adjusted phrase to be more clear and meaningful in PT #585
Conversation
This file contains hidden or 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
…iles for HDFS directories with many files ## What changes were proposed in this pull request? InMemoryFileIndex.listLeafFiles should use listLocatedStatus for DistributedFileSystem. DistributedFileSystem overrides the listLocatedStatus method in order to do it with 1 single namenode call thus saving thousands of calls to getBlockLocations. Currently in InMemoryFileIndex, all directory listings are done using FileSystem.listStatus following by individual calls to FileSystem.getFileBlockLocations. This is painstakingly slow for folders that have large numbers of files because this process happens serially and parallelism is only applied at the folder level, not the file level. FileSystem also provides another API listLocatedStatus which returns the LocatedFileStatus objects that already have the block locations. In FileSystem main class this just delegates to listStatus and getFileBlockLocations similarly to the way Spark does it. However when HDFS specifically is the backing file system, DistributedFileSystem overrides this method and simply makes one single call to the namenode to retrieve the directory listing with the block locations. This avoids potentially thousands or more calls to namenode and also is more consistent because files will either exist with locations or not exist instead of having the FileNotFoundException exception case. For our example directory with 6500 files, the load time of spark.read.parquet was reduced 96x from 76 seconds to .8 seconds. This savings only goes up with the number of files in the directory. In the pull request instead of using this method always which could lead to a FileNotFoundException that could be tough to decipher in the default FileSystem implementation, this method is only used when the FileSystem is a DistributedFileSystem and otherwise the old logic still applies. ## How was this patch tested? test suite ran Closes #24672 from rrusso2007/master. Authored-by: rrusso2007 <[email protected]> Signed-off-by: Dongjoon Hyun <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
…iles for HDFS directories with many files
What changes were proposed in this pull request?
InMemoryFileIndex.listLeafFiles should use listLocatedStatus for DistributedFileSystem. DistributedFileSystem overrides the listLocatedStatus method in order to do it with 1 single namenode call thus saving thousands of calls to getBlockLocations.
Currently in InMemoryFileIndex, all directory listings are done using FileSystem.listStatus following by individual calls to FileSystem.getFileBlockLocations. This is painstakingly slow for folders that have large numbers of files because this process happens serially and parallelism is only applied at the folder level, not the file level.
FileSystem also provides another API listLocatedStatus which returns the LocatedFileStatus objects that already have the block locations. In FileSystem main class this just delegates to listStatus and getFileBlockLocations similarly to the way Spark does it. However when HDFS specifically is the backing file system, DistributedFileSystem overrides this method and simply makes one single call to the namenode to retrieve the directory listing with the block locations. This avoids potentially thousands or more calls to namenode and also is more consistent because files will either exist with locations or not exist instead of having the FileNotFoundException exception case.
For our example directory with 6500 files, the load time of spark.read.parquet was reduced 96x from 76 seconds to .8 seconds. This savings only goes up with the number of files in the directory.
In the pull request instead of using this method always which could lead to a FileNotFoundException that could be tough to decipher in the default FileSystem implementation, this method is only used when the FileSystem is a DistributedFileSystem and otherwise the old logic still applies.
How was this patch tested?
test suite ran
Closes apache#24672 from rrusso2007/master.
Authored-by: rrusso2007 [email protected]
Signed-off-by: Dongjoon Hyun [email protected]
What changes were proposed in this pull request?
(Please fill in changes proposed in this fix)
How was this patch tested?
(Please explain how this patch was tested. E.g. unit tests, integration tests, manual tests)
(If this patch involves UI changes, please attach a screenshot; otherwise, remove this)
Please review https://spark.apache.org/contributing.html before opening a pull request.