ClickHouseSinkSchema

The tables and views of one aggregate type, in database, named after the type (<ctx>_<agg>, prefixOf):

• <ctx>_<agg>_state_events, the base table the Sink inserts into: a ReplacingMergeTree partitioned by the month of commit_time, ordered by (tenant_id, aggregate_id, version), one row per commit (ClickHouseRows); with a TTL on commit_time when AggregateSinkOptions.historyTtl is set.

• <ctx>_<agg>_state, a ReplacingMergeTree(version) ordered by (tenant_id, aggregate_id), the current state: filled by the materialized view <ctx>_<agg>_state_mv from the base table.

• <ctx>_<agg>_events, a ReplacingMergeTree ordered by (tenant_id, aggregate_id, version, sequence), one row per event: filled by <ctx>_<agg>_events_mv, which ARRAY JOINs the base table's events; under the same TTL.

• <ctx>_<agg>_state_flat, a view of _state FINAL: the system columns and the state schema's top-level scalars as typed columns (FlatColumns); its definition's flatFingerprint and those of the writers before it are kept in its comment (FlatLineage), so that during a rolling upgrade an older writer leaves a newer writer's view as it is.

Without the history (AggregateSinkOptions.history false) only _state and _state_flat exist, and the Sink inserts into _state directly (target).

Every table remembers the last deduplicationWindow insert blocks (non_replicated_deduplication_window), so an insert retried with the same token is dropped from the base table and, the inserts setting deduplicate_blocks_in_dependent_materialized_views, from the views' tables too. Beyond that window (a replay from an earlier checkpoint, an old owner beside the new one, a batch made up differently), the history tables' ReplacingMergeTree engine, keyed by ref@version (and the event's sequence), merges a replayed row away: a read with FINAL counts every commit and event once.