[jira] [Commented] (HIVE-8457) MapOperator initialization fails when multiple Spark threads is enabled [Spark Branch]

["Chao (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-8457?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14182221#comment-14182221 ] 

Chao commented on HIVE-8457:
----------------------------

There are two problems I found:
1. one thread may be reused for multiple queries, and therefore, if one {{IOContext}} is bound to one thread, and the {{IOContext}} is not initialized, we need to clear it when we close the map record handler;
2. synchronization is needed - the thread who doesn't have {{IOContext}} initialized may try to retrieve before the thread who have {{IOContext}} initialized saving it. Right now, a thread will save {{IOContext}} right after initializing it. But, this is not a perfect solution.

> MapOperator initialization fails when multiple Spark threads is enabled [Spark Branch]
> --------------------------------------------------------------------------------------
>
>                 Key: HIVE-8457
>                 URL: https://issues.apache.org/jira/browse/HIVE-8457
>             Project: Hive
>          Issue Type: Bug
>          Components: Spark
>            Reporter: Chao
>            Assignee: Chao
>         Attachments: HIVE-8457.1-spark.patch
>
>
> Currently, on the Spark branch, each thread it is bound with a thread-local IOContext, which gets initialized when we generates an input {{HadoopRDD}}, and later used in {{MapOperator}}, {{FilterOperator}}, etc.
> And, given the introduction of HIVE-8118, we may have multiple downstream RDDs that share the same input {{HadoopRDD}}, and we would like to have the {{HadoopRDD}} to be cached, to avoid scanning the same table multiple times. A typical case would be like the following:
> {noformat}
>      inputRDD     inputRDD
>         |            |
>        MT_11        MT_12
>         |            |
>        RT_1         RT_2
> {noformat}
> Here, {{MT_11}} and {{MT_12}} are {{MapTran}} from a splitted {{MapWork}},
> and {{RT_1}} and {{RT_2}} are two {{ReduceTran}}. Note that, this example is simplified, as we may also have {{ShuffleTran}} between {{MapTran}} and {{ReduceTran}}.
> When multiple Spark threads are running, {{MT_11}} may be executed first, and it will ask for an iterator from the {{HadoopRDD}} will trigger the creation of the iterator, which in turn triggers the initialization of the {{IOContext}} associated with that particular thread.
> *Now, the problem is*: before {{MT_12}} starts executing, it will also ask for an iterator from the
> {{HadoopRDD}}, and since the RDD is already cached, instead of creating a new iterator, it will just fetch it from the cached result. However, *this will skip the initialization of the IOContext associated with this particular thread*. And, when {{MT_12}} starts executing, it will try to initialize the {{MapOperator}}, but since the {{IOContext}} is not initialized, this will fail miserably. 



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