Database Clickhouse Weaviate

Goose tracks applied migrations automatically in the goose_db_version table (no manual registry).

Migration execution safety (agents)

Same rule as Postgres: do not run ch:* or ch:schema:dump unless the user explicitly asked in this conversation.

Commands (run from repo root):

# Apply all pending migrations
pnpm --filter @platform/db-clickhouse ch:up

# Roll back last migration
pnpm --filter @platform/db-clickhouse ch:down

# Show migration status
pnpm --filter @platform/db-clickhouse ch:status

# Create a new migration (creates timestamp-named files in both unclustered/ and clustered/)
pnpm --filter @platform/db-clickhouse ch:create <migration_name>

# Convert timestamp migrations to sequential order (run before merging a PR)
pnpm --filter @platform/db-clickhouse ch:fix

# Roll back ALL migrations (equivalent to drop)
pnpm --filter @platform/db-clickhouse ch:drop

# Reset ClickHouse volume and re-migrate (nuclear option)
pnpm --filter @platform/db-clickhouse ch:reset

# Seed sample span data
pnpm --filter @platform/db-clickhouse ch:seed

Creating migrations (hybrid versioning)

1. ch:create <name> — creates 20260305120000_name.sql in both unclustered/ and clustered/
2. Fill in both files (see rules below)
3. Before merging the PR, run ch:fix — renames timestamp files to the next sequential number (e.g. 00002_name.sql) and commits the renamed files

Migration file rules

• Each migration is a single .sql file with -- +goose Up and -- +goose Down sections
• Always include -- +goose NO TRANSACTION (ClickHouse does not support transactions)
• ClickHouse migration history is append-only in this repository. Do not edit existing Goose migration files; add a new migration in both unclustered/ and clustered/ instead.
• For additive changes to existing tables, prefer ordinary ALTER TABLE or additive projection migrations with sensible defaults unless the change truly requires a table rebuild.
• unclustered/: use standard table engines (e.g. ReplacingMergeTree)
• clustered/: add ON CLUSTER default and use Replicated* engines

Clustered migration reliability (replica lag / Code 517)

In clustered ClickHouse, replicas can temporarily lag DDL metadata propagation. A migration can fail with:

• code: 517
• Code: 517
• doesn't catchup with latest ALTER query updates

Use these authoring rules to reduce failures:

• Keep migrations idempotent (IF EXISTS / IF NOT EXISTS) so retries are safe.
• Prefer additive schema changes over destructive rewrites.
• Keep DDL batches small; avoid chaining many dependent ALTER statements in one migration.
• For tightly-coupled changes on the same table in replicated clusters, prefer one ALTER TABLE ... with multiple actions over multiple dependent ALTER statements.
• If statement B depends on metadata introduced by statement A, prefer splitting them into separate migration files.
• Avoid coupling view rebuilds and many base-table changes in one large migration when possible.
• Run one migration runner per environment (never concurrent ch:up against the same cluster).

Execution safety:

• packages/platform/db-clickhouse/clickhouse/scripts/up.sh retries transient replica lag errors from goose ... up.
• In clustered mode, migration sessions set alter_sync, distributed_ddl_task_timeout, and replication_wait_for_inactive_replica_timeout to improve DDL convergence.
• Retry tuning env vars:
  • CLICKHOUSE_MIGRATION_MAX_RETRIES (default 20)
  • CLICKHOUSE_MIGRATION_RETRY_DELAY_SECONDS (default 5)
  • CLICKHOUSE_MIGRATION_MAX_RETRY_DELAY_SECONDS (default 30)
• Clustered DDL tuning env vars:
  • CLICKHOUSE_MIGRATION_ALTER_SYNC (default 2)
  • CLICKHOUSE_MIGRATION_DISTRIBUTED_DDL_TASK_TIMEOUT_SECONDS (default 300)
  • CLICKHOUSE_MIGRATION_REPLICA_WAIT_TIMEOUT_SECONDS (default 300)

Weaviate collections and migrations

Use the dedicated Weaviate package for connection and schema bootstrapping:

• Connection API: packages/platform/db-weaviate/src/client.ts — createWeaviateClient() and createWeaviateClientEffect() connect and perform health checks. For the general platform pattern (Effect-first client, tagged errors, env, layer wiring), see architecture-boundaries — Platform adapters: Effect-based clients.
• Collection definitions: packages/platform/db-weaviate/src/collections.ts — define all collections in code via defineWeaviateCollections([...]).
• Migration logic: packages/platform/db-weaviate/src/migrations.ts — idempotent: checks collections.exists() before create and tolerates "already exists" race conditions.
• Manual migration command: pnpm --filter @platform/db-weaviate wv:migrate — entrypoint is packages/platform/db-weaviate/src/migrate.ts.

Rules

• Do not define Weaviate collections in app/domain packages.
• Do not add ad-hoc Weaviate migration scripts outside packages/platform/db-weaviate.
• Keep collection schema changes centralized in src/collections.ts and rely on the package migration flow.

Weaviate migrations (agents)

Do not run wv:migrate unless the user explicitly asked in this conversation.