doris
doris copied to clipboard
apache
[Optimize](Random distribution) Improve the performance of tablet sin…
…k and delta writer of writing blocks
The current distribution model for Doris is as follows:
OlapTableSink seperate the original Block into serveral subblocks of each node(BE) by tablets distribution and distributes subblocks to storage engine of backends, then the storage engine will seperate the subblock into multiple tablets channel and each delta writer will handle partial of the block.
This model causes blocks to be split according to tablets, and the splitting process can be a relatively heavy operation. After splitting, the blocks are distributed to different DeltaWriters (Memtables) through RPCs to TabletChannels. The distribution operation on TabletChannels is also a relatively heavy operation. If the distribution property of the table is RANDOM distribution, then we have the opportunity to distribute the blocks according to the complete block during distribution. The advantage of doing so is to reduce memory copying and improve write locality, similar to appending the entire block to the memtable.
This optimze could save 10% ~ 20% CPU cost of RANDOM distribution table load when enable load_to_single_tablet Whats more, is that we could even write to the local delta writer from OlapTableSink in the single_replica_load mode
Proposed changes
Issue Number: close #17388
Problem summary
Describe your changes.
Checklist(Required)
- [ ] Does it affect the original behavior
- [ ] Has unit tests been added
- [ ] Has document been added or modified
- [ ] Does it need to update dependencies
- [ ] Is this PR support rollback (If NO, please explain WHY)
Further comments
If this is a relatively large or complex change, kick off the discussion at [email protected] by explaining why you chose the solution you did and what alternatives you considered, etc...
![github-actions[bot] avatar](https://avatars.githubusercontent.com/in/15368?v=4 "Published by a pub.dev verified publisher"){kind=small}
TeamCity pipeline, clickbench performance test result: the sum of best hot time: 33.53 seconds stream load tsv: 460 seconds loaded 74807831229 Bytes, about 155 MB/s stream load json: 40 seconds loaded 2358488459 Bytes, about 56 MB/s stream load orc: 74 seconds loaded 1101869774 Bytes, about 14 MB/s stream load parquet: 31 seconds loaded 861443392 Bytes, about 26 MB/s https://doris-community-test-1308700295.cos.ap-hongkong.myqcloud.com/tmp/20230306101319_clickbench_pr_109289.html
