A background worker vanishes during an import. The service restarts and the website looks normal again. Was the VPS short of memory? Possibly, but a restart, an exit code or a red chart is not enough to establish that. A useful investigation connects the failed process to a specific timestamp, memory boundary and termination record. That connection tells you whether to change capacity, concurrency or application behaviour.
Start with the disappearance, not the upgrade button
Record when the job stopped, which timezone the application uses, what operation was running and whether other services failed at the same moment. Keep the job identifier internally, but remove customer details and secrets before sharing an incident summary. Also check whether a deployment, scheduled restart or administrator action happened nearby. A normal restart can look alarming in a dashboard that shows only process uptime.
Consider an illustrative import that succeeds with one worker but stops when several imports overlap. This points toward a concurrency question; it does not yet prove a memory leak. A leak usually requires evidence of memory that keeps accumulating under comparable work. A legitimate large batch can instead produce a brief peak. Both need attention, but the remedies differ.
Three explanations to keep separate
- Kernel OOM: look for a matching kernel record that identifies the terminated task and memory context.
- A service or container boundary: the application may reach its own configured memory limit while the wider machine still has headroom.
- Another termination: an application exception, deployment, manual signal or a userspace memory-pressure service may have stopped it.
Linux documents the kernel OOM mechanism; systemd also documents a separate userspace service, systemd-oomd. The presence of either feature in documentation does not show that it is enabled on your VPS. Check the installed environment. Exit status 137, where interpreted as termination by SIGKILL, still does not identify who sent that signal.
Collect a small, comparable evidence window
On a server you administer, inspect the relevant service journal and kernel records around the incident using your existing read-only tools. Avoid exporting the whole journal. Match process names and timestamps, then compare memory, swap activity and the number of concurrent jobs for that same window. A memory reading taken after the restart can be reassuring precisely because the failed process is no longer using memory.
Write down the scope of every measurement: whole VPS, container, service or individual process. Do not compare a container limit with a host-wide available-memory figure as though they were the same denominator. If you cannot see a layer, mark it unknown and ask its administrator for the narrow evidence needed. Our VPS monitoring guide helps organise those observations.
Choose a remedy that fits the evidence
If overlap creates the peak, test a bounded worker count or smaller batch in an isolated environment first. If the same process grows across repeated comparable jobs, investigate application ownership and retained objects. If ordinary, healthy work consistently exceeds a verified limit, a capacity change becomes a defensible option. Preserve the failed-job state and retry rules: restarting everything may repeat an email, export or other side effect.
Swap deserves its own assessment. It can change how pressure behaves, but it is not a promise of acceptable latency or a replacement for understanding the workload. Read when swap helps and when it hides a problem before making it the universal answer.
The useful support request
Send the affected service, redacted evidence window, workload size, concurrent-job count and the resource boundary you actually measured. State whether the issue is repeatable and who can approve a change. When comparing VPS options, include the administration responsibilities as well as RAM. Success means the same legitimate work completes without unexplained termination and within acceptable latency, not merely that the next chart contains more free memory.
Technical references: Linux memory-management concepts and systemd-oomd documentation.



