[Starlingx-discuss] [glance] Analysis of patch c73ab48 (Enhanced-error-messages) for StartlingX upstreaming

Wed Nov 14 04:18:57 UTC 2018

Hi Stefan

Thanks for the info. The patch you mentioned is not in the upstreaming patch list. I take a look at this patch, it looks more like a workaround for the bug of ceph. It checks the available space in ceph, and goes ahead to create image only if there's enough space in ceph.

-          This patch assumed image size always be given, but this is not the case in current upstream glance. I know there're patches in cinder and glanceclient to specify image size, I'm working on upstreaming these patches. These patches can improve the uploading performance(about 30%) because no resize will happen when creating an image with exact and large enough size. But cinder is not the only client of glance, at least we cannot assume this, so we still cannot assume image size always be given. Maybe the assumption is correct before starlingx be open sourced.

-          Even if image size always be given, this patch may also cannot handle all the cases. E.g.

o   Some images be created simultaneously by glance

o   The quote of ceph pool happened to be changed

o   The ceph pool be consumed by other app though this may rarely happen

I'm not sure if these are the reasons of not triaged this patch to the upstreaming patch list.
Go back to the patch c73ab48 (Enhanced-error-messages), this assumes rbd.NoSpace exception will happen, right? So do you agree to catch this exception in rbd.py?

Regards
Liang

From: Dinescu, Stefan [mailto:Stefan.Dinescu at windriver.com]
Sent: Tuesday, November 13, 2018 11:41 PM
To: Fang, Liang A <liang.a.fang at intel.com>; Badea, Daniel <Daniel.Badea at windriver.com>
Cc: starlingx-discuss at lists.starlingx.io
Subject: RE: [Starlingx-discuss] [glance] Analysis of patch c73ab48 (Enhanced-error-messages) for StartlingX upstreaming

I think you are missing a glance_store patch for starlingx: https://github.com/openstack/stx-upstream/blob/master/openstack/python-glance-store/centos/patches/0001-Check-ceph-cluster-free-space-before-creating-image.patch

This patch added aditional checks for ceph space requirments.

I am saying this due to how you described the "self._create_image function" where it is missing a parameter added by the above patch.

In our use-case, the image-size was always passed to the rbd driver code, so the resize function (which is only called when the size is zero) was never called.

How the algorithm should work:
- grab the size of the image that is about to be uploaded (we tested this through cli and horizon)
- before uploading any data, check that ceph actually has space for that data
- only after the check passes, we actually begin writing data to ceph

We did this, because ceph actually blocked almost completely when a pool was full, requiring manual intervention to recover.

The LVM backend behaved properly in such cases, so no fix was written for that.

________________________________
From: Fang, Liang A [liang.a.fang at intel.com]
Sent: Tuesday, November 13, 2018 5:23 PM
To: Dinescu, Stefan; Badea, Daniel
Cc: starlingx-discuss at lists.starlingx.io<mailto:starlingx-discuss at lists.starlingx.io>
Subject: RE: [Starlingx-discuss] [glance] Analysis of patch c73ab48 (Enhanced-error-messages) for StartlingX upstreaming
Hi Stefan

I take a look at this issue today, the hang happened inside ceph. So I modified ceph code (function resize) to just raise exception rbd.NoSpace at the entry of the function. This issue can be simulated now.

Regarding this patch, as I said previously, rbd (ceph) is just one type of the multiple types of backend, so I prefer to catch the rbd.NoSpace in glance_store (rbd.py), handle the exception there, and then re-raise a more common exception (StorageFull) to glance. Do you think this makes sense? If so, I will close the review in glance, and create a new review in glance_store.

Regards
Liang
From: Dinescu, Stefan [mailto:Stefan.Dinescu at windriver.com]
Sent: Monday, November 12, 2018 10:13 PM
To: Fang, Liang A <liang.a.fang at intel.com<mailto:liang.a.fang at intel.com>>; Badea, Daniel <Daniel.Badea at windriver.com<mailto:Daniel.Badea at windriver.com>>
Cc: starlingx-discuss at lists.starlingx.io<mailto:starlingx-discuss at lists.starlingx.io>
Subject: RE: [Starlingx-discuss] [glance] Analysis of patch c73ab48 (Enhanced-error-messages) for StartlingX upstreaming

Hello,

I havent' tested that part of the code in some time now, but from what I remember it was part of a larger set of changes to prevent ceph from hanging if not enough space was available.

How it should work (but as I said, I haven't tested this in some time):
-  decrease the glance image pool quota (this is just so it's easier to test)
-  upload images to glance until space if almost full
- the image that would have filled the pool should fail to upload entirely.

It seems you are doing the correct procedure from what you described, so this probably requires a more in-depth investigation to see why it fails.
________________________________
From: Fang, Liang A [liang.a.fang at intel.com]
Sent: Saturday, November 10, 2018 11:23 AM
To: Badea, Daniel; Dinescu, Stefan
Cc: starlingx-discuss at lists.starlingx.io<mailto:starlingx-discuss at lists.starlingx.io>
Subject: [Starlingx-discuss] [glance] Analysis of patch c73ab48 (Enhanced-error-messages) for StartlingX upstreaming
Hi Daniel and Stefan

Please refer to https://review.openstack.org/#/c/616880/
This patch is to provide detailed exception info when no space in rbd. But in my environment, if rbd has no space, the function call to rbd will not return, there's no exception raised.

Details:
In /usr/local/lib/python2.7/dist-packages/glance_store/_drivers/rbd.py

-          Function call will hang at line 484 (_create_image) if no space to create image at the beginning
[cid:image001.png at 01D47C0A.94F63E70]

-          Function call will hang at line 506 (image.resize) if no space at the middle of uploading.

[cid:image002.png at 01D47C0A.94F63E70]

Could you please give some clue to reproduce this patch (raise rbd.NoSpace)?
Couple of other thought:

1.       This patch is to catch exception at v1 api, but v1 is about to be removed by community.

2.       rbd is just one type of the multiple types of backend, so I think catch specific backend's exception at a common code is not so good. glance_store/rbd.py may be the better place to catch this exception. In fact rbd.py already did similar thing, see:

[cid:image003.png at 01D47C0A.94F63E70]

Thanks
Liang
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.starlingx.io/pipermail/starlingx-discuss/attachments/20181114/ecc54539/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.png
Type: image/png
Size: 2947 bytes
Desc: image001.png
URL: <http://lists.starlingx.io/pipermail/starlingx-discuss/attachments/20181114/ecc54539/attachment-0003.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image002.png
Type: image/png
Size: 1524 bytes
Desc: image002.png
URL: <http://lists.starlingx.io/pipermail/starlingx-discuss/attachments/20181114/ecc54539/attachment-0004.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image003.png
Type: image/png
Size: 8559 bytes
Desc: image003.png
URL: <http://lists.starlingx.io/pipermail/starlingx-discuss/attachments/20181114/ecc54539/attachment-0005.png>