate(yolosearch_storage_process_object_read_bytes_sum[5m])    # bytes per second
rate(yolosearch_storage_process_object_read_bytes_count[5m])  # requests per second

The counting wrapper encloses the complete object store, so the totals include ingest publication, catalog access, compaction, cache fills, and query reads — not only request-scoped search traffic.

Reads count Stat, FullGet, RangeGet, List, Exists, and Size; writes count Put, Create, CompareAndSwap, and Delete.

cache is block, table, or file; outcome is hit, miss, or eviction; state is resident or limit.

The block-entry family has only state="resident": the decoded-block cache is bounded by bytes, not by an independent entry count. Rebinding an already registered process owner is idempotent, so overlapping generation readers do not multiply either limits or reported occupancy.

The console process exposes six label-free gauges and nothing else. They describe the bridge, not any index.

Every data-plane registry also registers the Prometheus Go client's process and Go collectors, so process_* and go_* families are present with their usual meanings.

• Observability — scrape configuration and the shipped alerts
• Sizing — reading these as a capacity signal
• Glossary — the vocabulary the label values use