The PD API leader currently loads every keyspace group from etcd once per second when reconciling TSO node membership.
With millions of keyspaces, repeatedly loading and decoding every group creates unnecessary etcd traffic, allocations, and GC pressure, which can affect PD availability.
The reconciliation should be event-driven. Healthy groups should require no etcd read, and a possibly under-replicated group should be revalidated and updated individually in a transaction.
Related TiDB Cloud issue: https://github.com/tidbcloud/pd-cse/issues/559
The PD API leader currently loads every keyspace group from etcd once per second when reconciling TSO node membership.
With millions of keyspaces, repeatedly loading and decoding every group creates unnecessary etcd traffic, allocations, and GC pressure, which can affect PD availability.
The reconciliation should be event-driven. Healthy groups should require no etcd read, and a possibly under-replicated group should be revalidated and updated individually in a transaction.
Related TiDB Cloud issue: https://github.com/tidbcloud/pd-cse/issues/559