Andrew Wang created HDFS-12612:
----------------------------------
Summary: DFSStripedOutputStream#close will throw if called a
second time with a failed streamer
Key: HDFS-12612
URL: https://issues.apache.org/jira/browse/HDFS-12612
Project: Hadoop HDFS
Issue Type: Bug
Components: erasure-coding
Affects Versions: 3.0.0-beta1
Reporter: Andrew Wang
Found while testing with Hive. We have a cluster with 2 DNs and the XOR-2-1
policy. If you write a file and call close() twice, it throws this exception:
{noformat}
17/10/04 16:02:14 WARN hdfs.DFSOutputStream: Cannot allocate parity
block(index=2, policy=XOR-2-1-1024k). Not enough datanodes? Exclude nodes=[]
...
Caused by: java.io.IOException: Failed to get parity block, index=2
at
org.apache.hadoop.hdfs.DFSStripedOutputStream.allocateNewBlock(DFSStripedOutputStream.java:500)
~[hadoop-hdfs-client-3.0.0-alpha3-cdh6.x-SNAPSHOT.jar:?]
at
org.apache.hadoop.hdfs.DFSStripedOutputStream.writeChunk(DFSStripedOutputStream.java:524)
~[hadoop-hdfs-client-3.0.0-alpha3-cdh6.x-SNAPSHOT.jar:?]
{noformat}
This is because in DFSStripedOutputStream#closeImpl, if the stream is closed,
we throw an exception if any of the striped streamers had an exception:
{code}
protected synchronized void closeImpl() throws IOException {
if (isClosed()) {
final MultipleIOException.Builder b = new MultipleIOException.Builder();
for(int i = 0; i < streamers.size(); i++) {
final StripedDataStreamer si = getStripedDataStreamer(i);
try {
si.getLastException().check(true);
} catch (IOException e) {
b.add(e);
}
}
final IOException ioe = b.build();
if (ioe != null) {
throw ioe;
}
return;
}
{code}
I think this is incorrect, since we only need to throw in this situation if we
have too many failed streamers. close should also be idempotent, so it should
throw the first time we call close if it's going to throw at all.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
