Latency Principles

This skill provides a systematic approach to understanding and reducing latency in software systems, based on the book "Latency" by Pekka Enberg.

Core Concepts

Latency is the time delay between a cause and its observed effect. It is a distribution, not a single number.

Quick Reference

• Little's Law: Concurrency = Throughput * Latency. Use to size queues and thread pools.
• Amdahl's Law: Speedup is limited by the serial part of the task. Use to estimate max benefit of parallelization.
• Tail Latency: The experience of the 99th percentile users. In high-fanout systems, tail latency dominates.

For detailed definitions and laws, see references/principles.md.

Diagnosis & Optimization

When facing latency issues, follow this systematic approach:

1. Measure First: Identify where the time is going. Don't guess.
2. Check the Usual Suspects: Use the checklist to rule out common issues.
   • See references/diagnostic_checklist.md for a comprehensive list covering Hardware, OS, and Architecture.
3. Model the System:
   • Is it a throughput problem (queuing)? -> Apply Little's Law.
   • Is it a serial execution problem? -> Apply Amdahl's Law.
   • Is it a tail latency problem (variability)? -> Check fanout and shared resource contention.

Common Sources of Latency

• Physics: Distance (Speed of light).
• Hardware: CPU frequency scaling, cache misses, NUMA.
• OS: Context switching, interrupts, scheduler latency.
• Runtime: Garbage collection pauses, JIT compilation.

When to use Reference Files

• references/principles.md: When you need deep theoretical background, definitions of laws, or understanding of latency compounding.
• references/diagnostic_checklist.md: When you are actively debugging a slow system and need a list of things to check.