database-architect（数据库架构师）

Database Architect Skill

Step 1: Understand Data Requirements

Gather requirements:

1. Entities: What data needs to be stored?
2. Relationships: How do entities relate (1:1, 1:N, N:M)?
3. Access Patterns: How will data be queried?
4. Volume: Expected data size and growth rate
5. Consistency: ACID requirements vs eventual consistency

Step 2: Design Schema

For Relational Databases:

1. Normalize: Start with 3NF to reduce redundancy
2. Define Primary Keys: Use surrogate keys (UUID/SERIAL) or natural keys
3. Define Foreign Keys: Establish referential integrity
4. Consider Denormalization: Only for proven performance needs

For NoSQL Databases:

1. Model for Queries: Design documents/collections around access patterns
2. Embed vs Reference: Embed for 1:1/1:few, reference for 1:many
3. Shard Key Selection: Choose keys that distribute evenly

Step 3: Plan Indexes

Index strategy based on query patterns:

-- Example: Users table with common queries
CREATE INDEX idx_users_email ON users(email);           -- Exact match
CREATE INDEX idx_users_name ON users(last_name, first_name);  -- Range/sort
CREATE INDEX idx_users_created ON users(created_at DESC);     -- Ordering

Index Guidelines:

• Index columns used in WHERE, JOIN, ORDER BY
• Consider composite indexes for multi-column queries
• Avoid over-indexing (slows writes)
• Use covering indexes for read-heavy queries

Step 4: Plan Migrations

Create versioned migrations:

migrations/
  001_create_users.sql
  002_add_email_index.sql
  003_create_orders.sql

Migration Best Practices:

• Always include up and down migrations
• Test migrations on production-like data
• Plan for zero-downtime migrations
• Backup before running migrations

Step 5: Optimize Queries

Analyze and improve slow queries:

1. Use EXPLAIN ANALYZE: Understand execution plans
2. Identify Table Scans: Replace with index scans
3. Optimize JOINs: Ensure indexes on join columns
4. Batch Operations: Use bulk inserts/updates
5. Connection Pooling: Reduce connection overhead

Step 6: PostgreSQL 17 Features (2024–2026)

Leverage PostgreSQL 17 capabilities where applicable:

Performance improvements:

• New VACUUM memory management — up to 20x lower memory footprint; vacuum now runs faster on busy systems
• Streaming I/O interface accelerates sequential scans on large datasets
• BRIN indexes support parallel builds
• B-tree indexes are more efficient for IN clause queries
• Optimized CTE (Common Table Expression) planning

SQL/JSON enhancements (PG 17):

• JSON_TABLE() — converts JSON data into relational table representation
• JSON constructors and identity functions (JSON(), JSON_SCALAR(), JSON_ARRAY(), JSON_OBJECT())
• Use jsonpath for expressive path-based queries over JSONB columns

Incremental backups:

• pg_basebackup supports incremental backup; combine with pg_upgrade for zero-data-loss major version upgrades

Logical replication improvements:

• Failover control for logical replication slots
• pg_createsubscriber creates logical replicas from physical standbys
• pg_upgrade now preserves logical replication slots across major version upgrades

Security:

• New MAINTAIN privilege — grants targeted maintenance rights without full superuser access
• sslnegotiation=direct client option for direct TLS handshake (avoids round-trip)

COPY improvements:

• COPY... ON_ERROR ignore — continues import on row-level errors instead of aborting

Step 7: pgvector for AI Embeddings

Store and query vector embeddings alongside relational data to avoid a separate vector database:

-- Install extension
CREATE EXTENSION IF NOT EXISTS vector;

-- Table with embedding column (1536 dims for OpenAI text-embedding-3-small)
CREATE TABLE documents (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    content TEXT NOT NULL,
    embedding vector(1536),
    created_at TIMESTAMPTZ DEFAULT NOW()
);

-- IVFFlat index for approximate nearest neighbor (ANN) search
-- lists = sqrt(row_count) is a good starting value
CREATE INDEX idx_documents_embedding ON documents
    USING ivfflat (embedding vector_cosine_ops) WITH (lists = 100);

-- HNSW index (faster queries, more memory; preferred for < 1M vectors)
CREATE INDEX idx_documents_embedding_hnsw ON documents
    USING hnsw (embedding vector_cosine_ops) WITH (m = 16, ef_construction = 64);

-- Similarity search (cosine distance)
SELECT id, content, 1 - (embedding <=> $1::vector) AS similarity
FROM documents
ORDER BY embedding <=> $1::vector
LIMIT 10;

When to use pgvector vs. dedicated vector DB:

• Up to ~10M vectors: pgvector is sufficient (sub-50ms queries with HNSW index)
• Above 10M vectors or requiring specialized ANN algorithms: consider Pinecone, Weaviate, or Qdrant
• pgvector advantage: same backups, replication, and connection pooling as the rest of PostgreSQL

Step 8: Table Partitioning Strategies

Use declarative partitioning for tables expected to exceed available RAM:

-- Range partitioning by date (common for time-series / logs)
CREATE TABLE events (
    id BIGSERIAL,
    created_at TIMESTAMPTZ NOT NULL,
    event_type TEXT NOT NULL,
    payload JSONB
) PARTITION BY RANGE (created_at);

-- Monthly partitions
CREATE TABLE events_2025_01 PARTITION OF events
    FOR VALUES FROM ('2025-01-01') TO ('2025-02-01');
CREATE TABLE events_2025_02 PARTITION OF events
    FOR VALUES FROM ('2025-02-01') TO ('2025-03-01');

-- Hash partitioning for even distribution (e.g., multi-tenant)
CREATE TABLE orders (
    id UUID NOT NULL,
    tenant_id UUID NOT NULL,
    total DECIMAL(12,2)
) PARTITION BY HASH (tenant_id);

CREATE TABLE orders_p0 PARTITION OF orders FOR VALUES WITH (modulus 4, remainder 0);
CREATE TABLE orders_p1 PARTITION OF orders FOR VALUES WITH (modulus 4, remainder 1);
CREATE TABLE orders_p2 PARTITION OF orders FOR VALUES WITH (modulus 4, remainder 2);
CREATE TABLE orders_p3 PARTITION OF orders FOR VALUES WITH (modulus 4, remainder 3);

Partition pruning: PostgreSQL automatically skips irrelevant partitions when the partition key appears in WHERE. Always include the partition key in queries.

Index on partitioned tables: Indexes created on the parent table are automatically created on all child partitions.

Step 9: JSONB Patterns at Scale

-- Generated columns promote hot JSONB fields to indexed native columns
CREATE TABLE customers (
    id BIGSERIAL PRIMARY KEY,
    data JSONB NOT NULL,
    -- Promote frequently filtered fields to B-tree indexed generated columns
    country TEXT GENERATED ALWAYS AS (data->>'country') STORED,
    signup_date DATE GENERATED ALWAYS AS ((data->>'signup_date')::DATE) STORED
);
CREATE INDEX idx_customers_country ON customers (country);
CREATE INDEX idx_customers_signup ON customers (signup_date);

-- GIN index for containment / key-existence queries
CREATE INDEX idx_customers_data_gin ON customers USING GIN (data);

-- Partial GIN index for large tables (index only active records)
CREATE INDEX idx_customers_data_active ON customers
    USING GIN (data) WHERE (data->>'status') = 'active';

-- jsonpath query example (PG 17)
SELECT * FROM customers
WHERE data @? '$.tags[*] ? (@ == "premium")';

Step 10: Connection Pooling

Use a connection pooler in front of PostgreSQL for all production deployments:

PgBouncer (lightweight, battle-tested):

# pgbouncer.ini
[databases]
mydb = host=127.0.0.1 port=5432 dbname=mydb

[pgbouncer]
pool_mode = transaction       ; transaction pooling for stateless apps
max_client_conn = 1000
default_pool_size = 25
server_pool_size = 5

Supavisor (cloud-native, multi-tenant, Elixir-based):

• Designed for serverless / edge functions with thousands of short-lived connections
• Supports both session and transaction pooling modes
• Used by default in Supabase deployments; available as self-hosted

Pooling modes:

</execution_process>

<best_practices>

1. Normalize First: Optimize later based on data
2. Index Thoughtfully: Based on actual query patterns
3. Use Migrations: Never modify schema directly
4. Monitor Performance: Use database profiling tools
5. Plan for Scale: Consider partitioning for large tables
6. Upgrade to PostgreSQL 17: Benefit from new VACUUM memory manager, SQL/JSON functions, and incremental backup support
7. Use pgvector for AI: Store embeddings in PostgreSQL with HNSW or IVFFlat indexes before reaching for a dedicated vector database
8. JSONB at Scale: Promote hot JSONB fields to generated columns with B-tree indexes; use GIN for containment queries
9. Always Pool Connections: Use PgBouncer (transaction mode) or Supavisor for all production PostgreSQL deployments
10. Partition Large Tables: Apply range (time-series), list (category), or hash (multi-tenant) partitioning for tables projected to exceed server RAM

</best_practices>

Design a schema for an e-commerce orders system

Example Response:

-- Users table
CREATE TABLE users (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    email VARCHAR(255) UNIQUE NOT NULL,
    created_at TIMESTAMP DEFAULT NOW()
);

-- Orders table
CREATE TABLE orders (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    user_id UUID REFERENCES users(id),
    status VARCHAR(20) DEFAULT 'pending',
    total_amount DECIMAL(10,2),
    created_at TIMESTAMP DEFAULT NOW()
);

-- Order items (1:N relationship)
CREATE TABLE order_items (
    id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    order_id UUID REFERENCES orders(id) ON DELETE CASCADE,
    product_id UUID NOT NULL,
    quantity INTEGER NOT NULL,
    unit_price DECIMAL(10,2)
);

-- Indexes for common queries
CREATE INDEX idx_orders_user ON orders(user_id);
CREATE INDEX idx_orders_status ON orders(status);
CREATE INDEX idx_order_items_order ON order_items(order_id);

</usage_example>

Rules

• Always justify denormalization with performance data
• Include rollback strategy for all migrations
• Document relationships and constraints

Related Workflow

This skill has a corresponding workflow for complex multi-agent scenarios:

• Workflow: .claude/workflows/database-architect-skill-workflow.md
• When to use workflow: For comprehensive database design including requirements analysis, schema design, query optimization, migration planning, and testing (multi-phase, multi-agent)
• When to use skill directly: For quick schema reviews or single-agent database tasks

Iron Laws

1. NEVER make schema changes without versioned migrations that include both UP and DOWN scripts — manual DDL in production is not recoverable.
2. ALWAYS normalize to at least 3NF before considering denormalization — never prematurely optimize without measured performance evidence.
3. ALWAYS plan indexes based on actual query patterns from EXPLAIN ANALYZE — never add indexes speculatively before profiling real workloads.
4. NEVER test or deploy a migration without running it against production-like data first — schema issues surface under realistic volume, not on empty tables.
5. ALWAYS use connection pooling (Supavisor or PgBouncer) in production — direct connections from serverless functions exhaust the database connection limit under load.

Anti-Patterns

Memory Protocol (MANDATORY)

Before starting:

cat .claude/context/memory/learnings.md

After completing:

• New pattern -> .claude/context/memory/learnings.md
• Issue found -> .claude/context/memory/issues.md
• Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: Your context may reset. If it's not in memory, it didn't happen.