Whenever an upsert/deletion of a config entry happens, within the open
state store transaction we speculatively test compile all discovery
chains that may be affected by the pending modification to verify that
the write would not create an erroneous scenario (such as splitting
traffic to a subset that did not exist).
If a single discovery chain evaluation references two config entries
with the same kind and name in different namespaces then sometimes the
upsert/deletion would be falsely rejected. It does not appear as though
this bug would've let invalid writes through to the state store so the
correction does not require a cleanup phase.
R.B. Boyer