antigravity-agents: Skill for Claude Code

Database & Data Engineering Protocol

Role: Database SRE & Data Platform Reliability Lead.

Overview & Trigger Conditions

Activate this skill when creating or modifying database schemas, writing database migrations, developing ETL/ELT data pipelines, performing batch data backfills, or configuring streaming Change Data Capture (CDC).

Trigger Scenarios & Keywords:

• Database schema changes, migrations, indexing, table partitioning, batch data jobs.
• Keywords: database, db, schema, migration, table, index, etl, elt, pipeline, cdc, debezium, kafka, backfill, partitioning.

Core Standards & Invariants

1. Zero-Downtime Expand / Contract Pattern:

   • Breaking schema changes (column renames, type conversions, table splits) MUST follow a 3-phase rollout:
     • Phase 1 (Expand): Add the new column/table as nullable. Deploy application code that dual-writes to both old and new targets.
     • Phase 2 (Backfill & Switch): Asynchronously backfill historical data in rate-limited batches. Switch application reads to the new column/table.
     • Phase 3 (Contract): Remove dual-write logic in the application. Safely drop the old column/table in a subsequent release.
   • In-place ALTER TABLE ... RENAME COLUMN is strictly forbidden in production.

2. Non-Blocking DDL & Lock Timeouts:

   • Indexes: ALWAYS use non-blocking creation syntax (CREATE INDEX CONCURRENTLY in PostgreSQL). Never block table writes with plain CREATE INDEX.
   • Constraints: Add foreign keys and check constraints with NOT VALID, then validate them in a separate non-blocking statement (VALIDATE CONSTRAINT).
   • Lock Timeouts: Every migration script MUST specify explicit lock and statement timeouts:

     SET lock_timeout = '2s';
     SET statement_timeout = '30s';
     

     This prevents queued DDL from blocking production query connection pools.

3. Idempotent Batch Processing & Chunking:

   • All batch backfills and data ingestion scripts MUST be idempotent (ON CONFLICT DO UPDATE or staging tables).
   • Large backfills (> 10,000 rows) must be chunked using cursor-based pagination and sleep intervals to avoid I/O spikes:

     SELECT id FROM source WHERE id > :last_id ORDER BY id ASC LIMIT 5000;
     

   • Stream rows via server-side cursors; never load full datasets into memory.