[
https://issues.apache.org/jira/browse/ARROW-3781?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16685727#comment-16685727
]
Antoine Pitrou commented on ARROW-3781:
---------------------------------------
I don't think it's dependent on filesystem latency. Unless the filesystem
implementation is broken, writing should be asynchronous (i.e. the `Write` call
returns before the OS actually flushed the buffer to disk or to the network).
The point of the buffer is to avoid paying the cost of a system call (and
userspace/kernel transition) for every tiny write.
But we can make the buffer size configurable regardless.
> [C++] Configure buffer size in arrow::io::BufferedOutputStream
> --------------------------------------------------------------
>
> Key: ARROW-3781
> URL: https://issues.apache.org/jira/browse/ARROW-3781
> Project: Apache Arrow
> Issue Type: Improvement
> Components: C++
> Reporter: Wes McKinney
> Priority: Major
> Fix For: 0.12.0
>
>
> This is hard-coded to 4096 right now. For higher latency file systems it may
> be desirable to use a larger buffer. See also ARROW-3777 about performance
> testing for high latency files
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
