Metrics

Athena exposes Prometheus metrics at GET /metrics on port 8080 (no auth required). Two naming families exist: cognitive-pipeline metrics are unprefixed; promoter, LTM, blob, and analytics metrics carry the memos_ prefix.

Pipeline (worker / STM path)

Metric	Type	Labels	Meaning
stm_cache_ops_total	counter	op, result	STM Redis operations; result="error" signals Redis trouble
embedding_latency_seconds	histogram		Embedding-creation latency (provider round trip)
milvus_op_latency_seconds	histogram		Milvus operation latency
cosine_similarity_distribution	histogram		Similarity scores from chain-break checks (0.0–1.0 buckets); shows where your traffic sits relative to the 0.52/0.72 gates
cosine_gate_decisions_total	counter	decision (high/low/gray_zone), result (continue/new_chain)	Chain-break gate outcomes; a high gray_zone share means high LLM arbitration cost
llm_fallback_calls_total	counter	reason (gray_zone/embedding_failure/rate_limited/circuit_breaker), result	LLM fallback calls; rate_limited/circuit_breaker reasons are your outage signal
dialogue_chain_decision_latency_seconds	histogram		End-to-end chain-break decision latency

Promoter and heat

Metric	Type	Meaning
memos_promoter_chains_evaluated_total	counter	Chains scored per promoter pass (cumulative)
memos_promoter_chains_promoted_total	counter	Chains that crossed the threshold; the ratio to evaluated is your promotion selectivity
memos_heat_score_distribution	histogram	Heat scores at evaluation time (0.1-wide buckets); watch before/after tuning any HEAT_* parameter

LTM writes (graph extraction)

Metric	Type	Meaning
memos_extractor_llm_duration_seconds	histogram	Triple-extraction LLM latency
memos_extractor_schema_failures_total	counter	LLM returned invalid JSON; promotion aborted for that chain
memos_ltm_arango_upsert_duration_seconds	histogram	AQL UPSERT latency
memos_ltm_nodes_written_total	counter	Nodes upserted
memos_ltm_edges_written_total	counter	Edges upserted
memos_ltm_edge_interceptions_total	counter	Rogue relations auto-corrected to RELATES_TO; a rising rate means the extractor is drifting from the whitelist

LTM reads (retrieval enrichment)

Metric	Type	Meaning
memos_ltm_fetch_duration_seconds	histogram	Graph traversal latency during retrieval
memos_ltm_nodes_read_total / memos_ltm_edges_read_total	counter	Traversal result volume
memos_ltm_fetch_errors_total	counter	AQL traversal errors
memos_ltm_fetch_no_results_total	counter	Entity start nodes existed but traversal returned nothing; high values suggest a sparse graph or over-strict confidence filtering

Blob storage

Metric	Type	Labels	Meaning
memos_blob_storage_ops_total	counter	operation (upload/download/delete), provider, status	Object-store operations
memos_blob_payload_bytes	histogram		Payload sizes; watch for events that should have been blobs

Analytics

Metric	Type	Meaning
memos_analytics_pregel_duration_seconds	histogram	Community-detection runtime
memos_analytics_bridge_calc_duration_seconds	histogram	Bridge-scoring runtime
memos_analytics_bridges_found_total	counter	Bridge entities identified per run

A starter alert set

- alert: AthenaRedisErrors
  expr: rate(stm_cache_ops_total{result="error"}[5m]) > 0
- alert: AthenaLLMDegraded
  expr: rate(llm_fallback_calls_total{reason=~"rate_limited|circuit_breaker"}[10m]) > 0.1
- alert: AthenaExtractionFailing
  expr: rate(memos_extractor_schema_failures_total[30m]) > 0.05
- alert: AthenaLTMTraversalErrors
  expr: rate(memos_ltm_fetch_errors_total[10m]) > 0

Standard Go runtime and process metrics (go_*, process_*) are exported alongside; scrape configuration is ordinary Prometheus (/metrics, no credentials).