fix(azure): fence paginated membership reads
Azure Table continuation pages are separate Query Entities requests, so flattening them can combine membership rows from different topology versions even though IMembershipTable.ReadAll requires an atomic result.
This adds !Start and ~End, whose keys sort before and after all IPv4 and IPv6 membership row keys. Membership inserts and updates atomically write the changed silo row, the legacy VersionRow, and both boundary rows in one Azure entity group transaction.
Paginated reads compare the boundary membership versions from the completed query. Matching values prove that no boundary-aware membership update committed while the pages were read, preserving the single-query non-conflict path. A mismatch retries the query up to a bounded limit before throwing InconsistentStateException.
During rolling upgrades, older silos update only VersionRow. If VersionRow is newer than both boundary versions, the reader accepts the potentially torn snapshot instead of retrying, preserving availability while the cluster transitions. Tables without boundary rows likewise retain legacy behavior and skip the atomicity check. UpdateIAmAlive and defunct-row cleanup remain intentionally outside the topology fence.
References:
- Entity group transactions
- Performing entity group transactions
TableTransactionActionType- Query Entities
- Query timeout and pagination