sqlite-optimization（SQLite 优化）

SQLite Performance Optimization

Quick Start: Baseline Configuration

Set these PRAGMAs on every connection for most applications:

PRAGMA journal_mode = WAL;          -- Non-blocking concurrent reads/writes
PRAGMA synchronous = NORMAL;        -- Fast commits, safe for WAL mode
PRAGMA foreign_keys = ON;           -- Data integrity
PRAGMA cache_size = -64000;         -- 64MB cache (negative = KB)
PRAGMA temp_store = MEMORY;         -- In-memory temp tables/sorting
PRAGMA busy_timeout = 5000;         -- 5s retry on lock contention
PRAGMA mmap_size = 2147483648;      -- 2GB memory-mapped I/O (benchmark first)

Run periodically:

PRAGMA optimize;  -- Update query planner statistics

Optimization Workflow

1. Diagnose with EXPLAIN QUERY PLAN

EXPLAIN QUERY PLAN SELECT ...;

Key indicators:

• SCAN TABLE → Full table scan (O(N)) — needs index or query rewrite
• SEARCH TABLE USING INDEX → Index lookup (O(log N)) — good
• COVERING INDEX → Index-only scan, no table lookup — optimal
• USE TEMP B-TREE → Sorting not covered by index — consider index with ORDER BY columns

2. Schema Design

Use INTEGER PRIMARY KEY for rowid alias:

CREATE TABLE users (
    id INTEGER PRIMARY KEY,  -- Alias for rowid, no extra index
    email TEXT NOT NULL,
    created_at TEXT
);

Choose appropriate types:

• INTEGER for IDs, counters, booleans
• REAL for floats
• TEXT for strings
• Avoid large BLOBs in hot tables; use separate table or external files

WITHOUT ROWID for junction tables:

CREATE TABLE user_roles (
    user_id INTEGER,
    role_id INTEGER,
    PRIMARY KEY (user_id, role_id)
) WITHOUT ROWID;  -- Clustered index, no rowid overhead

3. Indexing Strategy

Create indexes for real query patterns:

-- For: WHERE user_id = ? AND created_at >= ? ORDER BY created_at DESC
CREATE INDEX idx_orders_user_created
ON orders (user_id, created_at DESC);

Covering indexes eliminate table lookups:

-- For: SELECT user_id, total FROM orders WHERE user_id = ?
CREATE INDEX idx_orders_cover
ON orders (user_id, created_at DESC, total);

Partial indexes for subsets:

CREATE INDEX idx_active_users
ON users (email) WHERE status = 'active';

Expression indexes:

CREATE INDEX idx_users_lower_email
ON users (lower(email));

Avoid indexing:

• Small tables (< hundreds of rows)
• Low-cardinality columns alone (booleans)
• Columns rarely in WHERE/JOIN/ORDER BY

4. Write Optimization

Always batch writes in transactions:

BEGIN TRANSACTION;
INSERT INTO logs (...) VALUES (...);
-- ... many more inserts
COMMIT;

Throughput jumps from ~50 inserts/sec (autocommit) to 50,000+ inserts/sec.

Use prepared statements:

stmt = conn.prepare("INSERT INTO logs (ts, level, msg) VALUES (?, ?, ?)")
for row in data:
    stmt.execute(row)

Optimal batch size: 1,000–10,000 rows balances throughput vs memory.

5. Concurrency Configuration

WAL mode enables concurrent readers + one writer:

• Readers don't block writers
• Writers don't block readers
• Uses .wal and .shm files alongside main DB

Handle contention:

PRAGMA busy_timeout = 5000;  -- Retry for 5 seconds before SQLITE_BUSY

Architecture pattern: Single writer thread/process + many reader connections.

6. Maintenance

VACUUM after large deletes:

VACUUM;  -- Rebuilds and defragments DB file

• Run during maintenance windows (blocking operation)
• Requires free disk space ≈ DB size

Auto-vacuum for incremental reclaim:

PRAGMA auto_vacuum = INCREMENTAL;
VACUUM;  -- Apply change
-- Then periodically:
PRAGMA incremental_vacuum;

Keep statistics fresh:

PRAGMA optimize;  -- Smart, selective ANALYZE
-- Or full:
ANALYZE;

Reference Files

• pragma-reference.md — Complete PRAGMA settings with tradeoffs
• indexing-guide.md — Advanced indexing patterns and diagnostics
• architecture.md — SQLite internals: B-trees, VDBE, Pager, concurrency mechanics

Common Anti-Patterns

Environment-Specific Notes

Mobile (iOS/Android):

• WAL + synchronous = NORMAL
• Background maintenance jobs for VACUUM
• Local storage only (not SD card)

Server/Desktop:

• Centralize writes in single thread if possible
• Set PRAGMAs immediately after connection open
• Consider connection pooling with writer serialization