Question

A bad deploy of the `pricing` service is rolled back at 16:05 after a latency regression. The rollback uses the same rolling strategy and is reported 'complete' at 16:09. But errors don't fully clear — they settle at ~4% with `proto: cannot parse invalid wire-format data` on the `quote` RPC, intermittent and seemingly random across requests. Dashboards: the deploy tool shows 100% of `pricing` pods on the old (rolled-back) version; but the *caller*, `checkout`, was deployed an hour earlier and is still on its new version. The new `pricing` had added a field to the protobuf response (field number 9, a new nested message); `checkout`'s new version started reading field 9. After rollback, old `pricing` no longer sends field 9, but ~4% of in-flight gRPC connections are sticky/long-lived. How do you triage and mitigate?