Aller au contenu
actor-ts
Français

Replicated snapshots

Ce contenu n’est pas encore disponible dans votre langue.

In single-writer event sourcing, snapshots save the state at a particular seqNr — recovery loads the snapshot, replays events after that seqNr.

In replicated event sourcing, the picture is more complex — there’s no single linear seqNr; instead, events are partially ordered across replicas via vector clocks. Snapshots must carry the vector clock alongside the state.

Snapshot contents:
  - state         (after applying all events causally seen at the time)
  - vector clock  ({ A: 100, B: 95, C: 60 })

On recovery:

1. Load snapshot.
2. Read events from journal AFTER the snapshot's vc.
3. Apply each (re-running conflict resolution if any are concurrent).
4. Ready.

The vc lets the recovering replica skip events that causally precede the snapshot (already incorporated).

When to snapshot

Section titled “When to snapshot”

Same picking heuristics as single-writer ES, but bias toward more frequent:

  • Replicated workloads accumulate events from multiple replicas at once.
  • The journal grows faster (N replicas × per-replica rate).
  • Recovery has to re-run conflict resolution for each not-snapshotted concurrent event.
class Account extends ReplicatedEventSourcedActor<...> {
  override snapshotPolicy() {
    return everyNEvents(100);   // every 100 events
  }
}

For replicated entities accumulating 1000 events/day across all replicas, snapshot every 100 means at most 100 events to re-process on recovery — sub-second.

What gets serialized

Section titled “What gets serialized”

A ReplicatedSnapshot carries more than state + clock:

{
  state:                State,
  vc:                   { A: 100, B: 95, C: 60 },  // current vector-clock view
  seenIds:              [ ... ],  // dedupe set: (replica, seqAtReplica)
  events:               [ ... ],  // canonical sorted history — the heaviest field
  localSeq:             205,  // _localSeq at snapshot time
  journalSeqAtSnapshot: 4096,  // recovery resumes from here + 1
  takenBy:              'A',  // replica that took it (informational)
  takenAt:              1716297600000  // wall-clock ms (informational)
}

The snapshot is a normal snapshot blob persisted to the standard snapshot store; existing stores (in-memory, SQLite, object-storage) handle it without changes. Note it carries the full deduped events history and the seenIds dedupe set — not just state + clock — so an out-of-order remote arrival can refold without re-reading the whole journal. events is the heaviest field: a 100k-event actor snapshots 100k envelopes on disk.

Recovery flow

Section titled “Recovery flow”
preStart():
  ↓ load latest snapshot
  ↓     state = snapshot.state
  ↓     vc    = snapshot.vc
  ↓ read events from journal
  ↓     for each event in journal:
  ↓       if event.vc <= snapshot.vc: skip (already incorporated)
  ↓       else if event.vc concurrent with vc: invoke resolver
  ↓       else: apply via onEvent
  ↓ ready

The “skip” case is what makes snapshots bound recovery time — events written before the snapshot are skipped.

Snapshot stores

Section titled “Snapshot stores”

All standard snapshot stores work:

  • InMemorySnapshotStore — tests.
  • SqliteSnapshotStore — single-node (rare for replicated ES).
  • ObjectStorageSnapshotStore — shared across replicas.

For replicated ES with multiple replicas across regions, a shared snapshot store is critical — each replica recovers faster if it can load the latest snapshot from any replica, not just its own.

{
  const objectStorageSnapshotStoreOptions = ObjectStorageSnapshotStoreOptions.create().withBackend(new S3ObjectStorageBackend(S3ObjectStorageOptions.create() /* shared bucket */));
  journal:       sharedJournal,
  snapshotStore: new ObjectStorageSnapshotStore(objectStorageSnapshotStoreOptions),
}

Vector clocks grow with retired replicas

Section titled “Vector clocks grow with retired replicas”

Long-lived deployments accumulate retired replicas in vector clocks:

vc { A: 1000, B: 500, C: 200, RETIRED-D: 50, RETIRED-E: 30 }

The retired replicas’ components are inert but take space + slow comparisons.

Vector-clock garbage collection is out of scope for v1 — there is no pruning hook, and a snapshot stores the full, unpruned vector clock. VC entries grow with the set of replicas ever seen: fine for a stable cluster, but a node-churn-heavy deployment will eventually want compaction. Keep this in mind before relying on aggressive replica rotation.

Concurrent writes during snapshot

Section titled “Concurrent writes during snapshot”
Replica A writes event_A at t1.
Replica A snapshots at t2 (sees state with event_A).
   snapshot.vc = { A: 1 }


   Meanwhile, replica B was concurrently writing event_B at t1.5.
   event_B has vc { B: 1 }; not in snapshot.


   Replica A reads event_B at t3:
       snapshot.vc { A: 1 } vs event_B.vc { B: 1 } → concurrent
       invoke resolver, apply.

Concurrent events that arrive after the snapshot are handled at read-time by the resolver — same as without snapshots.

Performance

Section titled “Performance”

Snapshot writes for replicated ES are slightly heavier than single-writer:

  • Vector clock serialization — typically 50-200 bytes extra per snapshot.
  • Resolver state merging — if the snapshot is taken during concurrent-write reconciliation, the merge runs first.

Negligible in most cases. Bigger snapshots come from the state itself.

Where to next

Section titled “Where to next”