Ivan Andika created HDDS-15711:
----------------------------------
Summary: Streaming Write Pipeline increment container blockCount
twice
Key: HDDS-15711
URL: https://issues.apache.org/jira/browse/HDDS-15711
Project: Apache Ozone
Issue Type: Sub-task
Reporter: Ivan Andika
Assignee: Ivan Andika
Streaming close sends two EOF {{{}PutBlock{}}}s for the same block, while
normal async write sends one.
Normal async path:
* BlockOutputStream.executePutBlock only calls {{{}putBlockAsync(...){}}}.
* Close with buffered data sends one EOF {{{}PutBlock{}}}.
* Close after an earlier non-EOF putBlock sends one forced EOF
{{{}PutBlock{}}}; the earlier non-EOF putBlock has already created the block /
pending cache state, so EOF does not increment blockCount again.
Streaming path:
* BlockDataStreamOutput.executePutBlock (line 425)] first sends an EOF
putBlock through {{{}executePutBlockClose(...){}}}.
* Then the same method also calls {{putBlockAsync(xceiverClient, blockData,
close, token)}}
* On the server, the data-stream close request is dispatched via
{{{}STREAM_LINK{}}}, and the async request is dispatched as a normal
{{{}PutBlock{}}}; both reach, then {{{}persistPutBlock{}}}.
So for a small streaming write that does not hit the intermediate data-stream
flush boundary, close can race two EOF full-block commits for the same local
ID. Both can see “block does not exist” before either batch commits, and both
increment source {{{}blockCount{}}}.
I agree the broad write lock in {{persistPutBlock}} is too expensive for this.
Better fix direction: remove that lock and make streaming close not submit two
successful EOF {{{}PutBlock{}}}s for the same block. The cleanest target is
probably {{{}BlockDataStreamOutput.executePutBlock{}}}: when {{{}close ==
true{}}}, use the data-stream close putBlock as the authoritative EOF commit
and avoid the extra {{{}putBlockAsync{}}}, except for the existing fallback
path when {{executePutBlockClose}} fails.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
