[#1594] improvement(client):support generating larger block size during shuffle map task by spill partial partitions data #1670
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
added 4 commits
April 28, 2024 10:35
… partial partitions data
… partial partitions data
# Conflicts: # client-spark/common/src/main/java/org/apache/spark/shuffle/writer/WriteBufferManager.java
added 6 commits
April 28, 2024 14:13
# Conflicts: # client-spark/common/src/main/java/org/apache/spark/shuffle/writer/WriteBufferManager.java
… partial partitions data
… partial partitions data
|
@rickyma Could you help me review this pull request? |
|
Could you paste some test results to the community for this feature? |
rickyma
reviewed
Apr 29, 2024
| List<Integer> partitionList = | ||
| new ArrayList<Integer>() { | ||
| { | ||
| addAll(buffers.keySet()); |
There was a problem hiding this comment.
I think this can be simplified as:
List<Integer> partitionList = new ArrayList<>(buffers.keySet());
| } | ||
| }; | ||
| if (bufferSpillRatio < 1.0) { | ||
| Collections.sort( |
There was a problem hiding this comment.
This can be simplified as:
partitionList.sort(Comparator.comparingInt(o -> buffers.get(o) == null ? 0 : buffers.get(o).getMemoryUsed()).reversed());
| @@ -316,6 +350,10 @@ public synchronized List<ShuffleBlockInfo> clear() { | |||
| + dataSize | |||
| + "], memoryUsed[" | |||
| + memoryUsed | |||
| + "],number of blocks[" | |||
There was a problem hiding this comment.
Nit, add a space here, for a better log output:
],number of blocks[ -> ], number of blocks[
| @@ -316,6 +350,10 @@ public synchronized List<ShuffleBlockInfo> clear() { | |||
| + dataSize | |||
| + "], memoryUsed[" | |||
| + memoryUsed | |||
| + "],number of blocks[" | |||
| + result.size() | |||
| + "],flush ratio[" | |||
There was a problem hiding this comment.
Nit, add a space here, for a better log output:
],flush ratio[ -> ], flush ratio[
| LOG.info( | ||
| String.format( | ||
| "ShuffleBufferManager spill for buffer size exceeding spill threshold," | ||
| + "usedBytes[%d],inSendListBytes[%d],spill size threshold[%d]", |
There was a problem hiding this comment.
Nit, a better log output:
" usedBytes[%d], inSendListBytes[%d], spill size threshold[%d]",
added 4 commits
April 29, 2024 11:06
… partial partitions data
jerqi
reviewed
Apr 29, 2024
client-spark/common/src/main/java/org/apache/spark/shuffle/writer/WriteBufferManager.java
Outdated
Show resolved
Hide resolved
… partial partitions data
zuston
approved these changes
May 14, 2024
This was referenced May 30, 2024
zuston
added a commit
that referenced
this pull request
May 30, 2024
…1756) ### What changes were proposed in this pull request? 1. When the spill ratio is `1.0` , the process of calculating target spill size will be ignored to avoid potential race condition that the `usedBytes` and `inSendBytes` are not thread safe. This could guarantee that the all data is flushed to the shuffle server at the end of task. 2. Adding the `bufferManager's` buffer remaining check ### Why are the changes needed? Due to the #1670 , the partial data held by the bufferManager will not be flushed to shuffle servers in some corner cases, this will make task fail fast rather than silently data loss that should thanks the #1558 ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? Existing tests.
zuston
added a commit
to zuston/incubator-uniffle
that referenced
this pull request
Dec 9, 2024
…umber (apache#1756) ### What changes were proposed in this pull request? 1. When the spill ratio is `1.0` , the process of calculating target spill size will be ignored to avoid potential race condition that the `usedBytes` and `inSendBytes` are not thread safe. This could guarantee that the all data is flushed to the shuffle server at the end of task. 2. Adding the `bufferManager's` buffer remaining check ### Why are the changes needed? Due to the apache#1670 , the partial data held by the bufferManager will not be flushed to shuffle servers in some corner cases, this will make task fail fast rather than silently data loss that should thanks the apache#1558 ### Does this PR introduce _any_ user-facing change? No. ### How was this patch tested? Existing tests.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
when spilling shuffle data, we just spill part of the reduce partition datas which hold the major space.
so, in each spilling process, the WriteBufferManager.clear() method should implement one more logic: sort the to-be spilled buffers by their size and select the top-N buffers to spill.
Why are the changes needed?
related feature #1594
Does this PR introduce any user-facing change?
No.
How was this patch tested?
new UTs.