PostgresqlTooManyConnections

warningPostgreSQL (postgres_exporter)

PostgreSQL is using most of its max_connections slots, and new clients will soon get "sorry, too many clients already".
Severity warning
Source prometheus-community/postgres_exporter 0.15+
Key metrics pg_stat_activity_count (labels datname, state), pg_settings_max_connections

What it means

The exporter counts backends in pg_stat_activity and reads the max_connections setting. The alert fires when the total has stayed close to the limit for several minutes. Once the limit is hit, only superusers can use the few superuser_reserved_connections; every application login fails.

Each PostgreSQL connection is a full OS process with its own memory, so raising the limit is rarely the right first move.

Common causes

  • Connection leaks: code that opens connections and never returns them to the pool.
  • Many application replicas each with a generous pool size (after a scale-out, pools multiply).
  • Sessions stuck idle in transaction, holding a slot and often locks.
  • Slow queries or lock waits making every request hold its connection longer.
  • No pooler (PgBouncer, pgcat, RDS Proxy) in front of a large fleet.

First checks

  1. Break the connections down by database, user, application and state:
    SELECT datname, usename, application_name, state, count(*)
FROM pg_stat_activity
GROUP BY 1, 2, 3, 4
ORDER BY count(*) DESC
LIMIT 20;
  2. Compare with the limits:
    SHOW max_connections;
SHOW superuser_reserved_connections;
  3. Find long idle and idle-in-transaction sessions:
    SELECT pid, usename, client_addr, state, now() - state_change AS in_state_for
FROM pg_stat_activity
WHERE state IN ('idle', 'idle in transaction')
ORDER BY state_change
LIMIT 20;
  4. See when the growth started and which database drives it:
    sum by (instance, datname, state) (pg_stat_activity_count)

Fixing it

For immediate relief, terminate clearly abandoned sessions: SELECT pg_terminate_backend(<pid>); (target old idle in transaction sessions first). Then fix the source: lower per-replica pool sizes, set idle_in_transaction_session_timeout, and put a transaction-mode pooler in front of the database. Raise max_connections only with a matching memory budget; it requires a restart.

Related alerts

← all alert runbooks