Linux: out of memory – killed process (OOM killer)

What the message looks like

Out of memory: Killed process 2314 (php-fpm8.3) total-vm:812344kB, anon-rss:402112kB, file-rss:0kB, shmem-rss:20480kB, UID:33 pgtables:1240kB oom_score_adj:0

Common causes

• Too many PHP-FPM workers or database buffers for the available RAM.
• A memory leak in an application that grows over hours.
• No or too little swap, so load spikes are not absorbed.
• cgroup limit (MemoryMax, Docker --memory) reached – then “Memory cgroup out of memory”.
• One-off jobs like backups, imports or image processing need a lot of memory briefly.

Check

1. Find OOM events in the kernel log
   journalctl -k --since '24 hours ago' | grep -iE 'out of memory|oom-kill|Killed process'Copy
2. Show free memory and swap
   free -h; swapon --showCopy
3. Show processes using the most memory
   ps -eo pid,user,rss,comm --sort=-rss | head -n 10Copy
4. Show the memory limits of the affected service
   systemctl show php8.3-fpm.service -p MemoryMax -p MemoryCurrent -p OOMPolicyCopy

Fix

• Match pm.max_children to RAM: available memory divided by average worker size.
• Limit database buffers, e.g. innodb_buffer_pool_size to fit the rest of the system.
• Add some swap so short spikes do not cause an immediate kill.
• Protect critical services with OOMScoreAdjust=-500 in the unit, cap non-critical ones with MemoryMax.
• Monitor memory usage over time to spot leaks early.

FAQ

Why was the database killed although PHP used the memory?

The OOM killer picks by size and oom_score_adj, not by who caused it. Large services like MariaDB are therefore often hit first.

Does disabling overcommit help?

Rarely. With vm.overcommit_memory=2 applications fail when allocating instead. Better set per-service limits and match usage to RAM.