PHP-FPM Reached pm.max_children can interrupt a specific task: PHP-FPM logs server reached pm.max_children and requests queue. This guide shows how to locate the actual failing component and choose a proportionate repair. It does not assume that one error message identifies a defective part or that reinstalling the whole platform is necessary.
Start here: Check active workers, slow request log and per-process memory footprint. A targeted comparison often provides better information than resetting every application, driver or service at once. Preserve your original settings so you can compare outcomes after each step.
Before you begin: preserve access and data
Safety boundary: Blindly increasing pm.max_children may exhaust shared host memory. If this device or service is managed by an employer, hosting provider or school, follow its approved change procedure. Back up unique content first whenever the proposed repair would affect stored records, firmware, encryption or account access.
Write down the exact product and release, the action that fails and the first time it happened. A useful first comparison is: Correlate traffic spikes with WordPress actions, PHP endpoints or blocked I/O. If possible, capture one passing case alongside the failure; those two results define what the fix must change.
Step-by-Step Troubleshooting
Begin with low-risk checks; continue to advanced diagnostics only if the evidence points to the affected component.
Reproduce the exact failure and preserve the context
Before changing settings, reproduce this specific symptom: PHP-FPM logs server reached pm.max_children and requests queue. Record the exact dialog, code or event text, product model and software version. Note which action triggers it and whether you can repeat it. Check active workers, slow request log and per-process memory footprint. Keep a screenshot with private data redacted; it will establish whether a later correction changed the original failure, not just a related symptom.
Run the lowest-risk comparison test
Correlate traffic spikes with WordPress actions, PHP endpoints or blocked I/O. Change only one variable between attempts: connection, user, input, application version, peripheral or storage target, whichever the check calls for. If the comparison succeeds, the difference is evidence about scope—not proof that every related system setting needs resetting. Record both results and the exact time of each test.
Apply the supported reversible correction
Reduce slow application work and tune caching before raising child count. Make the smallest documented change that addresses the observed failure. Keep the original setting or package version written down. If a restart is required, save active work first; avoid introducing multiple fixes at once because that makes a successful result impossible to attribute to a single cause.
Retest the original task and confirm stability
Queue depth drops and peak requests complete without OOM kills. Return to the same network, account, file or game where the problem first occurred. Repeat the same operation at least twice when practical, including one fresh launch or reconnect. If symptoms recur, restore the previous setting and move to the evidence-based checks below rather than applying unrelated tweaks.
Collect deeper evidence from the affected component
Calculate safe worker count from available RAM and use slowlog to trace handlers. Collect versioned logs, timestamps and the precise component identifier before attempting advanced modifications. Compare the event to the timeline recorded in Beginner checks. If you cannot reproduce the symptom reliably, gather several occurrences first to avoid treating harmless background log entries as the cause.
journalctl -u php-fpm --since "1 hour ago" --no-pagerDistinguish configuration errors from a component defect
Use both checks together: Check active workers, slow request log and per-process memory footprint; Correlate traffic spikes with WordPress actions, PHP endpoints or blocked I/O. A difference between a known-good account, cable, network, environment or image and the failing one helps distinguish configuration from hardware or service faults. Where diagnostics return a specific error, follow that component’s supported repair path, not a generic reset command.
Make a scoped advanced repair with rollback ready
Only if the evidence points to the identified subsystem, plan a targeted correction: Reduce slow application work and tune caching before raising child count. First save configuration, project files or business records that would be affected. Keep authentication and encryption protections enabled. Blindly increasing pm.max_children may exhaust shared host memory. For enterprise or production equipment, schedule the change and obtain the relevant administrator approval.
Verify recovery and prepare an escalation record
Complete this success test: Queue depth drops and peak requests complete without OOM kills. Monitor the same event, service, application or connection after the fix; confirm the outcome persists rather than merely observing that an error dialog disappeared. If unsuccessful, provide support with model, versions, exact symptoms, controlled comparisons, sanitized logs and rollback details.
Read the diagnostic results
Use the following evidence-to-action table instead of performing every possible repair. Each row suggests what the observed result actually establishes and which question to ask next.
| Observation | Evidence to collect | Recommended decision |
|---|---|---|
| Failure is reproducible | Check active workers, slow request log and per-process memory footprint | Keep exact code, version and triggering action; do not infer an unrelated cause. |
| Comparison isolates a difference | Correlate traffic spikes with WordPress actions, PHP endpoints or blocked I/O | Test only the component or setting that differs from the working case. |
| Advanced logs identify a subsystem | Calculate safe worker count from available RAM and use slowlog to trace handlers | Use the supported procedure for that subsystem and preserve a rollback. |
| After repair, original task succeeds | Queue depth drops and peak requests complete without OOM kills | Confirm persistence after reopening, reconnecting or restarting as appropriate. |
Why the obvious fix may not work
A seemingly successful workaround is not necessarily the underlying correction. For this issue, first determine whether php-fpm logs server reached pm.max_children and requests queue occurs on one account, one device, one connection or every attempt. The result of “Correlate traffic spikes with WordPress actions, PHP endpoints or blocked I/O” determines how narrowly the next change should be applied.
If the first attempt fails again, record whether “Reduce slow application work and tune caching before raising child count” actually ran and whether the same error returned. If it did, move to this specific advanced path: Calculate safe worker count from available RAM and use slowlog to trace handlers. A new, different error after a change is useful information: it may mean the original blocking stage was passed, but another dependency now needs attention.
How to verify the repair worked
Pass condition: Queue depth drops and peak requests complete without OOM kills. Repeat the operation with the same input, account or connection that originally failed. For intermittent faults, complete several cycles and review new logs; a single green status icon without a successful real-world task is not sufficient evidence.
If the test fails: undo the last reversible change, retain the saved diagnostic output and escalate with the exact failure text and model/release information. In particular, blindly increasing pm.max_children may exhaust shared host memory The next support contact should be able to reproduce your checks without guessing which settings were altered.
Solution Checklist
Frequently asked questions
Is this definitely a hardware failure or a product-wide outage?
No. The symptom alone—PHP-FPM logs server reached pm.max_children and requests queue—does not establish either cause. Start with “Check active workers, slow request log and per-process memory footprint” and determine whether the fault follows the account, connection, input, software version or physical component.
When should I stop troubleshooting myself?
Stop when the next action risks data loss, unapproved permissions or an unverified firmware or security change. For this case, the essential limit is: Blindly increasing pm.max_children may exhaust shared host memory. Bring your notes and diagnostic evidence to qualified support instead of repeatedly resetting the device or service.
Further reading
Consult Cloudflare Developer Documentation for current version and model-specific procedures. This is a general authoritative documentation entry point, not necessarily a page that confirms the exact error string in this guide. Verify product-specific instructions before applying advanced changes.