Improve ClickHouse

Operation

Implement the requested change and verify its directly affected behavior.

Goal

Correct the selected ClickHouse ingestion, table or query behavior using its actual engine, keys and merge semantics.

Domain rules

• Choose engine/partition/ORDER BY from concrete access and lifecycle patterns; avoid high-cardinality partitions and broad FINAL as a routine correctness fix.
• Make ingestion idempotency explicit with stable source identity and supported dedup semantics; do not rely on a short dedup window for indefinite retries.
• Rewrite queries to prune early, bound joins/aggregations/results, and use projections/dictionaries/materialized views only with owned refresh/consistency semantics.
• Stage schema/view/backfill changes and throttle mutations or rebuilds; preserve compatibility for existing parts and distributed replicas.
• Set users, quotas, profiles, pools, and workload limits from aggregate deployment capacity and expose system-table signals.

Validation

• Run DDL/query/ingestion integration tests against the supported ClickHouse version/topology and inspect system tables/query logs.
• Exercise duplicate/retried batches, partial failure, late data, replica/distributed-queue disruption, merge pressure, mutation/backfill, and query concurrency.
• Measure scanned rows/bytes, memory/spill, parts, merge backlog, insert/query latency, and result correctness with representative data shape.

Pin the failing insert/query and its expected logical result before changing the table, query or workload setting. Preserve source identity across retry and stage physical-layout changes with their actual migration cost. Feed the candidate DDL/settings, fixture and query to the ingestion/load check. Compare pre/post-merge visibility when it matters, without forcing a production merge as a test. A failed result or budget returns to that owner; accepted evidence proceeds to Patch Checking.