[HIVE-17802] Remove unnecessary calls to FileSystem.setOwner() from FileOutputCommitterContainer

• Type: Bug
• Status: Patch Available
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.2.0, 3.0.0
• Component/s: HCatalog
• External issue ID: YHIVE-751
• Assignee: Chris Drome
• Reporter: Mithun Radhakrishnan
• Created: 13/Oct/17 18:47
• Updated: 19/Oct/17 08:53

Description

For large Pig/HCat queries that produce a large number of partitions/directories/files, we have seen cases where the HDFS NameNode groaned under the weight of FileSystem.setOwner() calls, originating from the commit-step. This was the result of the following code in FileOutputCommitterContainer:

private void applyGroupAndPerms(FileSystem fs, Path dir, FsPermission permission,
                  List<AclEntry> acls, String group, boolean recursive)
    throws IOException {
...
    if (recursive) {
      for (FileStatus fileStatus : fs.listStatus(dir)) {
        if (fileStatus.isDir()) {
          applyGroupAndPerms(fs, fileStatus.getPath(), permission, acls, group, true);
        } else {
          fs.setPermission(fileStatus.getPath(), permission);
          chown(fs, fileStatus.getPath(), group);
        }
      }
    }
  }

  private void chown(FileSystem fs, Path file, String group) throws IOException {
    try {
      fs.setOwner(file, null, group);
    } catch (AccessControlException ignore) {
      // Some users have wrong table group, ignore it.
      LOG.warn("Failed to change group of partition directories/files: " + file, ignore);
    }
  }

One call per file/directory is far too many. We have a patch that reduces the namenode pressure.

Issue Links

• is blocked by: HIVE-17803 With Pig multi-query, 2 HCatStorers writing to the same table will trample each other's outputs