You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 5bcdb8c
Browse filesBrowse the repository at this point in the historyBrowse files
fix(postgres): let in-place restore re-bootstrap instead of wedging
An in-place RestoreJob (targetApplicationRef omitted) re-bootstraps the
Postgres app under its own name. On stock v1.6.2 it never comes back up,
for two independent reasons in the same restore patch - both surface only
in place, because a to-copy restore targets a different app.
1. Archive-vs-recovery serverName collision. The driver set the restored
cluster's WAL-archive serverName equal to the recovery source
(bootstrap.serverName = bootstrap.oldName = sourceServerName) while the
chart pinned the archiver serverName to the release name. In place those
coincide, so the cluster archived onto the source's own WAL prefix and
barman-cloud-check-wal-archive refused to start ("Expected empty
archive"). Fixed by giving the restored cluster a fresh
bootstrap.newServerName, keyed by the RestoreJob UID (unique per
restore, stable across reconciles); the chart archives there while
externalClusters still recovers from bootstrap.serverName. This is what
the manual CNPG strategies hit (serverName defaults to the cluster
name); the cozy-default flow dodged it (serverName is <ns>-<app>).
2. useSystemBucket guard. The restore populates explicit S3 coordinates
(destinationPath / s3CredentialsSecret / endpointCA) but left
backup.useSystemBucket=true, so the chart's "bootstrap.enabled is
incompatible with useSystemBucket" guard failed the HelmRelease upgrade
and the cluster never re-rendered. This is what the cozy-default
(system-bucket) flow hits on an in-place restore. Fixed by clearing
useSystemBucket in the restore patch, since the app is now on the
explicit-creds path. The driver's app view (postgresapp.Backup) gains
the field, without omitempty so a false survives the merge patch.
Both are exercised only by an in-place restore; the postgres backup e2e
drove only to-copy, so neither surfaced in CI. A step 50 in-place restore
now guards reason 1 (manual-strategy flow); reason 2 is covered by a Go
unit assertion (a system-bucket in-place e2e needs the platform bucket in
CI and is left as a follow-up).
Signed-off-by: Andrey Kolkov <[email protected]>
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
0 commit comments