github attune-io/attune v0.1.38

28 minutes ago

Highlights

A request increase is no longer sent when the node cannot hold the pod once init containers, pod-level requests, and overhead are counted.

A canary app starts a new observation watch after a safety revert of an older resize, instead of promoting on the previous clock.

Bug fixes

  • A neighbor with a large init container, pod-level requests, or overhead was treated as smaller than it is. The node-capacity pre-check now uses the same effective request as the scheduler. A small init container still does not block an increase by itself, and this pod's current spec is subtracted once (#1169).

  • A canary app could promote immediately after a safety revert of an older cycle. If that app had resized again since the reverted cycle, the old observation clock stayed. A successful revert now clears that app's clock. The next in-place success starts a new watch. A failed revert, and an app that is already fully promoted, stay as they were (#1168).

  • Freeze still lowered CPU when a startup boost expired and tenant metrics were rejected. Expiry now waits until the namespace is not frozen. A pending safety revert on that path still runs (#1164).

  • A closed resize window looked like it blocked startup boosts. Steady-state resizes stay inside the window. Boost apply and expiry, and the CREATE webhook, do not. Freeze still stops boosts. The log Outside resize window, skipping resize can appear in the same cycle as a boost (#1145).

  • Recommend mode said a schedule has no effect when template updates still follow that window. With templatePersistence.when set to OnRecommendation, the warning now says the schedule does not gate resizes and still gates those template updates (#1163).

  • A safety revert the API server had already stored could be reported as a failure when the reconcile was canceled. The confirm read uses its own timeout. On Kubernetes 1.33, an in-place memory limit decrease on revert is still refused when the resize policy is NotRequired (#1163).

  • A resize could be attempted on a pod that had already been replaced. The resize is skipped. The error names the live UID and the UID Attune expected (#1165).

  • The operator-down alert stayed quiet when every replica stopped reporting. AttuneOperatorDown fires when no scrape shows the operator up. One live replica, including a standby that is not the leader, does not page (#1142).

  • The fleet report was missing when watchNamespaces left out the operator namespace. The operator reads that ConfigMap from the API and still creates it when it is absent (#1143).

  • A long workload name could fail the cooldown annotation or the recommendation export. Names that already fit are unchanged. Longer names are shortened so the key and the export label stay within 63 characters (#1144).

  • Negative CPU or memory bounds were accepted. Admission rejects a negative minAllowed or maxAllowed. A dry run no longer stores a fingerprint that blocks the first live GitOps pull request. kubectl attune diff uses argoproj.io/v1alpha1 for an Argo Rollout (#1162).

  • The Helm chart page on Docker Hub showed the operator README. It now shows the chart overview (#1128).

Upgrading

No new required fields. Existing policy YAML still applies. Upgrade the chart to 0.1.38, or set the image tag to 0.1.38 or v0.1.38. The image is built with Go 1.27.4.

A request increase that used to be sent may now be skipped when a neighbor's init container, pod-level request, or overhead already fills the node. The reason is total pod requests would exceed node allocatable or node free request budget exceeded by neighbors.

A closed resize window does not stop startup boosts. Freeze does.

Kubernetes support stays 1.32 and newer. On 1.33, an in-place memory limit decrease is still refused when the resize policy is NotRequired. Use RestartContainer or move to 1.34 or newer.

Guide: Upgrading.

Contributors

Thanks to external contributors in this release:

  • @konih for reporting and fixing the CI gate when change detection does not succeed (#1134, #1141)
  • @konih for reporting and fixing the operator-down alert (#1138, #1142)
  • @konih for reporting and fixing the fleet report when the operator namespace is not watched (#1137, #1143)
  • @konih for reporting and fixing cooldown keys and export labels that exceeded 63 characters (#1135, #1136, #1144)
  • @konih for reporting and fixing the canary watch after a late revert (#1166, #1168)
  • @konih for reporting and fixing neighbor effective requests in the node-capacity pre-check (#1167, #1169)
  • @konih for reporting the low-severity review findings (#1139)
  • @konih for reporting whether startup boosts should follow the resize window (#1140)

Full changelog

v0.1.37...v0.1.38

Don't miss a new attune release

NewReleases is sending notifications on new releases.