mirror of
https://github.com/cozystack/cozystack.git
synced 2026-11-09 05:25:14 +09:00
Andrey Kolkov
33fa5e51b0
Two defects that break the etcd-operator on an in-cluster 1.5 -> 1.6 upgrade (#2859 swapped the upstream chart for the cozystack-authored one): - #3242: the controller Deployment's spec.selector.matchLabels changed ({instance,name} -> {name,control-plane}). spec.selector is immutable, so `helm upgrade` cannot patch the existing Deployment and the whole HelmRelease upgrade fails ("field is immutable") -- the operator and the v1alpha2 CRDs it serves never come up, which blocks the etcd v1alpha2 adoption. Fresh installs use the new selector directly, so fresh-install CI does not catch it. Add a pre-upgrade hook that deletes the Deployment ONLY when its live selector is the pre-1.6 one (lacks control-plane=controller-manager), so Helm recreates it cleanly. No-op when the selector already matches (rc.1 -> later) and never runs on a fresh install (pre-upgrade only). Keeping the selector stable in the chart is not an option: rc.1 already shipped the new selector, so aligning it back would merely move the immutable break to rc.1 -> next. - Raise the manager's cold-start memory limit floor 128Mi -> 256Mi (and the VPA minAllowed to match). The steady-state working set is ~250Mi (the VPA's own recommendation); at 128Mi a Pod that starts before the VPA admission webhook rewrites it (e.g. a Deployment recreated out of band, or the webhook briefly unavailable during upgrade) OOMKills into a crash loop. Defense in depth so the operator never depends on VPA timing merely to avoid crashing; the VPA still scales it further under load up to maxAllowed. Verified live on a 1.5 -> 1.6 adoption: the delete-Deployment step plus the VPA re-applying 256Mi+ let the operator come up and the adoption complete. Refs: #3242, #3243 Signed-off-by: Andrey Kolkov <androndo@gmail.com>