Mail background database work

A folder check whose IMAP STATUS shows nothing new ends without a commit: it claims no resource fence and records no sync run. That needs a server with CONDSTORE, a folder whose import finished with no flag walk, reconciliation or search under way, a full reconciliation less than six hours old, and settled stored state: an active binding and connection without errors and an active resource and mailbox health. The resource's last-sync time still moves on, at most once a minute, so health and operator views keep reporting a recent check for quiet mailboxes. The binding's last-use time follows committed batches only.

Completed sync runs in operator counts and cld mail sync runs therefore show batches that changed something. A failed check still records a failed run. The Drafts folder, servers without CONDSTORE, a folder whose stored state needs recovery, and every folder a waiting sync command asked for still take a full batch with a completed run. IMAP push heartbeats and provider lease renewal keep their existing cadence.

Each Mail process logs Mail folder sync activity every ten minutes in which folder syncs ran. It counts sync requests by trigger (scheduled, push, reconcile, command, maintenance, mailbox, continuation, busy, rate_limited) and batches by outcome (committed, unchanged, deferred, failed). A coalesced job keeps the input of its first request, so requests are counted when they are made.

An unchanged folder cursor, folder sync status, mailbox health or connection error state no longer causes a row update. Binding-folder UID state is updated only when UID validity, next UID or highest modification sequence changes; its last_verified_at follows that state update. It is not the binding verification timestamp used for rediscovery. Changes, error recovery and completed reconciliation windows are still recorded immediately.

The mailbox and connection updatedAt timestamps now follow actual metadata changes. Idle checks no longer move a mailbox forward in lists ordered by its last update; sync freshness still comes from the resource's last-sync time.

Completed sync runs are retained for 30 days. The hourly runtime-history cleanup leaves running rows alone and removes expired terminal rows in 1,000-row batches, with at most 100 batches and a five-second budget per run. A backlog can outlast 30 days; this is an age policy, not a fixed table-size limit. No migration deletes history or reclaims disk space. Investigate a persistently growing backlog before changing retention or scheduling a separate database maintenance operation.

After deployment, compare interval deltas in pg_stat_user_tables.n_tup_upd for folders, binding_folder_refs, provider_connections and mailboxes in the mail schema. Compare them with completed mail.sync_runs counts for the same period and pg_stat_statements calls for the corresponding updates. Calls still occur to validate and lock current state, but unchanged rows are not rewritten. Cumulative production totals include older versions.