Skip to content

gpu page: host health table with CPU/RAM/Disk bars and drill-down - #114

Merged
Thangnguyenvn98 merged 1 commit into
agent/gpu-temp-power-badgesfrom
agent/gpu-host-metrics-ui
Sep 3, 2026
Merged

Thangnguyenvn98 merged 1 commit into
agent/gpu-temp-power-badgesfrom
agent/gpu-host-metrics-ui

Conversation

@Thangnguyenvn98

Copy link
Copy Markdown
Collaborator

Stacked on #113 (base retargets to main automatically when it merges).

What

The approved "table + drill-down" prototype design, rewritten properly on the real stack:

  • Host Summary table gains CPU / RAM / Disk progress bars per host, next to the existing health dot, Stale/Unreporting badges, GPU Utilization bar, GPU Temp, and the per-GPU memory mini-bars (kept verbatim)
  • The Disk cell shows the worst job-critical mount (roles workspace/images/data/system); other mounts never drive it; tmpfs mounts are labeled RAM-backed
  • Row click expands a drill-down: reporter status + last error (broken reporter vs broken machine), k8s node-condition chips with a distinct Cordoned marker, per-mount disk bars with role/device/fstype/size (including errored mounts), and per-GPU util/mem/temp
  • Hosts without host metrics yet (reporter not rolled out) render em-dashes, not broken cells
  • /api/gpu/latest gains a hosts array via a skip-scan latest-row-per-host query over host_snapshots, keeping the page's single 30s SWR poll

Test plan

  • npm test: 98/98 (24 new: query normalization, worst-alertable-disk selection, graceful-missing-host path, 11 renderToStaticMarkup component tests)
  • npm run lint, tsc --noEmit, npm run build all clean

Add a host-level read layer (queryHostLatest, skip-scan latest row per
hostname over host_snapshots) and fold it into /api/gpu/latest so the
page keeps its single 30s SWR cadence. The Host Summary table becomes a
GpuHostTable component: health dot, GPU util/temp, verbatim per-GPU
memory mini-bars, and new CPU/RAM/Disk progress bars, where the Disk
cell shows the worst alertable mount (workspace/images/data/system;
'other' never drives it) and tmpfs mounts are labeled RAM-backed. Rows
expand to a drill-down with reporter status and last_error, node
conditions (including a Cordoned marker for cordoned-but-Ready nodes),
per-mount disk detail with errors, and per-GPU breakdown. Hosts whose
reporter has not rolled out host metrics render em-dash empty states.

Signed-off-by: Thang Nguyen <thang.nguyen@inferact.ai>
@vercel

vercel Bot commented Sep 3, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
vllm-ci-dashboard Ready Ready Preview Sep 3, 2026 1:39am UTC

Request Review

@Thangnguyenvn98
Thangnguyenvn98 merged commit 42d67d1 into agent/gpu-temp-power-badges Sep 3, 2026
3 checks passed
Thangnguyenvn98 pushed a commit that referenced this pull request Sep 3, 2026
Signed-off-by: Thang Nguyen <thang.nguyen@inferact.ai>
Co-authored-by: Thang Nguyen <thang.nguyen@inferact.ai>
@Thangnguyenvn98
Thangnguyenvn98 deleted the agent/gpu-host-metrics-ui branch September 3, 2026 02:15

This branch was successfully deployed

1 active deployment
Preview — 52ef0799 Deployed Sep 3, 2026 by vercel[bot]
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.

1 participant