Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
Description
Motivation
Our transaction protocol assumes that all required request validations, lock acquisitions and similar activities are performed on a primary replica prior to command replication, meaning that it's not necessary to await replication for every request one by one rather it's required to await them all at once in pre-commit phase. Most of what is required for such all at once await has already been implemented.
Definition of Done
- It's required to do the command replication in an async manner, meaning that it's necessary to return the result to the client right after replication is triggered. Currently, we return replication result in PartitionReplicaListener#applyCmdWithExceptionHandling and await it in ReplicaManager#onReplicaMessageReceive
CompletableFuture<?> result = replica.processRequest(request); result.handle((res, ex) -> { ... msg = prepareReplicaResponse(requestTimestamp, res); ... clusterNetSvc.messagingService().respond(senderConsistentId, msg, correlationId);
- And of course it's required to await all commands replication at once in pre-commit. We already have such logic in ReadWriteTransactionImpl#finish
protected CompletableFuture<Void> finish(boolean commit) { ... CompletableFuture<Void> mainFinishFut = CompletableFuture .allOf(enlistedResults.toArray(new CompletableFuture[0])) .thenCompose( ... return txManager.finish( ...
however, it should use not the result from primary, but the replication completion one.
Implementation Notes
I believe it's possible to implement it in a following way:
- ReplicaManager should await only primary related actions like lock acquisition and store the replication future in a sort of map. It's possible to use safeTime as request Id.
- Transaction should send replicationAwaitRequest in an async manner right after replicationResponse from primary was achieved.
- enlistedResults should be switched to replicationAwaitResponse.
Attachments
Issue Links
- links to