[scotwells]

Requested by Scot Schuchert-Wells

What

The bgpadvertisement and trafficpolicybinding controllers in galactic-controller report reconcile errors for races that resolve on their own:

• The BGPAdvertisement projection updates NetworkInterface, claim and VPCAttachment status that network-services-operator also writes, each side from a cached copy. Whoever writes second gets a conflict.
• A TrafficPolicyBinding is often deleted with its interface between the read and the status apply, so the apply returns not-found.
• A status write to an interface, claim or attachment that was deleted mid-pass returns not-found.

These errors keep the reconcile error floor alert firing. They don't point to any real fault.

Change

• bgpadvertisement: a conflict now requeues with backoff instead of returning an error. A status write to an interface, claim or attachment that was deleted mid-pass is a no-op.
• trafficpolicybinding: a not-found status apply means the binding is gone, so the reconcile ends cleanly.
• Every other error still surfaces.

Tests cover each case and confirm that other errors still surface. Each new test fails without the fix.

A companion network-services-operator change, datum-cloud/network-services-operator#884, handles the same races on its side.

[ecv]

Approve.

Conflicts requeue through the rate-limited workqueue (bounded backoff, fresh cached read next pass), and NotFound is skipped only on status writes after the object is gone, never on a path that does deletion work.

• No blind overwrite: the writes stay Status().Update carrying the read resourceVersion, so a lost race still returns Conflict and is retried, not forced. The per-attachment loop is idempotent, so re-running after a mid-loop conflict is safe.
• bgpadvertisement_controller.go skips only on NotFound from a status write (clearWithdrawnAdvertisement, attachment/interface/claim projection). It does not key off DeletionTimestamp, so withdrawal clearing and the APIReader confirmation on the NotFound-at-read path are unchanged. Nothing that must run during deletion is skipped.
• #1369 (attachment without a node maps to all tenant advertisements) fans more reconciles at the same attachments, so this conflict path will be hit more often. It stays fine: Conflict is a debug log plus requeue, not an error metric.
• reconcilers.go: NotFound on the binding SSA status patch returns nil. Right, since the binding has no status once gone.

Follow-ups:

• Requeue: true has no cap beyond the workqueue's per-item exponential backoff; a persistent conflict (two controllers fighting over a condition) would retry forever at the max delay with only a V(1) log. Consider a counter metric for conflict requeues.
• The test covers only the interface-status write; attachment and claim writes share the same handling but have no test.