In the BF, we noticed that while preallocating the snapshot for multi-statement transactions during startup recovery as part of reconstructPreparedTransactions(), which is single-threaded, the preallocation might fail and require a retry due to catalog invalidation, even if no other real conflicting threads are present. We found that a concurrent TimestampMonitor thread can drop the physical idents but doesn’t update the on-disk mdb_catalog, which can invalidate the in-memory catalog. I believe dropping physical ident shouldn't invalidate the catalog.