Anoop Sam John created HADOOP-17308:
---------------------------------------
Summary: WASB : PageBlobOutputStream succeeding flush even when
underlying flush to storage failed
Key: HADOOP-17308
URL: https://issues.apache.org/jira/browse/HADOOP-17308
Project: Hadoop Common
Issue Type: Bug
Reporter: Anoop Sam John
Assignee: Anoop Sam John
In PageBlobOutputStream, write() APIs will fill the buffer and
hflush/hsync/flush call will flush the buffer to underlying storage. Here the
Azure calls are handled in another thread
{code}
private synchronized void flushIOBuffers() {
...
lastQueuedTask = new WriteRequest(outBuffer.toByteArray());
ioThreadPool.execute(lastQueuedTask);
....
}
private class WriteRequest implements Runnable {
private final byte[] dataPayload;
private final CountDownLatch doneSignal = new CountDownLatch(1);
public WriteRequest(byte[] dataPayload) {
this.dataPayload = dataPayload;
}
public void waitTillDone() throws InterruptedException {
doneSignal.await();
}
@Override
public void run() {
try {
LOG.debug("before runInternal()");
runInternal();
LOG.debug("after runInternal()");
} finally {
doneSignal.countDown();
}
}
private void runInternal() {
......
writePayloadToServer(rawPayload);
...........
}
private void writePayloadToServer(byte[] rawPayload) {
......
try {
blob.uploadPages(wrapperStream, currentBlobOffset, rawPayload.length,
withMD5Checking(), PageBlobOutputStream.this.opContext);
} catch (IOException ex) {
lastError = ex;
} catch (StorageException ex) {
lastError = new IOException(ex);
}
if (lastError != null) {
LOG.debug("Caught error in
PageBlobOutputStream#writePayloadToServer()");
}
}
}
{code}
The flushing thread will wait for the other thread to complete the Runnable
WriteRequest. Thats fine. But when some exception happened while
blob.uploadPages, we just set that to lastError state variable. This variable
is been checked for all subsequent ops like write, flush etc. But what about
the current flush call? that is silently being succeeded.!!
In standard Azure backed HBase clusters WAL is on page blob. This issue causes
a serious issue in HBase and causes data loss! HBase think a WAL write was
hflushed and make row write successful. In fact the row was never gone to
storage.
Checking the lastError variable at the end of flush op will solve the issue.
Then we will throw IOE from this flush() itself.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
