Metrics
Prometheus text format on /metrics, served by the headless worker (set TIERDB_METRICS_PORT) and by the console binary (on TIERDB_CONSOLE_PORT) identically.
Daemon
| Metric |
Meaning |
tierdb_leader |
1 on the leading worker, 0 on standbys (fleet sum should be exactly 1) |
tierdb_cycle_duration_seconds |
Duration of the last work cycle |
tierdb_last_cycle_timestamp_seconds |
Epoch time of the last completed cycle (staleness alert) |
Per table
| Metric |
Meaning |
tierdb_cutline_tier_key{table} |
The cut-line T |
tierdb_cutline_snapshot{table} |
The pinned lake snapshot S |
tierdb_delta_backlog_rows{table} |
Correction rows awaiting compaction (growth = fold blocked or under-provisioned) |
tierdb_lake_commits_total{table} |
Lake snapshots committed |
tierdb_mirror_lag_bytes{table} |
Mirrored: WAL bytes between the server position and the mirror frontier |
tierdb_mirror_flushes_total{table} |
Mirrored: pump flushes committed |
tierdb_load_total{table,state} |
Stream loads by outcome (committed, staged, replay, rejected, conflict) |
tierdb_load_rows_total{table,path} |
Stream-loaded rows by path (heap, delta, spool) |
tierdb_load_staged_labels{table} |
Staged loads awaiting adoption into Iceberg |
tierdb_load_adoption_lag_seconds{table} |
Age of the oldest staged load (growth = adoption blocked) |
Lake health
One gauge per counter the format plugin reports into tierdb.lake_stats, republished as tierdb_lake_<counter>{table}. The counter names are format-owned. Iceberg reports files, delete_files, bytes, records, snapshots, manifests, and delete_ratio. tierdb_lake_warnings{table} is the number of active health warnings, alert on it being nonzero.
Per replication slot
| Metric |
Meaning |
tierdb_slot_active{slot} |
Whether a consumer currently streams the slot |
tierdb_slot_retained_wal_bytes{slot} |
WAL pinned by the slot. Alert on growth, see the WAL guard |
Housekeeping
| Metric |
Meaning |
tierdb_expired_pins_deleted_total |
Expired read pins removed by the sweep |
Suggested alerts
sum(tierdb_leader) != 1: no leader, or two.time() - tierdb_last_cycle_timestamp_seconds > 120: the daemon stalled.tierdb_slot_retained_wal_bytes above your WAL budget: see the runbook.tierdb_delta_backlog_rows growing monotonically: compaction blocked.tierdb_mirror_lag_bytes sustained high: the pump can't keep up.