fredia commented on pull request #23:
URL: https://github.com/apache/flink-benchmarks/pull/23#issuecomment-1086609117


   Thanks for the feedback! @Myasuka 
   The configurations and results for rocksdb state backend are: 
   
   <div class="lake-content" typography="classic">
   
     | rescale type | delete type | overlap threshold | time
   -- | -- | -- | -- | --
   1 | 1->2 | native deleteRange | 0 | 1811.478 ±   82.262  ms/op
   2 | 1->2 | native deleteRange | 0.75 | 17640.542 ± 1784.368  ms/op
   3 | 1->2 | scan and delete | 0 | 21790.634 ± 216.322 ms/op
   4 | 1->2 | scan and delete | 0.75 | 19595.801 ± 523.404  ms/op
   5 | 2->1 | native deleteRange | 0 | 19825.918 ± 1261.372  ms/op
   6 | 2->1 | native deleteRange | 0.75 | 20121.673 ± 1298.390  ms/op
   7 | 2->1 | scan and delete | 0 | 18628.801 ± 289.178  ms/op
   8 | 2->1 | scan and delete | 0.75 | 18520.173 ± 616.218  ms/op
   
   </div>
   
   For RocksDB state backend:
   For rescale-in cases(line5~line8), the actual overlap fraction is 1.0, 
regardless of whether the threshold is 0 or 0.75, one handle is selected as the 
init handle to initialize the base DB, and then another handle is put into the 
base DB one by one. There is nothing to delete, using `scan and delete` or 
`native deleteRange` has no effects. So, we can see the results in line5 to 
line8 are about the same.
   
   For rescale-out cases(line1~line4), the actual overlap fraction is 0.5,
   - if threshold=0, the handle(only one) is selected as the init handle to 
initialize the base DB, and then half of the handle would be deleted. In this 
situation, `native deleteRange` is better than `scan and delete`(line1 and 
line3), this also proves the performance improvement of 
[FLINK-21321](https://issues.apache.org/jira/browse/FLINK-21321).
   - if threshold=0.75, it would build an empty DB as base DB, and put half of 
the handle into base DB, thus using `scan and delete` or `native deleteRange` 
has no effects(line2 and line4).
   
   For file system state backend, I set fileStateSizeThreshold=0 and got the 
results:
   
   ```
   # Run complete. Total time: 00:17:11
   
   Benchmark                                     (rescaleType)  Mode  Cnt      
Score     Error  Units
   FsStateBackendRescalingBenchmarkExecutor.rescale     RESCALE_IN  avgt   30  
15890.784 ± 461.543  ms/op
   FsStateBackendRescalingBenchmarkExecutor.rescale    RESCALE_OUT  avgt   30   
7858.031 ± 276.433  ms/op
   ```
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to