For every retryable write we update the session entry in config.transactions, this creates long update chains in WT, a large volume of writes to the WT history store, etc. In general we only need the persisted data during startup to reconstruct the in-memory state, so we could persist the data less frequently and reconstruct the missing pieces from the oplog.

Caveats: resharding relies on point-in-time access to config.transactions, rollback to stable timestamp would also need some changes.