[jira] Updated: (CASSANDRA-789) Add configurable range sizes, paging to hadoop range queries

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

     [ https://issues.apache.org/jira/browse/CASSANDRA-789?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Johan Oskarsson updated CASSANDRA-789:
--------------------------------------

    Attachment: CASSANDRA-789.patch

Allows the record reader to fetch data in small batches, to avoid timeouts that a large get_range_slice request can create. 
This means users can increase the number of rows per input split to a number that utilizes the cluster well without the overhead of many mappers. 

> Add configurable range sizes, paging to hadoop range queries
> ------------------------------------------------------------
>
>                 Key: CASSANDRA-789
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-789
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Core
>            Reporter: Jonathan Ellis
>            Priority: Minor
>         Attachments: CASSANDRA-789.patch
>
>
> For very large (billions) numbers of keys, the current hardcoded 4096 keys per InputSplit could cause the split generator to OOM, since all splits are held in memory at once.  So we want to make 2 changes:
>  1) make the number of keys configurable*
>  2) make record reader page instead of assuming it can read all rows into memory at once
> Note: going back to specifying number of splits instead of number of keys is bad for two reasons.  First, it does not work with the standard hadoop mapred.min.split.size configuration option.  Second, it means we have no way of measuring progress in the record reader, since we have no idea how many keys are in the split.  If we specify number of keys, then even if we page we know (to within a small margin of error) how many keys to expect, even if we page.
> See CASSANDRA-775, CASSANDRA-342 for background.

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