Kubernetes memory, explained

See what is really consuming Pod memory.

Kubernetes tells you memory is high. KubeMemLens separates the observed charge into application memory, file cache, shmem and residual signals so you can investigate why.

Get startedWatch demoView on GitHub

v1.0.2 available · Tested on AWS EKS · Linux and cgroup v2

Local demo showing 13 Pods, memory composition, limits and the selected Pod’s detailsThe main dashboardFull resolution
Choose a Pod from the memory overview.
Pod dashboard for demo-search-index showing 77% file cache, retained history and suggested checksInside a PodFull resolution
Inspect composition, history and read-only next checks.

Screenshots from the local demo. Open at full resolution to read the terminal text.

01 · Composition

A single memory number is not a diagnosis.

kubectl top can tell you that memory is high. It cannot tell you whether the charge comes from application memory, filesystem cache, tmpfs, pressure or another cgroup signal.

Repository sample · cache-heavy

Total observed charge4.10 GiB
Anonymous memory
1.20 GiB
File cache
2.51 GiB
Shmem / tmpfs
90.00 MiB
Residual / other
307.20 MiB
Diagnosiscache-heavy
Severityinfo
Confidencemedium

02 · Workflow

From signal to evidence.

Stay in the terminal and move through the investigation with read-only commands. KubeMemLens reports evidence and caveats; it does not mutate workloads.

Explore the CLI
  1. 01

    Find

    kubectl memlens top pods -A --sort-by cache

    Bring the Pods with the strongest memory signals to the top.

  2. 02

    Explain

    kubectl memlens explain pod api-123 -n production

    Inspect composition, limits, pressure and the evidence behind the diagnosis.

  3. 03

    Preserve

    kubectl memlens capture -n production --pod api-123 --include-history -o incident.json

    Write a redacted incident capture that you can review and compare.

In action

Similar totals.
Different memory.

Follow two Pods from the overview into their memory breakdown. Press Enter to inspect a Pod and Esc to return to your selection.

55 seconds · 4K · Silent · Local kind cluster with 13 synthetic PodsDownload video
Read the demo walkthrough

The video has no audio. Its on-screen walkthrough shows:

  1. Select demo-assets-cache in the main dashboard and press Enter.
  2. Inspect its composition: 84% file cache and 14% anonymous memory.
  3. Move between panels with Tab and open suggested checks. These are hints; nothing runs automatically.
  4. Press Esc to return to the same selected Pod, then open demo-upload-buffer.
  5. Compare the similar totals: about 76 MiB and 75 MiB. The second Pod is 85% shared memory / tmpfs.

These are recorded examples from local synthetic workloads, not live cluster data.

03 · Evidence

Tested on AWS EKS.

Memory inspection, capture and replay, access isolation, recovery and cleanup were exercised on AWS EKS. Use KubeMemLens in your workflow and tell us what you find.

EKS testingAL2023 · containerd · amd64 · two managed nodesEnvironment details ↗
FeedbackFound a bug or have a suggestion?Open a GitHub issue ↗

Exact supported and unsupported combinations are maintained in the compatibility contract.

Start here

Start investigating memory.

v1.0.2 adds a Pod dashboard with memory composition, context and history, plus a Go security update. Press Enter to inspect a Pod and Esc to return to your selection. Send feedback or report a bug on GitHub.

helm upgrade --install kube-memlens \
  oci://ghcr.io/danushkastanley/charts/kube-memlens \
  --version 1.0.2 \
  --namespace kube-memlens \
  --create-namespace \
  --set-string image.repository=ghcr.io/danushkastanley/kube-memlens \
  --set-string image.digest=sha256:d70146181fbc8a1d4c9220c78f3bf28bc9604919ef6ba5e1aaab699e055cecc5

Read the installation guideDownload v1.0.2