Skip to content

grafana/osdc: add hf-cache OOM-killed-mounts panel#916

Merged
huydhn merged 3 commits into
pytorch:mainfrom
huydhn:hf-cache-oom-dashboard-panel
Jul 19, 2026
Merged

grafana/osdc: add hf-cache OOM-killed-mounts panel#916
huydhn merged 3 commits into
pytorch:mainfrom
huydhn:hf-cache-oom-dashboard-panel

Conversation

@huydhn

@huydhn huydhn commented Jul 17, 2026

Copy link
Copy Markdown
Contributor

Adds a timeseries panel hf-cache Mounts Broken by rclone OOM [cluster] to the OSDC dashboard.

count by(cluster) (kube_pod_container_status_last_terminated_reason{cluster=~"$cluster", namespace="hf-cache", container="rclone", reason="OOMKilled"} == 1)

hf-cache runs one rclone DaemonSet pod per node serving /mnt/hf_cache, so an OOMKilled rclone container = a broken node-wide FUSE mount (usually wedged in CreateContainerError until the node is recycled). This surfaces, per cluster, the small-tier OOM count that until now needed manual kubectl queries.

Preview

Dashboard

@huydhn
huydhn requested review from jathu and jeanschmidt July 17, 2026 16:50
@jathu

jathu commented Jul 17, 2026

Copy link
Copy Markdown
Collaborator

Do you have a demo link?

@huydhn

huydhn commented Jul 17, 2026

Copy link
Copy Markdown
Contributor Author

Do you have a demo link?

Ah, forgot to link it, It's here, the bottom chart. I'm in the middle of adding highlight for this one. The threshold would be flat 0

@huydhn huydhn changed the title grafana/osdc: add hf-cache broken-mount (rclone OOM) panel grafana/osdc: add hf-cache OOM-killed-mounts panel Jul 17, 2026
@huydhn
huydhn force-pushed the hf-cache-oom-dashboard-panel branch from 70dfbd1 to b94478a Compare July 17, 2026 22:20
huydhn added 2 commits July 18, 2026 16:12
Counts nodes whose per-node hf-cache rclone mount was OOM-killed
(kube_pod_container_status_last_terminated_reason, container=rclone,
reason=OOMKilled) — a broken node-wide /mnt/hf_cache FUSE mount. Surfaces
the small-tier OOM stat that previously required manual kubectl queries.
…#917)

pytorch#917 raised the CPU catch-all and 1-GPU reserves to 640Mi in deploy.sh but the
smoke test's MOUNT_TIERS still asserted 256Mi/512Mi, failing 'deploy-fast / Smoke
tests'. Sync the smoke-test expectations to deploy.sh.
@huydhn
huydhn force-pushed the hf-cache-oom-dashboard-panel branch 2 times, most recently from c259a63 to 4a7ac1b Compare July 18, 2026 23:25
…+ soft-capped axis

- thresholds: green@0, red@1 (line+area) so any OOM'd mount (>=1 node) turns red
- query 'or vector(0)': no series (0 OOM) renders as 0/green instead of 'No data'
- decimals: 0 so the y-axis shows integers only (node counts)
- axisSoftMax: 5 keeps the calm-view axis readable for a few nodes, but still
  expands for a real wave (no clipping / no understating an incident)
palette-classic-by-name kept for per-cluster series.
@huydhn
huydhn force-pushed the hf-cache-oom-dashboard-panel branch from 4a7ac1b to 95dbaa4 Compare July 19, 2026 07:06
@huydhn
huydhn added this pull request to the merge queue Jul 19, 2026
Merged via the queue into pytorch:main with commit 89792e6 Jul 19, 2026
13 checks passed
@huydhn
huydhn deleted the hf-cache-oom-dashboard-panel branch July 19, 2026 07:09
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants