[
https://issues.apache.org/jira/browse/HBASE-28850?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17882776#comment-17882776
]
Andrew Kyle Purtell edited comment on HBASE-28850 at 9/18/24 6:04 PM:
----------------------------------------------------------------------
I am not saying this approach is wrong but it will change the performance of
replicateEntries. Before, the replicateEntries call will return as soon as the
local edits on the sink are _scheduled_ for application. After, the
replicateEntries will block for the entire time it takes to confirm all local
edits in the batch are applied, including waiting for retries and backoff if
there is local server unavailability. And that may cause a reduction in
replication throughput because of the new backpressure on the source, as we
block in replicateEntries for what I assume will be a longer time on average.
On the other hand, it can be a reasonable design decision to not return
"success" to the source unless all edits in the batch are confirmed to be
applied.
It would make failure handling more reliable because we may miss local
exceptions, and return "success" back to the source prematurely, if we are not
waiting for all futures to complete.
was (Author: apurtell):
I am not saying this approach is wrong but it will change the performance of
replicateEntries. Before, the replicateEntries call will return as soon as the
local edits on the sink are _scheduled_ for application. After, the
replicateEntries will block for the entire time it takes to confirm all local
edits in the batch are applied, including waiting for retries and backoff if
there is local server unavailability. And that may cause a reduction in
replication throughput because of the new backpressure on the source, as we
block in replicateEntries for what I assume will be a longer time on average.
On the other hand, it can be a reasonable design decision to not return
"success" to the source unless all edits in the batch are confirmed to be
applied. It would make failure handling more reliable because we may miss
exceptions and not fail the batch back to the source if we are not waiting for
all futures to complete.
> Only return from ReplicationSink.replicationEntries while all background
> tasks are finished
> -------------------------------------------------------------------------------------------
>
> Key: HBASE-28850
> URL: https://issues.apache.org/jira/browse/HBASE-28850
> Project: HBase
> Issue Type: Improvement
> Components: Replication, rpc
> Reporter: Duo Zhang
> Assignee: Duo Zhang
> Priority: Major
> Labels: pull-request-available
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
