Database Patterns

Reference knowledge for relational and NoSQL database engineering — schema design, query optimization, indexing, partitioning, replication, NoSQL modeling, monitoring, and backup.

When to Use

• Designing or evolving a schema; reviewing a migration PR
• Tuning a slow query (EXPLAIN, indexing)
• Modeling access patterns in MongoDB, DynamoDB, Cassandra, Redis
• Planning partitioning, sharding, or replication topology
• Configuring monitoring, slow-query logging, vacuum/bloat checks
• Designing backup, restore, RTO/RPO strategy

When NOT to Use

• ORM-layer modeling in app code → python-engineer / java-engineer (reference this skill)
• Data pipelines / warehousing → data-engineer
• Database infra provisioning → devops-engineer

SQL Database Engines

PostgreSQL

• timestamptz not timestamp; text not varchar; numeric(p,s) for money
• pg_stat_statements, pg_stat_user_tables, pg_stat_activity for diagnostics
• pgaudit for audit trail; pgcrypto for column-level encryption
• pg_repack for online bloat reorganization; REINDEX CONCURRENTLY for index bloat
• Streaming replication: async for read replicas, sync for zero data loss
• PITR via pg_basebackup + WAL archiving
• Failover: Patroni, pg_auto_failover, or cloud-managed

MySQL

• InnoDB default; check innodb_buffer_pool_size sizing (60-80% RAM)
• Use EXPLAIN FORMAT=JSON for execution plan; performance_schema for diagnostics
• Slow query log: long_query_time, log_queries_not_using_indexes
• Replication: row-based (RBR) default; GTID for failover; semi-sync for durability
• ProxySQL for read/write split and connection pooling
• Avoid utf8 (3-byte) — use utf8mb4 (4-byte) for full Unicode

SQL Server

• Clustered index = physical row order; pick narrow, increasing, unique key
• Statistics auto-updated; use UPDATE STATISTICS after bulk loads
• sys.dm_exec_query_stats, sys.dm_db_missing_index_details for tuning
• Always On Availability Groups for HA; log shipping for DR
• Use OPTION (RECOMPILE) sparingly — parameter sniffing is usually preferable
• TDE for encryption at rest; Always Encrypted for column-level

Schema Design

• Normalization: 3NF by default. Denormalize only with measured read-performance justification
• Primary keys: Surrogate (BIGSERIAL, UUID) for most tables. Natural keys only when immutable and unique
• Foreign keys: Always explicit ON DELETE/ON UPDATE. CASCADE for child lifecycle, RESTRICT for references, SET NULL for optional
• Naming: snake_case, plural tables (users, orders). FK: {table_singular}_id
• Constraints: NOT NULL by default. CHECK for value ranges. UNIQUE for business keys
• Data types: Most specific type for storage + correctness
• Audit columns: created_at timestamptz DEFAULT now(), updated_at timestamptz on every table

SQL Optimization

• EXPLAIN ANALYZE: Always use actual execution. Compare estimated vs actual rows — gaps = stale statistics
• Seq scan elimination: Add indexes. Check work_mem for hash joins and sorts
• Join order: Optimizer handles most cases. 10+ joins → use join_collapse_limit or explicit order
• CTEs: PG 12+ inlines by default. Use MATERIALIZED only for optimization fences
• Subquery vs JOIN: Prefer JOIN for correlated subqueries. EXISTS over IN for large subsets
• Batch ops: INSERT ... ON CONFLICT (upsert). COPY for bulk. Avoid row-by-row loops
• Locking: FOR UPDATE SKIP LOCKED for queues. NOWAIT for fast fail. Advisory locks for app coordination
• Statistics: ANALYZE after bulk loads. Increase default_statistics_target for skewed distributions
• No SELECT * in production code — specify columns explicitly

Indexing Strategies

• B-tree (default): Equality + range. Composite: equality columns first, then range
• Hash: Equality-only, faster than B-tree for = but no range
• GIN: Full-text search, JSONB @>, arrays. gin_trgm_ops for LIKE '%term%'
• GiST: Spatial (PostGIS), range types, nearest-neighbor
• BRIN: Large append-only tables with natural ordering (time-series). Tiny index size
• Partial indexes: WHERE active = true — index only filtered rows
• Covering indexes: INCLUDE (col) for index-only scans
• Expression indexes: CREATE INDEX ON t (lower(email)) for case-insensitive lookups
• Composite order: Most selective first for equality; leftmost prefix rule applies
• Maintenance: REINDEX CONCURRENTLY for bloat. Monitor pg_stat_user_indexes for unused indexes
• Justify cost: Every index must have a documented query pattern. Indexes slow writes and waste storage

Partitioning + Replication

Partitioning

• Range: Time-series (month/year), numeric ranges — most common
• List: Categorical (region, status, tenant_id) — good for multi-tenant
• Hash: Even distribution when no natural range/list. High-throughput writes
• Pruning: Queries must include partition key in WHERE for pruning
• Maintenance: Pre-create future partitions. Detach and archive old ones
• Constraints: PK and UNIQUE must include partition key

Replication

• Streaming replication (PostgreSQL): Async for read replicas, sync for zero data loss. Monitor replication lag
• Connection routing: Read queries to replicas, writes to primary. Use PgBouncer or application-level routing
• Failover: Automated via Patroni, pg_auto_failover, or cloud-managed. Test failover regularly
• Logical replication: For selective table replication, version upgrades, multi-master
• RTO/RPO: Define Recovery Time Objective and Recovery Point Objective. Design backup frequency and failover speed accordingly

NoSQL

MongoDB

• Schema design follows access patterns, not normalization. Embed for 1:1 and 1:few. Reference for 1:many and many:many
• Indexes: compound indexes follow ESR rule (Equality → Sort → Range). Use explain() with executionStats
• Aggregation pipeline for complex queries. Avoid $lookup on large collections — denormalize instead
• Sharding: choose shard key based on cardinality + write distribution + query isolation. Avoid monotonic keys (ObjectId) as shard key
• Write concern: {w: "majority"} for durability; j: true for journaled
• Replica sets: minimum 3 voting members; use hidden/delayed secondaries for analytics

Redis

• Data structures: Strings (cache), Hashes (objects), Lists (queues), Sets (tags), Sorted Sets (leaderboards), Streams (event log)
• Key naming: {service}:{entity}:{id}:{field} — consistent, scannable, eviction-friendly
• TTL on every key — never store data without expiration unless explicitly persistent
• Memory management: maxmemory-policy (allkeys-lru for cache, noeviction for queues). Monitor used_memory vs maxmemory
• Persistence: RDB snapshots + AOF for durability; AOF-only for stricter RPO
• Cluster mode for horizontal scaling; sentinel for HA with single primary

Cassandra / ScyllaDB

• Model by query, not entity. Each query pattern = one table. Denormalization expected
• Partition key = distribution, clustering key = sort within partition
• Avoid partitions >100MB. Time bucketing for growing partitions
• LOCAL_QUORUM for strong consistency, ONE for eventual
• Avoid secondary indexes on high-cardinality columns; use materialized views or denormalized tables
• Tombstones: gc_grace_seconds tuning critical for deletion-heavy workloads

DynamoDB

• Single-table design: PK + SK overloading. GSIs for alternative access patterns
• Avoid hot partitions — random suffix on PK for high-throughput items
• ProjectionExpression to minimize read cost. Batch for bulk ops. TTL for auto-expiration
• On-demand vs provisioned capacity — choose based on traffic predictability
• DAX for sub-millisecond read cache
• Streams for change data capture into Lambda / Kinesis

Time-Series + Graph

Time-Series

• TimescaleDB (PG extension): hypertables = automatic time-range partitioning + compression
• InfluxDB: tag/field separation, downsampling via continuous queries
• Use BRIN indexes on append-only time columns in PostgreSQL
• Retention policies: drop or compress partitions older than N days
• Pre-aggregate (continuous aggregates) for dashboard queries

Graph

• Neo4j: model relationships as first-class. Cypher queries for path/traversal
• Property graphs vs RDF/SPARQL — choose based on schema rigidity
• Index on node label + property for entry points
• Avoid unbounded variable-length paths in production queries

Monitoring + Capacity Planning

• Key metrics: QPS, latency (p50/p95/p99), active connections, replication lag, cache hit ratio, disk I/O, bloat
• PostgreSQL: pg_stat_statements (query analysis), pg_stat_user_tables (seq scans, dead tuples), pg_stat_activity (locks)
• Slow query log: Enable with threshold (100ms). Review top offenders weekly
• Connection pooling: PgBouncer or pgpool-II. Monitor saturation and wait time
• Vacuum: Tune autovacuum per table. Monitor n_dead_tup. VACUUM FULL only in maintenance windows
• Bloat: Monitor table/index bloat. pg_repack for online reorganization
• Capacity: Project growth from current rate + business plan. Plan disk + memory + connection headroom 6–12 months ahead

Backup + Recovery

• PostgreSQL: pg_basebackup + WAL archiving for PITR. pg_dump for logical backups
• MySQL: mysqldump (logical), xtrabackup (physical, hot), binlog for PITR
• Verify restores on schedule — untested backups are not backups
• Off-site copy: Backups must survive primary site failure
• Encryption at rest for backup files
• RTO/RPO documented and tested per database tier

Database Security

• Least privilege: App accounts get only required permissions on specific tables. Never use superuser for applications
• Row-Level Security (RLS): PostgreSQL RLS for multi-tenant isolation. Enable FORCE ROW LEVEL SECURITY on the table owner
• Encryption: TLS for connections (sslmode=verify-full). Encryption at rest via disk/tablespace. Column-level (pgcrypto) for PII
• Audit logging: pgaudit for DDL/DML audit trail. Log all privileged operations
• SQL injection: Parameterized queries only. Never concatenate user input. Review stored procedures for dynamic SQL
• No secrets in SQL files: Connection strings, passwords, keys → env vars or secret managers

When this applies

Integration

• Consumed by: db-engineer (primary), data-engineer, python-engineer / java-engineer (ORM review), solution-architect
• External references: PostgreSQL, MySQL, SQL Server, MongoDB, Redis, Cassandra, DynamoDB official docs

Source: alex-voloshin-dev/ai-assets — distributed by TomeVault.