Glossary

Terms as Celeriant uses them. Where another database uses the same word differently, that is noted.

Aggregate : An independent event stream, addressed by org_id / aggregate_type_id / aggregate_id. Events within an aggregate are strictly ordered, with no gaps and no concurrent writers. The unit of ordering and of optimistic concurrency.

Aggregate version : The index of an aggregate's latest event batch. What you pass as expectedVersion to make a write conditional.

Event : An immutable, typed, ordered fact appended to an aggregate. Carries a payload (opaque bytes), an event type, timestamps, and sequence numbers.

Event batch : Events the server grouped together on write. Each batch has a 1-based index and shared metadata (client id, server timestamp). An aggregate's version is its latest batch index.

Event type : A (major, minor) pair. Major is for breaking schema changes, minor for backward-compatible ones. Tied to the schema registry.

Client sequence (ClientSeq) : A per-writer monotonic index for events on an aggregate. With idempotency enabled, the server uses it to reject duplicates. See Idempotent retries.

Optimistic concurrency (OCC) : Appending only if the aggregate is still at the version you read. On conflict the write is rejected whole. See Optimistic concurrency.

Consistency boundary : The set of aggregates a single write commits atomically. In Celeriant that can be several aggregates, as long as they share a shard. See Consistency boundaries.

Shard : A partition of the keyspace owned by one core. An aggregate's shard is id % shard_count (a modulo, not a hash), where id is the org, type, or aggregate id picked by the routing rule set at cluster init. Ordering and atomic multi-aggregate writes hold within a shard, so the routing rule is how you co-locate aggregates that must be written together.

Watch : A subscription to change notifications as writes land, per aggregate, type, or org. A watch tells you what changed; you then read the new events. See Watch and subscribe.

Leader / follower : The two nodes of a cluster. The leader takes writes and replicates to the follower; both fdatasync before the leader acknowledges. See Two-node cluster.

Lease : The right to be leader, arbitrated by an S3 conditional write. Failover is a lease handoff. See Leader election and S3 leases.

Audit chain : A cryptographic hash (BLAKE3) linking each event to its predecessor, making tampering detectable. See The audit chain.

Projection / read model : A queryable view built by folding the event log, kept in a read store like Postgres. The read side of CQRS. See Building a read model.