A flow control technique prevents exhaustion of storage resources in an exactly once semantics (EOS) system of a storage input/output stack executing on a node of a cluster. An EOS server may service transactions sent by an EOS client and issue replies with results to the EOS client. In order to replay the transactions during normal operation after recovery from a crash, the EOS server persistently stores the transactions in the storage resources until an acknowledgement of completion is received from the EOS client for each pending transaction. The EOS client may issue a checkpoint acknowledgement, e.g., as a prune record, after a periodic interval that marks the completion of all pending transactions issued prior to the record. The EOS server need only log the prune record (rather than each pending transaction) to thereby prevent exhaustion of the storage resources, while also minimizing logging overhead of the server. In response to the crash and during replay of the transactions, the EOS server may employ the prune records to ignore those transactions that have already been acknowledged by the EOS client, thereby reducing time required for replay of the transactions.