Almost every request touches Postgres. Spaces, workflows, users, run history, and live UI sync all depend on it. Alarm on the database itself, not only the application pods.
The chart’s in-cluster Postgres is the postgres StatefulSet with a 10 GiB PVC named pgdata-postgres-0. Production installs should prefer a managed PostgreSQL 18 instance with wal_level=logical, as described in the Helm installation guide. Use that provider’s CPU, memory, storage, and connection alarms when you do; the thresholds below still apply.
Alarms to set
Alarm | When to fire | What it means |
CPU sustained | Above 80% for 10 minutes | The instance is too small, or queries are backing up. |
CPU spike | Above 95% for 2 minutes | A burst. Repeat spikes mean the instance needs more CPU. |
Memory | Less than about 10% free, or the pod holding near its limit, for 15 minutes | Postgres will start evicting cache, then get OOM-killed. |
Connections | Sustained above about 80 connections (in-cluster default | Clients are leaking or the pool is undersized. On a managed instance, alarm at about 80% of that instance’s limit. |
Disk | The Postgres volume above 60% full | WAL and data share this volume. A full disk stops writes and can take the whole product down. |
For the in-cluster PVC, scrape kubelet volume stats on pgdata-postgres-0 the same way as Blobstore disk. If you run a Postgres exporter, also alarm on pg_stat_activity connection count and on remaining disk from pg_database_size.
Managed Postgres extras worth turning on: free storage, replication slot lag, and “oldest inactive logical replication slot.” 3B’s Zero sync uses logical replication. A stuck slot retains WAL until it fills the volume.
Confirming from the cluster
In-cluster, the database runs in postgres-0:
kubectl -n 3b exec postgres-0 -- psql -U postgres -c "SELECT count(*) FROM pg_stat_activity;"
kubectl -n 3b exec postgres-0 -- psql -U postgres -c "SELECT slot_name, active, pg_wal_lsn_diff(pg_current_wal_lsn(), restart_lsn) AS retained_bytes FROM pg_replication_slots;"
kubectl -n 3b exec postgres-0 -- psql -U postgres -c "SELECT pg_database_size('3b'), pg_database_size('zero_internal');"
Do not put the Postgres password on the command line. The chart’s postgres container is already authenticated for these as the postgres superuser.
What to do
CPU or memory pressure on in-cluster Postgres: move to a larger node, or cut over to a managed instance. The in-cluster database is a singleton and does not scale out.
Connections: find idle sessions in
pg_stat_activity, then raise the instance size or the pool — do not raisemax_connectionsas the first move.Disk: expand
pgdata-postgres-0the same way as Blobstore disk, using StatefulSetpostgresandpostgres.storage.size. On the single-server install, grow the VM disk instead.