[
https://issues.apache.org/jira/browse/HADOOP-10669?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Mikhail Bernadsky updated HADOOP-10669:
---------------------------------------
Description:
Found this debugging Nutch.
MapTask serializes keys and values to the same stream, in pairs:
keySerializer.serialize(key);
.....
valSerializer.serialize(value);
.....
bb.write(b0, 0, 0);
AvroSerializer does not flush its buffer after each serialization. So if it is
used for valSerializer, the values are only partially written or not written at
all to the output stream before the record is marked as complete (the last line
above).
<EDIT> Added HADOOP-10699_all.patch. This is a less intrusive fix, as it does
not try to flush MapTask stream. Instead, we write serialized values directly
to MapTask stream and avoid using a buffer on avro side.
was:
Found this debugging Nutch.
MapTask serializes keys and values to the same stream, in pairs:
keySerializer.serialize(key);
.....
valSerializer.serialize(value);
.....
bb.write(b0, 0, 0);
AvroSerializer does not flush its buffer after each serialization. So if it is
used for valSerializer, the values are only partially written or not written at
all to the output stream before the record is marked as complete (the last line
above).
> Avro serialization does not flush buffered serialized values causing data lost
> ------------------------------------------------------------------------------
>
> Key: HADOOP-10669
> URL: https://issues.apache.org/jira/browse/HADOOP-10669
> Project: Hadoop Common
> Issue Type: Bug
> Components: io
> Affects Versions: 2.4.0
> Reporter: Mikhail Bernadsky
> Attachments: HADOOP-10669.patch, HADOOP-10669_alt.patch
>
>
> Found this debugging Nutch.
> MapTask serializes keys and values to the same stream, in pairs:
> keySerializer.serialize(key);
> .....
> valSerializer.serialize(value);
> .....
> bb.write(b0, 0, 0);
> AvroSerializer does not flush its buffer after each serialization. So if it
> is used for valSerializer, the values are only partially written or not
> written at all to the output stream before the record is marked as complete
> (the last line above).
> <EDIT> Added HADOOP-10699_all.patch. This is a less intrusive fix, as it does
> not try to flush MapTask stream. Instead, we write serialized values directly
> to MapTask stream and avoid using a buffer on avro side.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
