Vihang Karajgaonkar created HIVE-21040:
------------------------------------------
Summary: msck does unnecessary file listing at last level of
partitions
Key: HIVE-21040
URL: https://issues.apache.org/jira/browse/HIVE-21040
Project: Hive
Issue Type: Improvement
Reporter: Vihang Karajgaonkar
Assignee: Vihang Karajgaonkar
Here is the code snippet which is run by {{msck}} to list directories
{noformat}
final Path currentPath = pd.p;
final int currentDepth = pd.depth;
FileStatus[] fileStatuses = fs.listStatus(currentPath,
FileUtils.HIDDEN_FILES_PATH_FILTER);
// found no files under a sub-directory under table base path; it is
possible that the table
// is empty and hence there are no partition sub-directories created
under base path
if (fileStatuses.length == 0 && currentDepth > 0 && currentDepth <
partColNames.size()) {
// since maxDepth is not yet reached, we are missing partition
// columns in currentPath
logOrThrowExceptionWithMsg(
"MSCK is missing partition columns under " +
currentPath.toString());
} else {
// found files under currentPath add them to the queue if it is a
directory
for (FileStatus fileStatus : fileStatuses) {
if (!fileStatus.isDirectory() && currentDepth < partColNames.size()) {
// found a file at depth which is less than number of partition keys
logOrThrowExceptionWithMsg(
"MSCK finds a file rather than a directory when it searches for
"
+ fileStatus.getPath().toString());
} else if (fileStatus.isDirectory() && currentDepth <
partColNames.size()) {
// found a sub-directory at a depth less than number of partition
keys
// validate if the partition directory name matches with the
corresponding
// partition colName at currentDepth
Path nextPath = fileStatus.getPath();
String[] parts = nextPath.getName().split("=");
if (parts.length != 2) {
logOrThrowExceptionWithMsg("Invalid partition name " + nextPath);
} else if
(!parts[0].equalsIgnoreCase(partColNames.get(currentDepth))) {
logOrThrowExceptionWithMsg(
"Unexpected partition key " + parts[0] + " found at " +
nextPath);
} else {
// add sub-directory to the work queue if maxDepth is not yet
reached
pendingPaths.add(new PathDepthInfo(nextPath, currentDepth + 1));
}
}
}
if (currentDepth == partColNames.size()) {
return currentPath;
}
}
{noformat}
You can see that when the {{currentDepth}} at the {{maxDepth}} it still does a
unnecessary listing of the files. We can improve this call by checking the
currentDepth and bailing out early.
This can improve the performance of msck command significantly especially when
there are lot of files in each partitions on remote filesystems like S3 or ADLS
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
