PHP From ZeroModern PHP course

deployment and operations

PHP-FPM Pool Tuning

FPM pool settings control worker count and lifecycle. Tuning starts with measured traffic, request latency, and process memory rather than copied numbers. A configuration that works for a small admin tool may be wrong for an API handling bursts of traffic.

Measure Before Changing Limits

  • Understand dynamic, static, and ondemand process management modes.
  • Estimate safe worker count from available memory and measured worker usage.
  • Monitor queueing, saturation, slow requests, and restarts.

Change One Variable At A Time

  • Measure baseline.
  • Change one setting in staging.
  • Load-test and observe.

Read The Symptoms

  • Too many workers can exhaust memory.
  • Too few workers increase queue latency.
  • Average memory can hide high-percentile spikes.

Capacity Estimate

safe workers ~= memory budget for PHP workers / measured high-percentile worker memory

Then verify with traffic, queue latency, saturation, and restart metrics.

Record the measured worker memory, host memory budget, expected traffic shape, and observed queue behaviour. A capacity estimate is a starting point for a staging test, not proof that production is safe.

Practice

Practice: Create An FPM Tuning Plan

An application starts queueing HTTP requests during traffic spikes. Write a tuning plan that gathers evidence before changing the pool limits.

Requirements

  • Understand dynamic, static, and ondemand process management modes.
  • Estimate safe worker count from available memory and measured worker usage.
  • Monitor queueing, saturation, slow requests, and restarts.
  • Measure baseline.
  • Change one setting in staging.
  • Load-test and observe.
Show solution

Measure request latency, FPM queueing, active workers, high-percentile worker memory, host memory, and restart behaviour. Estimate a safe worker budget from measured memory rather than average memory alone.

Change one pool setting in staging, repeat a representative load test, and compare the same signals. Increasing workers is not automatically an improvement: it can trade queue latency for host-wide memory pressure.