sendou.ink/docs/load-testing.md
2026-01-14 17:57:40 +02:00

7.2 KiB

Load Testing Guide

This guide explains how to profile and optimize slow routes in sendou.ink.

Prerequisites

Install autocannon globally:

npm install -g autocannon

Step 1: Identify Slow Routes

Data URLs vs Full Routes

When benchmarking a route like /leaderboards, the response includes server-side rendering (HTML generation). To isolate just the data loader performance, use React Router 7's data URLs:

# Full route (includes SSR)
autocannon -c 10 -d 10 http://localhost:301/leaderboards

# Data only (loader function only)
autocannon -c 10 -d 10 "http://localhost:301/leaderboards.data?_routes=features%2Fleaderboards%2Froutes%2Fleaderboards"

The data URL format is: {route}.data?_routes={encoded_route_path}

To discover the exact data URL for a route, open Chrome DevTools Network tab, navigate to the page, then filter by "Fetch/XHR" - you'll see the .data requests made during client-side navigation.

Running benchmarks

Use autocannon to benchmark endpoints:

# Basic benchmark (10 connections, 10 seconds)
autocannon -c 10 -d 10 http://localhost:301/leaderboards

# With more connections to simulate load
autocannon -c 50 -d 30 http://localhost:301/leaderboards

# Single request timing
curl -s -o /dev/null -w "Time: %{time_total}s\n" "http://localhost:301/leaderboards"

Reading autocannon output

┌─────────┬────────┬─────────┬─────────┬─────────┬────────────┬────────────┬─────────┐
│ Stat    │ 2.5%   │ 50%     │ 97.5%   │ 99%     │ Avg        │ Stdev      │ Max     │
├─────────┼────────┼─────────┼─────────┼─────────┼────────────┼────────────┼─────────┤
│ Latency │ 826 ms │ 1478 ms │ 7674 ms │ 8985 ms │ 1989.09 ms │ 1770.75 ms │ 8985 ms │
└─────────┴────────┴─────────┴─────────┴─────────┴────────────┴────────────┴─────────┘
  • 50% (median): Half of requests are faster than this
  • 97.5%/99%: Tail latency - worst case for most users
  • Avg: Average response time
  • Max: Slowest single request

Target: median < 100ms, 99th percentile < 500ms

Step 2: Profile SQL Queries

Create a profiling script to test queries directly against the prod database:

// profile-query.ts
import Database from "better-sqlite3";

const db = new Database("db-prod.sqlite3");

function timeQuery(name: string, fn: () => unknown) {
  const runs: number[] = [];
  let result: unknown;

  // Run 5 times for consistent measurement
  for (let i = 0; i < 5; i++) {
    const start = performance.now();
    result = fn();
    runs.push(performance.now() - start);
  }

  const avg = runs.reduce((a, b) => a + b, 0) / runs.length;
  const rowCount = Array.isArray(result) ? result.length : "N/A";
  console.log(`${name}: avg ${avg.toFixed(2)}ms (${rowCount} rows)`);
  return { avg, result };
}

// Example: Test a query
const stm = db.prepare(`
  SELECT * FROM "User" WHERE "id" = @id
`);

timeQuery("User lookup", () => stm.all({ id: 1 }));

db.close();

Run with:

npx tsx profile-query.ts

Step 3: Analyze Query Plans

Use EXPLAIN QUERY PLAN to understand how SQLite executes queries:

const plan = db.prepare(`
  EXPLAIN QUERY PLAN
  SELECT * FROM "User"
  JOIN "Skill" ON "Skill"."userId" = "User"."id"
  WHERE "Skill"."season" = @season
`).all({ season: 10 });

console.log(JSON.stringify(plan, null, 2));

Reading query plans

  • SCAN: Full table scan (slow for large tables)
  • SEARCH USING INDEX: Index lookup (fast)
  • USE TEMP B-TREE: Temporary sorting/grouping needed

Step 4: Check Indexes

// List indexes on a table
const indexes = db.prepare(`PRAGMA index_list("Skill")`).all();
console.log("Indexes:", JSON.stringify(indexes, null, 2));

// Get columns in an index
for (const index of indexes) {
  const info = db.prepare(`PRAGMA index_info("${index.name}")`).all();
  console.log(`${index.name}:`, JSON.stringify(info, null, 2));
}

// Table row counts
const count = db.prepare(`SELECT COUNT(*) as count FROM "Skill"`).get();
console.log(`Skill rows: ${count.count}`);

Common Optimizations

1. LEFT JOIN → INNER JOIN

If your WHERE clause filters on the joined table, LEFT JOIN is unnecessary and prevents index usage:

-- Slow: scans entire ReportedWeapon table
SELECT * FROM "ReportedWeapon"
LEFT JOIN "GroupMatch" ON ...
WHERE "GroupMatch"."createdAt" > @date

-- Fast: uses index on createdAt
SELECT * FROM "ReportedWeapon"
INNER JOIN "GroupMatch" ON ...
WHERE "GroupMatch"."createdAt" > @date

2. Add Missing Indexes

If query plan shows SCAN on a large table with a WHERE clause, add an index:

CREATE INDEX skill_season ON "Skill"("season");

3. Avoid ORDER BY in CTEs

-- Slow: sorts intermediate results
WITH q1 AS (
  SELECT ... ORDER BY count DESC
)
SELECT ... FROM q1

-- Fast: only sort final results
WITH q1 AS (
  SELECT ...
)
SELECT ... FROM q1 ORDER BY count DESC

4. Use HAVING Instead of Subquery Filter

-- Filter aggregates directly
SELECT "userId", count(*) as cnt
FROM "ReportedWeapon"
GROUP BY "userId"
HAVING count(*) >= 7

Step 5: Verify Fix

Always verify optimizations produce identical results:

const original = originalStm.all(params);
const optimized = optimizedStm.all(params);

console.log(`Original: ${original.length} rows`);
console.log(`Optimized: ${optimized.length} rows`);

// Compare results
const originalSet = new Set(original.map(r => JSON.stringify(r)));
const optimizedSet = new Set(optimized.map(r => JSON.stringify(r)));

const match = originalSet.size === optimizedSet.size &&
  [...originalSet].every(x => optimizedSet.has(x));

console.log(match ? "✓ Results match" : "✗ Results differ");

Step 6: Re-benchmark

After applying fixes, restart the dev server and re-run autocannon:

# Restart server to pick up changes
# (Ctrl+C and npm run dev)

# Re-benchmark
autocannon -c 10 -d 10 http://localhost:301/leaderboards

Example: Full Profiling Session

# 1. Benchmark the slow route
autocannon -c 10 -d 10 http://localhost:301/leaderboards

# 2. Create and run profiling script
cat > profile.ts << 'EOF'
import Database from "better-sqlite3";
const db = new Database("db-prod.sqlite3");

const start = performance.now();
const result = db.prepare(`YOUR_QUERY_HERE`).all({ /* params */ });
console.log(`Time: ${(performance.now() - start).toFixed(2)}ms`);
console.log(`Rows: ${result.length}`);

db.close();
EOF

npx tsx profile.ts

# 3. Clean up
rm profile.ts

Route File Locations

Route Code Location
/leaderboards app/features/leaderboards/
/q/* app/features/sendouq/
/to/* app/features/tournament/
/u/* app/features/user-page/
/builds/* app/features/builds/