Reliable Delivery

Audience: application developers and operators who need to know exactly what delivery guarantees EventMesh provides — ACK tracking, retries, the dead-letter queue, and what happens across crashes and restarts.

---

The contract: at-least-once

Every delivery is tracked by the runtime until the subscriber acknowledges it:

1. The runtime dispatches an event to a subscriber and records an in-flight delivery (persisted in a DeliveryStateStore, RocksDB by default).
2. The subscriber processes it, then ACKs (POST /events/ack with the deliveryId, or auto-ACK by returning from an SDK handler).
3. The ACK advances the subscriber's offset — the only thing that ever advances it.
4. No ACK within the timeout → the runtime retries with exponential backoff (+ jitter); after the attempt budget the event goes to the dead-letter queue (DLQ).

Because an unacknowledged event is re-delivered, each event is delivered at least once; subscribers must be idempotent (dedupe on event id, or use LOAD_BALANCE_STICKY-style ordering per key and business-level idempotency keys).

ACK semantics

• POST /events/ack body {"deliveryId": "..."} → 200 {"status":"acked"} or 404 for an unknown/already-acked id. (The Java SDK long-poll handler auto-ACKs on return; subscribeWithAck(topic, mode, Predicate) ACKs iff the predicate returns true.)
• The offset advances on ACK exactly once; double-ACK is a no-op.
• The backend broker is ACKed only after the subscriber ACKs (RocketMQ 5.x POP "deferred broker ACK"), so a crash between poll and ACK is covered by broker-side redelivery — not message loss.

Retries and dead-letter

Retry policy per delivery (defaults in ReliableDispatcher):

• Backoff: exponential from 1 s, capped at 16 s, ±20 % jitter to spread retry storms (DEFAULT_JITTER_RATIO).
• Attempt budget: 6 attempts (initial + 5 retries) by default.
• Terminal failure → the event is written to the DLQ ledger (DeadLetterStore, meta-backed when clustered) and the delivery retired.

Operators inspect and replay the DLQ through the admin plane: GET /admin/dlq/browse?topic=…&max=…, POST /admin/dlq/replay?topic=…&max=… (see Admin API).

Crash recovery — what restarts guarantee

On boot the runtime replays the persisted in-flight records (UniRuntime.alignPullOffsetsToAck + ReliableDispatcher.recover()):

• A record with a decodable event is re-dispatched to the subscriber's poll buffer with its attempt counter preserved — the client sees the event again (at-least-once), and the offset still only moves on a real ACK.
• A record that cannot be decoded (legacy/corrupt) is retired without advancing the offset; backend redelivery bounds it.
• Recovery is idempotent — records already live on this dispatcher are skipped, so calling it twice is safe.

The one thing recovery never does is pretend the absent client ACKed: that would convert an unacknowledged delivery into acknowledged progress and skip the event (the bug class fixed by issue #5379).

Multi-instance fencing

When several instances run with a meta store, each topic partition is owned by exactly one instance (PARTITION_OWNED_PULL). If ownership moves while a delivery is in flight, the old owner's ACK is fenced: it raises StaleOwnerException, writes nothing, and the broker redelivers to the new owner. Details: Control plane → delivery topology.

Delivery failure matrix (per backend)

Scenario	Behavior	Visible signal	Recovery
Local disk full (RocksDB offset/state write fails)	write returns false / throws	offsetWriteFailures counter; pendingDeliveries stops dropping	Free disk; deliveries stay in flight and retry
Meta unreachable during DLQ record	store throws; dispatcher keeps delivery in flight	WARN logs	Retry succeeds once Meta heals; records are idempotent
Push write failure (SSE/WS)	connection pump nacks the dispatcher	re-delivery counter	Event re-dispatched immediately
Process kill -9	WAL + persisted delivery state survive	—	Boot recovery re-dispatches in-flight records

Full per-store failure behavior: Control plane → state store failure matrix.

Where the code lives

Piece	Location
Dispatcher (ACK / retry / DLQ)	eventmesh-runtime/.../delivery/ReliableDispatcher.java
Delivery state (RocksDB)	eventmesh-runtime/.../state/RocksDBDeliveryStateStore.java
DLQ ledger (meta, clustered)	eventmesh-runtime/.../state/MetaBackedDeadLetterStore.java
Offset store	eventmesh-runtime/.../offset/RocksDBOffsetStore.java
Push transports (nack on failure)	eventmesh-runtime/.../push/ConnectionPushPump.java
Tests	RecoveryRedispatchTest, DeliveryRecoveryTest, CrossStoreFaultInjectionTest, AckTimeoutRedeliveryIntegrationTest