hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "stack (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-14267) In Mapreduce on HBase scenario, restart in TableInputFormat will result in getting wrong data.
Date Thu, 03 Sep 2015 00:23:45 GMT

    [ https://issues.apache.org/jira/browse/HBASE-14267?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14728278#comment-14728278
] 

stack commented on HBASE-14267:
-------------------------------

 bq. User should know the data can not be modified.

I read through the javadoc and we do not clearly state this.

When you modify the row, you are altering the cache'd row instance in Result?

  // We're not using java serialization.  Transient here is just a marker to say
  // that this is where we cache row if we're ever asked for it.
  private transient byte [] row = null;

I am unclear on how this plays in the restart of the Scan? Is the modified Result row used
to calculate where new Scan restarts?

Thanks [~qianxiZhang] Looks like some dirty debugging to figure the issue.

> In Mapreduce on HBase scenario, restart in TableInputFormat will result in getting wrong
data.
> ----------------------------------------------------------------------------------------------
>
>                 Key: HBASE-14267
>                 URL: https://issues.apache.org/jira/browse/HBASE-14267
>             Project: HBase
>          Issue Type: Bug
>          Components: Client, mapreduce
>            Reporter: Qianxi Zhang
>            Assignee: Qianxi Zhang
>         Attachments: HBASE_14267_trunk_v1.patch
>
>
> When I run a mapreduce job on HBase, I will modify the row got from Result.getRow(),
for example, reverse the row. Since my program is very complicated to handle data, it takes
long time, and the lease int Region server expired. 
> Result#195
> {code}
>   public byte [] getRow() {
>     if (this.row == null) {
>       this.row = (this.cells == null || this.cells.length == 0) ?
>           null :
>           CellUtil.cloneRow(this.cells[0]);
>     }
>     return this.row;
>   }
> {code}
> TableInputFormat will restart the scan from last row, but the row has been modified,
so it will read wrong data.
> TableRecordReaderImpl#218
> {code}
>       } catch (IOException e) {
>         // do not retry if the exception tells us not to do so
>         if (e instanceof DoNotRetryIOException) {
>           throw e;
>         }
>         // try to handle all other IOExceptions by restarting
>         // the scanner, if the second call fails, it will be rethrown
>         LOG.info("recovered from " + StringUtils.stringifyException(e));
>         if (lastSuccessfulRow == null) {
>           LOG.warn("We are restarting the first next() invocation," +
>               " if your mapper has restarted a few other times like this" +
>               " then you should consider killing this job and investigate" +
>               " why it's taking so long.");
>         }
>         if (lastSuccessfulRow == null) {
>           restart(scan.getStartRow());
>         } else {
>           restart(lastSuccessfulRow);
>           scanner.next();    // skip presumed already mapped row
>         }
>         value = scanner.next();
>         if (value != null && value.isStale()) numStale++;
>         numRestarts++;
>       }
>       if (value != null && value.size() > 0) {
>         key.set(value.getRow());
>         lastSuccessfulRow = key.get();
>         return true;
>       }
> {code}



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message