Kubernetes OOMKilled proxy versus origin split response

When this happens

Use this when a cdn, load balancer, reverse proxy, or waf sits in front of the service. Do not stop at the screen message; validate whether the edge/proxy or the origin server creates the failure first.

Symptom checklist

• OOMKilled appears repeatedly in the Kubernetes UI or logs.
• events, describe, probes, requests, image pull differs between successful and failed requests.
• The issue appears only after separating pod internal error from cluster scheduling issue.
• It often follows deploys, permission changes, configuration edits, or data refreshes.

Likely causes

• OOMKilled specifically changes the investigation surface for Kubernetes: verify the exact failing object, route, user, and timestamp before applying the broader pattern.
• The process exceeds memory or heap limits under real data size.
• Disk space, inode count, or mounted volume state fails before byte usage looks obvious.
• File descriptors leak across repeated requests.
• Logs, cache, or build artifacts grow without retention.
• A workload runs synchronously instead of being chunked.
• For the proxy versus origin split case, the first useful clue is whether the edge/proxy or the origin server creates the failure.

First 1-minute checks

1. Write down the first failure time, latest change, affected user, path, and object ID.
2. Compare events, describe, probes, requests, image pull for success and failure in the same window.
3. Test the hypothesis: resource exhaustion caused by memory pressure, disk or inode shortage, descriptor limits, cache growth, volume state, or oversized workload.
4. Classify this as proxy versus origin split: whether the edge/proxy or the origin server creates the failure.
5. Capture current values before changing configuration.

First evidence

Treat OOMKilled as a proxy versus origin split case. First collect evidence for whether the edge/proxy or the origin server creates the failure.

Output examples

Normal output

Direct origin and routed proxy return equivalent status and headers.

Failing output

Origin succeeds but proxy fails, or proxy succeeds while direct origin exposes the real backend error.

Output-to-action branches

• A CDN, load balancer, reverse proxy, or WAF sits in front of the service.
  Test direct origin and routed traffic separately, then change only the layer that differs.
• The working and failing outputs differ.
  Act on the differing layer first: For OOMKilled, apply the fix only after reproducing the same condition and saving the before/after evidence for this exact code.
• Command output is normal but users still fail.
  Separate browser cache, cookies, permissions, and network location before declaring it fixed.

Do not do this

• Do not assume the CDN error code is the application root cause.
• Do not change multiple layers before identifying the failing layer.
• Do not delete production data, grant broad permissions, or disable security controls as a first response.

Evidence quality

Auto-generated operator draft: includes issue-specific causes, commands, output branches, and unsafe-action warnings. Official-source links and real incident validation are queued for enrichment.

Commands to run first

grep -R "OOMKilled" ./logs

kubectl describe pod POD

kubectl get events --sort-by=.lastTimestamp

kubectl logs POD --previous

kubectl describe ingress INGRESS

df -h && df -i

ulimit -n

ps aux --sort=-%mem | head

du -sh ./* | sort -h | tail

curl -Iv https://example.com --resolve example.com:443:ORIGIN_IP

Fix order

1. Record the full OOMKilled message, failing URL, user, object ID, and latest change.
2. Collect issue-specific evidence for resource exhaustion caused by memory pressure, disk or inode shortage, descriptor limits, cache growth, volume state, or oversized workload.
3. Compare the failing case with a successful case before editing settings.
4. If this is the proxy versus origin split branch, Test direct origin and routed traffic separately, then change only the layer that differs.
5. Re-check with the same command and URL, then record the normal output.

Actions by cause

• For OOMKilled, apply the fix only after reproducing the same condition and saving the before/after evidence for this exact code.
• Measure memory, disk, inode, descriptor, and mounted volume state at failure time.
• Clean generated artifacts and add retention policies.
• Raise limits only after removing leaks or oversized batches.
• Chunk large jobs and stream outputs where possible.
• Add alerts before hard exhaustion.
• For the proxy versus origin split branch, Test direct origin and routed traffic separately, then change only the layer that differs.

Evidence links

• Kubernetes troubleshooting applications official
• Kubernetes debugging services official

Verification metadata

• operator-draft
• official-reference-linked
• 2026-07-23

Update queue

• Review cadence
  weekly-source-review
• Next enrichment
  Add one official-source check and one real output example for Kubernetes OOMKilled.

Environment-specific checks

• Shared hosting, proxies, VPNs, or CDN layers can change pod internal error from cluster scheduling issue results.
• Do not trust only the Kubernetes UI; compare command output.
• Japanese hosting panels may show completion before DNS or SSL fully propagates.
• Test from both office and external networks.

Prevent it next time

• Store normal examples for events, describe, probes, requests, image pull.
• Add probes, resource limits, secrets, service selectors, ingress to the release checklist.
• Keep recurring errors in the same note format.
• Split alerts by error rate, latency, certificates, disk, and permission changes.

Related notes

• Kubernetes CrashLoopBackOff
• Kubernetes ImagePullBackOff fix note
• Kubernetes ImagePullBackOff fix note 2
• Kubernetes ImagePullBackOff fix note 3
• Kubernetes ImagePullBackOff fix note 4
• Kubernetes ImagePullBackOff fix note 5
• Kubernetes ImagePullBackOff fix note 6
• Kubernetes CrashLoopBackOff first triage response