attune-io/attune v0.1.26 on GitHub

What's New in v0.1.26

v0.1.26 adds a waste GRADE on kubectl attune recommendations so you can see over-request and under-request at a glance. When Prometheus (or another metrics backend) goes quiet, recommendations are marked stale and the operator stops resizing, boosting, exporting, and opening GitOps PRs from that data.

Highlights

kubectl attune recommendations prints GRADE (A-F waste, or U when current is more than 10% under the rec).
Stale recommendations stay visible but do not drive apply paths until fresh samples arrive.
Ready reason MetricsUnavailable replaces PrometheusUnavailable for every metrics backend.

Waste grade

The recommendations table had no letter for request waste. GRADE is the worse of CPU or memory waste: A under 10% over (or up to 10% under), B 10-25%, C 25-50%, D 50-75%, F 75% or more. U when current is more than 10% below the recommendation, including a zero request. U wins if one resource is under and the other is waste. - while collecting, when the rec is stale, or when quantities cannot be compared. CSV adds a grade column before confidence_or_status (#604, #606).

Stale recommendations

Empty or gapped Prometheus data used to look like a live rec. The operator now sets status.recommendations[].stale when the newest finite sample is older than 3 * queryStep (default 15m), or when the last rec is reused after an empty query. Reuse expires after that same bound. Stale recs do not count toward Ready or savings (#610, #620).
Resize was not the only leftover consumer. Startup boost, CREATE initial sizing, template persistence, ConfigMap export, kubectl attune preview/diff/wizard, and GitOps drift all skip while stale is true. The CLI GRADE is - and CONFIDENCE / STATUS prints stale. The ConfigMap last-updated stamp does not move (#609, #611).

Safety and hold

A probe blip right after /resize triggered an immediate revert. Immediate post-apply safety now checks OOM and restart spikes only. NotReady, throttle, and SLO stay on the observation window (#627).
A one-sided sample gap (CPU present, memory missing, or the reverse) could write template requests. The operator holds the last rec instead. Hold prefers last rec over a live template snapshot and drops leftover template limits so they do not stick after the hold lifts (#625, #626).
A failed policy List looked like bootstrap (no conflicts). The operator now keeps last recs, sets Ready ConflictCheckFailed, and skips CREATE initial sizing for that policy until List succeeds (#660).
ResourceQuota names that are not requests.cpu / requests.memory were ignored. Common aliases such as cpu and memory are honored (#621).

GitOps and CloudWatch

CloudWatch PodName="api-*" matched a pod named api-*, not a prefix. SEARCH is exact. Prefix matching is now client-side (#620).
GitOps to a self-hosted forge on a private IP was blocked as SSRF. allowPrivateEndpoints permits RFC1918 and ULA hosts. Link-local IMDS (169.254.169.254 and fd00:ec2::254) stays blocked. Corporate HTTPS_PROXY no longer fails the GitOps dial pin (#620, #621).
GitLab merge requests were matched by title only. Matching now uses the configured base branch. Labels on an existing MR are not wiped on update. An empty rebase bootstrap is not treated as a new open (#622, #624).

Ready, defaults, and doctor

Ready said PrometheusUnavailable even for Datadog or CloudWatch. The reason is now MetricsUnavailable. Alerts that match the old reason should also match the new one. The operator still treats the old reason as bootstrap until the next reconcile overwrites it (#654, #655, #656).
AttuneDefaults Datadog or CloudWatch blocks were ignored (windows only). The policy inherits the provider pointer when it set none. HPA absolute targets are preserved. Query URLs are sanitized before use (#603).
Datadog collectors with the same app key but different query options shared a cache entry. Cache keys now include the app key plus query options. NaN/Inf samples increment attune_nan_inf_samples_total (#653).

Upgrade notes

Upgrade the chart to 0.1.26, or set image.tag to 0.1.26 or v0.1.26.
Pull ghcr.io/attune-io/attune:v0.1.26 or ghcr.io/attune-io/attune:0.1.26. Both tags point at the same digest.
Apply CRDs before helm upgrade if you need new schema fields (kubectl apply --server-side --force-conflicts from the release crds.yaml).
kubectl attune recommendations now has a GRADE column. Update any parser that assumed the old column order. CSV is ...,mem_rec,grade,confidence_or_status.
If Ready shows stale on a rec, fix metrics reachability. Resize and export resume when fresh samples arrive. Do not treat a frozen ConfigMap last-updated as current.
Alerts that match reason="PrometheusUnavailable" should also match reason="MetricsUnavailable".
GitOps to a private forge needs allowPrivateEndpoints. IMDS stays blocked.

See Upgrading for maxConcurrentResizes inherit and other 0.1.26 behavior notes.

Full changelog

v0.1.25...v0.1.26