History Pruning

Cohort's ledger and hold tables need retention too. Pruning is opt-in. Each retention is independent, defaults to null (keep forever), and must be at least one day when set:

{
  "Cohort": {
    "HistoryPruning": {
      "SucceededRunRetention": "90.00:00:00",
      "FailedRunRetention": "365.00:00:00",
      "InactiveHoldRetention": "90.00:00:00"
    }
  }
}

A hosted pruner runs a pass when the host starts and then every hour.

Runs

A settled run is deleted, with its sweep_run_entity_summary, sweep_run_row_detail and sweep_row_handler_status rows, once its SettledAt is older than its retention:

• A run that settled Succeeded, with every handler delivery succeeded, uses SucceededRunRetention.
• A run that settled Failed, PartiallyFailed or Cancelled, or a succeeded run with dead-lettered handler work, uses FailedRunRetention. That lets you keep failure evidence longer, or shorter, than routine history.

A run is never deleted while any of its handler work is Pending or InFlight, however old it is. A run still Started is never deleted either; the dispatcher recovers abandoned runs as failed first.

Holds

A hold is deleted once the earlier of its RemovedAt and ExpiresAt is older than InactiveHoldRetention. Active holds are never deleted, including holds with no expiry.

ListActiveAsync and HasActiveHoldAsync with an asOf older than InactiveHoldRetention can therefore answer "not held" for a hold that has since been pruned.

How it runs

• Each batch of up to 100 runs, or 100 holds, commits on its own. Batches are claimed with FOR UPDATE SKIP LOCKED, so replicas prune different batches.
• A failed pass is logged and retried an hour later.
• KillSwitch stops pruning after the batch in progress.
• Pruning is otherwise independent of Schedule and DryRun.