Normally one patch should fix one problem.

+    } catch(IOException e){
+      throw e;

I saw the above code twice in the patch.
Can you explain why it is needed ?

Please also run test suite and tell us the results.

@Ted Yu
OK, I attached the patch again,and also provide the test suite results.

Is it possible to write a testcase with the scenario you have told in the description?
Minor comment:

+        } 
+        else {
+          this.server.abort("Abort; we got an error after point-of-no-return");
+        }  

+ } else

else block should begin in the same line where if block ends.

Apache Jenkins build for 0.90 has 678 tests. The attached test report had fewer test cases.

Also, patch for 0.92/TRUNK is needed since the fix would go to those places as well.

@Ted Yu
I'll check the tests...
Also I'll attach the patch for 0.92 & TRUNK.

@Ted Yu
I count all tests in src/test,I'm sure all test are ran...

Please run patch for 0.92 or TRUNK (hopefully both) through unit tests.
Thanks

I'm +1 on this patch if tests pass. Its more conservative than what we currently have. Comments in code speculated that we should probably tend to the more conservative. Thanks for making the change bluedavy.

I attached patches & test reports for 0.90.4,0.92 and trunk.

+1 on patches.

Are the attached patches the full story?
I see only PONR moved, but no code to abort the server if timing out against .META. Where is that code?

In the trunk patch the comment seems to be mangled. The big PONR comment be moved above the comment speculating the that PONR should be moved here, just like in the 0.90 and 0.92 patches...

In fact should that comment now be removed in all patches?

If OfflineParentInMeta() times out, SplitRequest.run() would execute the following code:

          if (st.rollback(this.server, this.server)) {
            LOG.info("Successful rollback of failed split of " +
              parent.getRegionNameAsString());
          } else {
            this.server.abort("Abort; we got an error after point-of-no-return");

I agree that the comments should be consistent in all patches.

I see... Thanks Ted.

I fix the comments to keep consistent in all patches.

+1 for latest patches (assuming all tests pass)

wait for committer commit to the svn.

Committing this too.

0.90 patch fails to apply. @bluedavy, could you double check all patches against the latest of the 0.90, 0.92, trunk, respectively?

the patch-0.90 is for 0.90.4...

@bluedavy:
Please generate patch for latest 0.90 and rerun the test suite.

Thanks

em,OK,I renamed the current patch for 0.90.4.

@Lars
I attached the patch for latest 0.90,pls apply it again & commit,thks.

Committed to 0.90, 0.92, and trunk.

The 0.90 patch still did not apply to the current 0.90 branch. I applied the changes manually this time, but in the future it would be great to base patches of the latests state of the branch in SVN.

em,thks.

No problem. Thanks for the patch

Integrated in HBase-TRUNK #2333 (See https://builds.apache.org/job/HBase-TRUNK/2333/)
HBASE-4562 When split doing offlineParentInMeta encounters error, it'll cause data loss

larsh :
Files :

• /hbase/trunk/CHANGES.txt
• /hbase/trunk/src/main/java/org/apache/hadoop/hbase/regionserver/SplitTransaction.java

Integrated in HBase-0.92 #72 (See https://builds.apache.org/job/HBase-0.92/72/)
HBASE-4562 When split doing offlineParentInMeta encounters error, it'll cause data loss

larsh :
Files :

• /hbase/branches/0.92/CHANGES.txt
• /hbase/branches/0.92/src/main/java/org/apache/hadoop/hbase/regionserver/SplitTransaction.java