The broker exports transaction coordinator state from the same in-memory state used to serve requests. Recommended production alerts are:
cursus_transaction_recovery_ready != 1 for any ready broker. The broker
should normally fail startup before this can occur.cursus_transaction_oldest_active_seconds above the application’s maximum
transaction duration. This detects stuck open or committing transactions.cursus_transactions{state="committing"}. A brief
non-zero value is normal while a commit is being applied.cursus_transactions_expired, paired with journal size
and filesystem alerts. The standalone journal rewrites atomically after 256
records or 16 MiB and retains the latest state per transactional ID.Alert thresholds are workload-specific. Compare age and counts with broker readiness, storage errors, Raft leadership, and consumer metadata recovery metrics before taking recovery action. Metrics and diagnostic commands are read-only and never create groups, transactions, topics, or offsets.