klsince commented on a change in pull request #7969:
URL: https://github.com/apache/pinot/pull/7969#discussion_r783327022



##########
File path: 
pinot-core/src/main/java/org/apache/pinot/core/data/manager/BaseTableDataManager.java
##########
@@ -493,12 +445,139 @@ protected File getTmpSegmentDataDir(String segmentName) {
     return new File(_resourceTmpDir, segmentName);
   }
 
-  @VisibleForTesting
-  static boolean isNewSegment(SegmentZKMetadata zkMetadata, @Nullable 
SegmentMetadata localMetadata) {
-    return localMetadata == null || !hasSameCRC(zkMetadata, localMetadata);
+  /**
+   * Create a backup directory to handle failure of segment reloading.
+   * First rename index directory to segment backup directory so that original 
segment have all file
+   * descriptors point to the segment backup directory to ensure original 
segment serves queries properly.
+   * The original index directory is restored lazily, as depending on the 
conditions,
+   * it may be restored from the backup directory or segment downloaded from 
deep store.
+   */
+  private void createBackup(File indexDir) {
+    if (!indexDir.exists()) {
+      return;
+    }
+    File parentDir = indexDir.getParentFile();
+    File segmentBackupDir = new File(parentDir, indexDir.getName() + 
CommonConstants.Segment.SEGMENT_BACKUP_DIR_SUFFIX);
+    // Rename index directory to segment backup directory (atomic).
+    Preconditions.checkState(indexDir.renameTo(segmentBackupDir),
+        "Failed to rename index directory: %s to segment backup directory: 
%s", indexDir, segmentBackupDir);
+  }
+
+  /**
+   * Remove the backup directory to mark the completion of segment reloading.
+   * First rename then delete is as renaming is an atomic operation, but 
deleting is not.
+   * When we rename the segment backup directory to segment temporary 
directory, we know the reload
+   * already succeeded, so that we can safely delete the segment temporary 
directory.
+   */
+  private void removeBackup(File indexDir)
+      throws IOException {
+    File parentDir = indexDir.getParentFile();
+    File segmentBackupDir = new File(parentDir, indexDir.getName() + 
CommonConstants.Segment.SEGMENT_BACKUP_DIR_SUFFIX);
+    if (!segmentBackupDir.exists()) {
+      return;
+    }
+    // Rename segment backup directory to segment temporary directory (atomic).
+    File segmentTempDir = new File(parentDir, indexDir.getName() + 
CommonConstants.Segment.SEGMENT_TEMP_DIR_SUFFIX);
+    Preconditions.checkState(segmentBackupDir.renameTo(segmentTempDir),
+        "Failed to rename segment backup directory: %s to segment temporary 
directory: %s", segmentBackupDir,
+        segmentTempDir);
+    FileUtils.deleteDirectory(segmentTempDir);
+  }
+
+  private boolean tryLoadExistingSegment(String segmentName, 
IndexLoadingConfig indexLoadingConfig,
+      SegmentZKMetadata zkMetadata) {
+    // Try to recover the segment from potential segment reloading failure.
+    File indexDir = getSegmentDataDir(segmentName);
+    recoverReloadFailureQuietly(_tableNameWithType, segmentName, indexDir);
+
+    // Creates the SegmentDirectory object to access the segment metadata.
+    // The metadata is null if the segment doesn't exist yet.
+    SegmentDirectory segmentDirectory = tryInitSegmentDirectory(segmentName, 
indexLoadingConfig);

Review comment:
       You're right about the call stack ^ when loading a branch new segment.
   
   When server loads a brand new segment, there is no indexDir on disk yet, so 
SegmentLocalFSDirectory is initialized with an empty dir (no warning emit). But 
with an empty SegmentLocalFSDirectory object, we get to L502/L508 to return 
false from tryLoadExistingSegment() method. Then, the addOrReplaceSegment() 
method continues to download the new segment from deep store to finish loading. 
   
   This keeps the original logic flow intact. If you check the original code, 
it does recoverSegmentQuietly and loadSegmentQuietly, and if they fails, then 
go to download segment from deep store to finish loading.  




-- 
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]



---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to