Wang Xixu has posted comments on this change. ( 
http://gerrit.cloudera.org:8080/20680 )

Change subject: KUDU-3523 Use f_bsize as Kudu block size instead of st_blksize
......................................................................


Patch Set 4:

> Patch Set 3:
>
> Generally, both st_blksize and f_bsize are same when we are working on a 
> native filesystems on linux e.g. ext3/ext4.
>
> The use case that you have mentioned here seems to be unusual. I noticed the 
> filesystem type that you are using is of type overlayfs which IIUC is a 
> unionfs/layeredfs that probably had layers of different filesystem type. I 
> can't say for sure, but there is a possibility that in layeredfs, the 
> st_blksize may differ from the f_bsize depending on which filesystem driver 
> is used for get that information. I am planning to test this out on local 
> setup. Will try to share the details once I complete the test or you can 
> update here if you are already aware of it.
> 
> I am very much interested in knowing more about the setup you are using, what 
> all actual filesystems are constituting this setup, purpose of having a 
> overlayfs setup, directory structure used in overlayfs setup, etc.
>
> LBM needs hole punching support from underlying filesystem (ext4 does provide 
> that currently). However, I am not quite clear on what it would mean if Kudu 
> data is residing on overlayfs that has multiple different filesystems 
> underneath it and there is one filesystem that doesn't support hole-punching.
> Maybe it would be worth an exercise to evaluate behaviour of POSIX APIs (used 
> in Kudu) on overlayfs mount-point.
>
> IIUC, overlayfs is just an abstraction layer and all the requests go to 
> targeted underlying file-system so it shouldn't matter much but it is worth a 
> try though because what all underlying filesystems are supported for Kudu is 
> something that is not clearly visible when it is an overlayfs setup.
>
> TL;DR: You could provide more information on the following points:
> 1. Provide some more information about the directory setup and how it uses 
> overlayfs.
> 2. Check if all the underlying filesystems support hole punching (if Kudu is 
> using LBM).
> 3. Purpose of using overlayfs.
> 4. Can you try reproduction with same steps on a regular underlying 
> filesystem e.g. ext4 (just to rule out other possibilities)?

I reproduction it on a XFS filesystem, st_blkszie is 65536 and f_bsize is 4096.
It seems that this case is not unique for overlay filesystem. Please see the 
issue: #KUDU-3523.


-- 
To view, visit http://gerrit.cloudera.org:8080/20680
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings

Gerrit-Project: kudu
Gerrit-Branch: master
Gerrit-MessageType: comment
Gerrit-Change-Id: I04343caf5fe377a4fa9b57e6e450307e6b995928
Gerrit-Change-Number: 20680
Gerrit-PatchSet: 4
Gerrit-Owner: Wang Xixu <[email protected]>
Gerrit-Reviewer: Alexey Serbin <[email protected]>
Gerrit-Reviewer: Ashwani Raina <[email protected]>
Gerrit-Reviewer: Attila Bukor <[email protected]>
Gerrit-Reviewer: Kudu Jenkins (120)
Gerrit-Reviewer: Wang Xixu <[email protected]>
Gerrit-Reviewer: Yifan Zhang <[email protected]>
Gerrit-Reviewer: Yingchun Lai <[email protected]>
Gerrit-Comment-Date: Thu, 07 Dec 2023 02:49:11 +0000
Gerrit-HasComments: No

Reply via email to