Retry & Backoff

When a handler throws, the event is retried with a configurable backoff strategy.

Retry Flow

Handler throws
    │
    ├─ retry_count++
    │
    ├─ retryCount < maxRetries?
    │     ├─ yes → status=PENDING, wait for backoff delay
    │     └─ no  → status=FAILED, store last_error
    │
    └─ Next polling cycle picks up the event
       (only if backoff delay has elapsed)

Backoff Strategies

Exponential (default)

The delay doubles on every attempt:

delay = min(maxDelay, initialDelay * 2^(retry_count - 1))

With defaults (initialDelay: 1000ms, maxRetries: 5):

Attempt	Delay
1	1s
2	2s
3	4s
4	8s
5	FAILED (attempt limit reached)

Fixed

The delay is constant:

delay = initialDelay (every attempt)

With initialDelay: 1000ms:

Attempt	Delay
1	1s
2	1s
3	1s
...	1s

Configuration

OutboxModule.forRoot({
  prisma: PrismaService,
  retry: {
    maxRetries: 5,         // default: 5
    backoff: 'exponential', // 'fixed' | 'exponential' (default: 'exponential')
    initialDelay: 1000,    // default: 1000ms
  },
})

Option	Type	Default	Description
maxRetries	number	5	Max delivery attempts before FAILED
backoff	'fixed' | 'exponential'	'exponential'	Backoff strategy
initialDelay	number	1000	Base delay in ms

TIP

The maxRetries value is stored per-record in the max_retries column at emit time. Configuration changes during rolling deployments do not affect in-flight events.

Persisted retry scheduling

Version 0.3 stores the next retry time in next_attempt_at using PostgreSQL's clock when a failure is committed. Every poller claims from that same due time; differing local backoff settings do not recompute already scheduled retries. OutboxRecord.nextAttemptAt exposes the schedule.

retry.maxDelay defaults to 86,400,000 ms and caps exponential delay. The maximum permitted bound is 2,147,483,647 ms. maxRetries remains a per-record delivery-attempt limit.

Lease recovery

Active callbacks renew their PostgreSQL claim lease. Recovery runs every tenth poll cycle and requeues expired claims without incrementing retry_count. Completion requires the original claim token and an unexpired lease; stale completions cannot overwrite newer claims.

OutboxModule.forRoot({
  prisma: PrismaService,
  lease: {
    duration: 300_000,
    heartbeatInterval: 60_000,
    heartbeatFailureTolerance: 1,
  },
});

stuckThreshold is a deprecated alias for lease.duration. A hung callback with healthy heartbeats is not automatically timed out; application deadlines and unhealthy-process termination remain necessary.

Manual Reprocessing

Use an authorized operator service or fixed tenant scope instead of raw SQL resetting counters:

const result = await tenantAdmin.retry(eventId);
switch (result.outcome) {
  case 'applied': break;
  case 'not_found': /* missing or outside this tenant */ break;
  case 'conflict': /* state does not allow retry */ break;
  case 'lost_claim': /* concurrent transition won */ break;
}

Only FAILED rows move to PENDING. Retry preserves the lifetime counter, clears error/completion fields, and makes the row due now. markFailed() accepts only PENDING; admin operations never cancel an active PROCESSING callback. See Admin setup.

Monitoring

Query failed events to detect systemic issues:

-- Count failed events by type
SELECT event_type, COUNT(*) AS failed_count, MAX(created_at) AS latest
FROM outbox_events
WHERE status = 'FAILED'
GROUP BY event_type
ORDER BY failed_count DESC;

-- Recent failures with error messages
SELECT id, event_type, retry_count, last_error, created_at
FROM outbox_events
WHERE status = 'FAILED'
ORDER BY created_at DESC
LIMIT 20;