Metrics

Hub emits OpenTelemetry metrics from both hub-api and hub-worker. They are exported only when a metrics exporter is configured (see OTEL_METRICS_EXPORTER); with none set, metric recording is a no-op. Standard HTTP server metrics (request duration, counts) are also emitted for the API via OpenTelemetry’s net/http instrumentation — the tables below cover Hub’s own instruments.

Label cardinality is deliberately bounded: label values come from fixed sets, and an unexpected value collapses to other rather than opening a new series.

Enrichment pipelines

Embeddings, translation, sentiment, and emotions each emit the same five instruments under a per-type prefix — {type} is one of embedding, translation, sentiment, emotions:

Metric	Type	Labels	Description
hub_{type}_jobs_enqueued_total	counter	—	Jobs enqueued by the enqueue provider.
hub_{type}_provider_errors_total	counter	reason	Enqueue-path failures (for example enqueue_failed, settings_read_failed).
hub_{type}_outcomes_total	counter	status	Worker outcomes: success, retry, failed_final, skipped.
hub_{type}_worker_errors_total	counter	reason	Worker-side errors (for example {type}_api_failed, get_record_failed, rate_limited, superseded, tenant_write_conflict). input_over_cap counts translations refused before any provider call because value_text is over TRANSLATION_MAX_INPUT_BYTES; they are not {type}_api_failed.
hub_{type}_duration_seconds	histogram	status	Time to process one job.

Enrichment-wide instruments:

Metric	Type	Labels	Description
hub_provider_panics_total	counter	event_type	Recovered panics in an event fan-out provider. Expected to stay flat; a rising value means a provider is repeatedly crashing.
hub_enrichment_outputs_cleared_total	counter	output	Enrichment outputs (sentiment, emotions, translation) nulled by an edit’s eager-clear. A high rate with no matching re-enrichment points at backfill-recovery cases.
hub_enrichment_text_trimmed_total	counter	enrichment	Provider calls whose input was cut to its type’s configured cap (*_MAX_INPUT_BYTES), counted once per attempt. The call still runs — trimming is not a failure — so a rising rate means the cap is too low for this deployment’s records or its model. Translation never appears here: it refuses over-cap input instead (hub_translation_worker_errors_total{reason="input_over_cap"}).

Semantic search

Metric	Type	Labels	Description
hub_hnsw_iterative_scan_degraded	gauge (0/1)	—	1 when HNSW iterative scan is latched off (pgvector < 0.8 fallback; recall capped at ef_search until restart), else 0.

Webhooks

Metric	Type	Labels	Description
hub_webhook_jobs_enqueued_total	counter	—	Webhook delivery jobs enqueued.
hub_webhook_provider_errors_total	counter	reason	Enqueue-path failures.
hub_webhook_deliveries_total	counter	status	Delivery outcomes: success, retry, failed_final.
hub_webhook_disabled_total	counter	reason	Endpoints auto-disabled (410_gone, max_attempts).
hub_webhook_dispatch_errors_total	counter	reason	Dispatch-time errors.
hub_webhook_delivery_duration_seconds	histogram	status	Outbound webhook request duration.

Event publisher & job queues

Metric	Type	Labels	Description
hub_events_discarded_total	counter	—	Events dropped because the in-memory fan-out channel was full. Should be ~0; a rise means the publisher can’t keep up.
hub_message_publisher_fan_out_duration_seconds	histogram	—	Time to fan one event out to all providers.
hub_event_channel_depth	gauge	—	Current depth of the event fan-out channel.
hub_river_queue_depth	gauge	queue	Waiting River jobs per queue (available / retryable / scheduled).
hub_river_queue_oldest_age_seconds	gauge	queue	Age in seconds of the oldest waiting job per queue (0 when empty) — drives lag SLOs where depth alone cannot.

Caches

Metric	Type	Labels	Description
hub_cache_hits_total	counter	cache	Cache hits. cache is search_query_embedding or tenant_settings.
hub_cache_misses_total	counter	cache	Cache misses. Hit ratio = rate(hits) / (rate(hits) + rate(misses)) per cache.

Histogram buckets

All *_duration_seconds histograms share explicit second buckets up to 60s (…, 5, 7.5, 10, 30, 60), covering the 30s enrichment/LLM job timeouts so a final-attempt latency lands in a real bucket rather than +Inf.