hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Lars George (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HBASE-1969) HBASE-1626 does not work as advertised due to lack of "instanceof" check in MR framework
Date Thu, 19 Nov 2009 18:57:39 GMT

    [ https://issues.apache.org/jira/browse/HBASE-1969?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12780199#action_12780199
] 

Lars George commented on HBASE-1969:
------------------------------------

I thought about this. Could we create a class Change that is the base for Put and Delete and
implements a Writable as well, which only carries the changes, or in other words KeyValue's
with only as much additional meta data that is required to batch execute them on a table?

Then we would have to either 

a.1) add a Put.getChange() / Delete.getChange() 

a.2) implement Change as the base class for Put and Delete

a.3) have a Change change = new Change(put) so you do context.write(key, new Change(put));

Then on the table we 

b.1) either recreate the base classes from the Change instance and apply them as is done right
now

b.2) add support for Change to the HTable class so it can batch it and with the KV types knows
what to do.


Alternatively we can add a Change() class only to the mapreduce package that takes a Put or
Delete as the constructor so you also do

context.write(key, new Change(put));

and it transports the payload (a Put or Delete) to the TableOutputFormat where it is then
converted back and applied. Seems the least intrusive alternative.

What say you?



> HBASE-1626 does not work as advertised due to lack of "instanceof" check in MR framework
> ----------------------------------------------------------------------------------------
>
>                 Key: HBASE-1969
>                 URL: https://issues.apache.org/jira/browse/HBASE-1969
>             Project: Hadoop HBase
>          Issue Type: Bug
>    Affects Versions: 0.20.1
>            Reporter: Lars George
>
> The issue that HBASE-1626 tried to fix is that we can hand in Put or Delete instances
to the TableOutputFormat. So the explicit Put reference was changed to Writable in the process.
But that does not work as expected:
> {code}09/11/04 13:35:56 INFO mapred.JobClient: Task Id : attempt_200911031030_0004_m_000013_2,
Status : FAILED
> java.io.IOException: Type mismatch in value from map: expected org.apache.hadoop.io.Writable,
recieved org.apache.hadoop.hbase.client.Put
>         at org.apache.hadoop.mapred.MapTask$MapOutputBuffer.collect(MapTask.java:812)
>         at org.apache.hadoop.mapred.MapTask$NewOutputCollector.write(MapTask.java:504)
>         at org.apache.hadoop.mapreduce.TaskInputOutputContext.write(TaskInputOutputContext.java:80)
>         at com.worldlingo.hadoop.mapred.RestoreTable$RestoreMapper.map(RestoreTable.java:140)
>         at com.worldlingo.hadoop.mapred.RestoreTable$RestoreMapper.map(RestoreTable.java:69)
>         at org.apache.hadoop.mapreduce.Mapper.run(Mapper.java:144)
>         at org.apache.hadoop.mapred.MapTask.runNewMapper(MapTask.java:583)
>         at org.apache.hadoop.mapred.MapTask.run(MapTask.java:305){code}
> The issue is that the MapReduce framework checks not polymorphic for the type using "instanceof"
but with a direct class comparison. In MapTask.java you find this code
> {code}
>     public synchronized void collect(K key, V value, int partition
>                                      ) throws IOException {
>       reporter.progress();
>       if (key.getClass() != keyClass) {
>         throw new IOException("Type mismatch in key from map: expected "
>                               + keyClass.getName() + ", recieved "
>                               + key.getClass().getName());
>       }
>       if (value.getClass() != valClass) {
>         throw new IOException("Type mismatch in value from map: expected "
>                               + valClass.getName() + ", recieved "
>                               + value.getClass().getName());
>       }
>       ... {code}
> So it does not work using a Writable as the MapOutputValueClass for the job and then
hand in a Put or Delete! The test case TestMapReduce did not pick this up as it has this line
in it
> {code}
>       TableMapReduceUtil.initTableMapperJob(
>         Bytes.toString(table.getTableName()), scan,
>         ProcessContentsMapper.class, ImmutableBytesWritable.class, 
>         Put.class, job);{code}
> which sets the value class to Put
> {code}if (outputValueClass != null) job.setMapOutputValueClass(outputValueClass);{code}
> To fix this (for now) one can set the class to Put the same way or explicitly in their
code 
> {code}job.setMapOutputValueClass(Put.class);{code}
>  
> But the whole idea only seems feasable if a) the Hadoop class is amended to use "instanceof"
instead (lodge Hadoop MapRed JIRA issue?) or b) we have a combined class that represent a
Put *and* a Delete - which seems somewhat wrong, but doable. It would only really find use
in that context and would require the user to make use of it when calling context.write().
This is making things not easier to learn.
> Suggestions?

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message