You are viewing a plain text version of this content. The canonical link for it is here.

Posted to issues@arrow.apache.org by "Ziheng Wang (Jira)" <ji...@apache.org> on 2022/08/03 23:24:00 UTC

[jira] [Created] (ARROW-17299) Expose the Scanner kDefaultBatchReadahead and kDefaultFragmentReadahead parameters

Ziheng Wang created ARROW-17299:
-----------------------------------

             Summary: Expose the Scanner kDefaultBatchReadahead and kDefaultFragmentReadahead parameters
                 Key: ARROW-17299
                 URL: https://issues.apache.org/jira/browse/ARROW-17299
             Project: Apache Arrow
          Issue Type: New Feature
          Components: C++, Python
            Reporter: Ziheng Wang
            Assignee: Ziheng Wang


In the Scanner there are parameters kDefaultFragmentReadahead and kDefaultBatchReadahead that are currently set to fixed numbers that cannot be changed.

This is not great because tuning these numbers is the key to tradeoff RAM usage and network IO utilization during reading. For example on an i3.2xlarge instance on AWS you can get peak throughput only by quadrupling kDefaultFragmentReadahead from the default. 

The current settings are very conservative and assume a < 1Gbps network. Exposing them allow people to tune the Scanner behavior to their own hardware. 



--
This message was sent by Atlassian Jira
(v8.20.10#820010)