txn_global.last_ckpt_disaggregated_schema_epoch is declared wt_shared, but all three stores are plain assignments:

• checkpoint_txn.c:2083 and :2098 (checkpoint completion — runs on every checkpoint, disaggregated or not)
• conn_layered_checkpoint_pick_up.c:1547 (checkpoint pickup adoption)

while query_timestamp("get=last_disaggregated_schema_epoch") reads the field with a lock-free acquire load (txn_timestamp.c:210). The stores run under the checkpoint lock, the reader takes no lock, so the pair is a data race (UB per C11; ThreadSanitizer reports it — see the reproducer comment). All other readers are checkpoint-lock- or schema-lock-ordered with the stores; the query API is the only racy access.

Practical impact is bounded — an aligned 8-byte store does not tear on our platforms, so the caller can only observe a stale epoch — but the race trips TSAN in any workload combining checkpoints with this timestamp query. One possible fix: use release stores at the three store sites, matching the adjacent fields.