Skip to content

Refresh component versions and offline examples for 26.11 - #1428

Merged
michael-balint merged 9 commits into
masterfrom
dholt/26.11-version-bumps
Nov 4, 2026
Merged

michael-balint merged 9 commits into
masterfrom
dholt/26.11-version-bumps

Conversation

@dholt

@dholt dholt commented Oct 19, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Refresh published component versions and keep offline preparation aligned with role defaults:

  • Slurm 26.05.5; GPU Operator v26.7.2; device plugin and GPU Feature Discovery 0.20.1; MIG Manager 0.15.1.
  • Prometheus 3.15.0; Alertmanager 0.34.1; Grafana 13.2.4; node-exporter 1.12.1.
  • Distribution registry 3.1.2; registry proxy 0.7.0; Network Operator 26.7.0.
  • Retain master's kube-prometheus-stack 92.14.0 and its CRD-first recovery path from Prepare kube-prometheus-stack 92.14.0 for DeepOps 26.12 #1437; this PR no longer changes the monitoring installer.
  • Align offline chart, image and archive examples; add regression tests for pin consistency.
  • Preserve the Operator-managed driver pin at 580.126.20 rather than changing driver branches with the chart update.

Validation

On integration head 6a3cee6d20208759fa369f134f284a944d3c1181:

  • All 30 GitHub checks passed on this exact head.
  • Independent integration review passed with no blocking findings.
  • Role lint passed for 346 files; 12/12 playbook syntax checks passed.
  • 165/165 local validation tests passed with no skips, including offline pin checks and monitoring chart/image renders against kube-prometheus-stack 92.14.0.
  • Shell syntax, changed-role YAML parsing and whitespace checks passed.
  • The monitoring installer is identical to master; all 16 remaining PR files are identical to the previously reviewed version-refresh head.

Earlier live QA, not rerun on this integration head:

  • On ef258397, Slurm 26.05.5 build, an A30 GPU job via srun, a Pyxis container GPU job, login GPU guard and Slurm exporter checks passed with driver 580.178.04.
  • On a1e1f4cb, single-node Kubernetes, GPU Operator and a GPU pod running nvidia-smi passed with an A40 and preinstalled driver 580.178.04.
  • That Kubernetes run used the previous monitoring chart: nonpersistent monitoring deployed, Prometheus collected GPU/node/Kubernetes metrics, and Grafana loaded 30 dashboards with sampled queries returning data. It is not live proof of chart 92.14.0 on this combined head.

QA limits

  • No live cluster tests were rerun for this master merge. Earlier Slurm and GPU Operator results are retained because their version-refresh files are unchanged.
  • The Operator-managed driver pin was not exercised by the preinstalled-driver Kubernetes test.
  • Persistent monitoring and monitoring upgrades on this combined head, live MIG reconfiguration, and Network Operator secondary-network traffic are not established by these results.
  • Existing GPU Nodes host/node panels returned no data during the earlier Kubernetes QA; not every dashboard panel is validated.
  • Network Operator 26.7.0 requires Kubernetes 1.32 or newer; older cluster-version overrides need separate validation.
  • Existing monitoring installations must follow docs/k8s-cluster/monitoring-upgrade-92.md; a plain Helm upgrade is not a substitute for the CRD-first procedure.
  • No new operating-system, firmware or hardware support is claimed.

Ready for maintainer review; the QA limits above still apply.

dholt and others added 9 commits October 5, 2026 08:25
Update monitoring, registry proxy and Slurm pins from public upstream
releases. Keep the registry-cache example aligned with its role default.

Leave Network Operator and MIG Manager unchanged pending verification.

Signed-off-by: Doug Holt <[email protected]>
Keep Slurm archives and monitoring pull, save, mirror and override
examples aligned with the selected 26.11 component defaults.

Signed-off-by: Doug Holt <[email protected]>
Default kube-prometheus-stack to 88.6.5 so monitoring meets the
minimum version required by Run:ai platform mode.

Signed-off-by: Doug Holt <[email protected]>
Use published stable releases and align offline download examples.
Select Network Operator 26.7.0, the newest chart currently published
in the configured repository, rather than its newer source-only release.
Retain the monitoring chart floor pending major-upgrade review.

Co-Authored-By: Paperclip <[email protected]>
Signed-off-by: Doug Holt <[email protected]>
Preserve the version-refresh commits while integrating current master.
Retain Slurm 26.05.5 alongside the upstream opt-in Lua build setting.

Co-Authored-By: Paperclip <[email protected]>
Signed-off-by: Doug Holt <[email protected]>
Mirror GPU Operator v26.7.2 and registry 3.1.2 consistently across
connected-side preparation, transfer records, and offline imports.
Add regression checks against role defaults, including the explicit
driver override. Retain R580 to avoid an unvalidated driver-branch
upgrade as part of the chart refresh.

Co-Authored-By: Paperclip <[email protected]>
Signed-off-by: Doug Holt <[email protected]>
Preserve the version refresh while incorporating the latest upstream changes.

Co-Authored-By: Paperclip <[email protected]>
Include the driver role update and monitoring failure handling while
preserving the reviewed version-refresh commits.

Signed-off-by: Doug Holt <[email protected]>
Co-Authored-By: Paperclip <[email protected]>
Preserve the current monitoring installer and all remaining version
refresh changes without rewriting existing commits.
@dholt
dholt marked this pull request as ready for review November 3, 2026 16:28
@dholt
dholt requested a review from michael-balint November 3, 2026 16:28
@michael-balint
michael-balint merged commit 6428a10 into master Nov 4, 2026
30 checks passed
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