All registration goes through SetupWithManagerForResourceType (eager and
lazy paths in main.go); the legacy single-resource SetupWithManager and
its corev1 import were dead code that misled readers about how the
controller wires up.
M7: extractUnstructuredContent only hashed 'spec' when present, dropping
all other top-level content fields. Resources with both spec and data
(or any non-spec content) silently drifted until the next 10m resync.
Now hashes every non-Kubernetes-managed top-level field, matching the
fields updateUnstructuredMirror copies.
M6: when a source has a transform annotation, also hash the source's
labels and annotations (filtered of kubemirror.raczylo.com/* keys to
avoid the controller's own bookkeeping churning the hash). Templates
read these via TransformContext; without this a label change wouldn't
re-render the transformed mirror.
H3: text/template.Execute is not context-aware, so applyTemplateRule's
timeout cancels the select but leaks the executor goroutine. Added a
process-wide semaphore (cap 64) so a runaway template can't spawn an
unbounded number of stuck goroutines on every reconcile.
M4: zap dev mode (DPanic-on-error, console output, stacktraces on
warning) was hardcoded on. Defaulted to production; --zap-devel flag
remains for opt-in.
M5: WaitForInitialDiscovery was anchored on context.Background() with
its own WithTimeout, so SIGTERM during startup couldn't abort the wait.
Now anchors on signalCtx.
H2: scanAndRegister held d.mu (write lock) across registerController
and registerMirrorControllerOnly. Those calls enter controller-runtime's
manager state machine, which takes its own internal locks and can block
on cache sync — holding our application-level write lock across them
is a latent deadlock the moment any reentrant access happens (health
checks reading GetRegisteredCount, factories that introspect state).
Restructured into three phases: snapshot work under RLock, perform
registrations with NO lock held, then commit results under Lock.
Registration step routed through funcs to keep tests honest about
the lock state at the moment of invocation.
H4: NamespaceReconciler.reconcileMirror builds an ad-hoc SourceReconciler
to delegate mirror creation. The previous version left APIReader and
CircuitBreaker as nil, which silently disabled freshness verification
on the namespace-driven path (a label change to a target namespace
would mirror cached, possibly stale source data) and bypassed circuit
breaker accounting for those reconciles.
Construction extracted into newSourceReconciler so the forwarding is
covered by a unit test that pins both fields by identity.
C2: updateLastSyncStatus wrote the sync-status annotation on every
successful reconcile. Because the source's watch predicate is the
'enabled' label (server-side filter), that Update fires a watch event
that re-enters Reconcile. With reconciled/error counts varying across
cycles, the value differs each time, so the API server bumps RV and
the loop never quiesces. Now skips the Update when the value matches
the existing annotation.
C3: NamespaceReconciler's happy-path returned RequeueAfter=3s
unconditionally. Every namespace in the cluster re-reconciled every
3 seconds forever, generating constant List calls per source kind.
Now returns ctrl.Result{}; cache-staleness windows are handled by
the manager's resync period and source freshness verification.
H1: --verify-source-freshness used to default to false, so any source
update whose annotation was still in the informer cache (5-20s lag)
would resolve the wrong target list. cleanupOrphanedMirrors then ran
against the stale list and missed orphans (manifested in e2e as
'Orphaned mirror in kubemirror-e2e-app-1 not deleted within timeout'
after target-namespaces was changed). Defaulting to true fixes the
race; the trade-off is one extra API read per stale-cache reconcile.
M2: ResolveTargetNamespaces glob branch checked filter.IsAllowed but
not the opt-out map, so a namespace labeled allow-mirrors=false would
still receive a mirror through patterns like 'app-*'. The 'all' branch
already had the guard; the glob branch now does too. Direct namespace
listings still bypass opt-out by design (explicit author intent).
C1: deleteAllMirrors used to issue a blind Delete on every namespace
matching the source name+GVK, which would destroy unrelated resources
(e.g. a 'default' SA, 'ca-bundle' ConfigMap) sharing the source name.
Now reads each candidate, verifies managed-by label and source-reference
annotation, and only deletes confirmed mirrors.
M1: BlacklistedSecretTypes was declared but never enforced. Enabling
mirroring on a service-account-token / bootstrap-token / helm release
Secret would mirror credentials cluster-wide. Now refused at Reconcile.
M3: deleteAllMirrors swallowed per-namespace errors and returned nil,
so callers removed the finalizer even on partial failure (orphans).
Errors are now joined and returned.
lukaszraczylo