Need for new x86 Jenkin's Bot

classic Classic list List threaded Threaded
5 messages Options
slacka slacka
Reply | Threaded
Open this post in threaded view
|

Need for new x86 Jenkin's Bot

Christian,


The decision was to demote release builds, but NOT to remove x86 compatibility. If you take this as an excuse to also end x86 CI testing bots, you are effectively killing x86 through bit rot. This is not a hypothetical. As you know, @87 just recently caught a regression introduced by an easyHack to convert use of sal_uLong to better integer types. Without CI testing, these bugs will pile up effectively removing x86 compatibility.


In a past job, I have seen how building code against multiple architectures uncovers latent bugs and code portability. This philosophy made ARM and RISC ports much easier along with higher quality software.


If we chose to go down the path of killing x86, we should discuss it. While I agree that overhead of x86 release build may not be worth it, the advantages of maintaining a working x86 build are clear.  My suggestion would be to upgrade @87 to CentOS 7 or Debian 9.0 before bit rot makes this a difficult task. Currently master has no problem building on 32-bit Fedora 30.


-Luke


From: Christian Lohmaier <lohmaier+[hidden email]>
Sent: Tuesday, June 4, 2019 6:17 AM
To: Luke Benes
Subject: Re: @87 reboot?
 
Hi Luke,

On Sun, Apr 14, 2019 at 6:40 PM Luke Benes <[hidden email]> wrote:
>
> Christian,
> No rush, just a polite ping. Looks like @87 is stuck.

It's not really stuck per se, it is just that the old baseline is no
longer suitable for master/libreoffice-6-3 (glib2 requirement not
fullfilled) - with 6.2 we demoted 32bit linux builds provided by TDF
(they for example already don't include all vclplugins, most notably
gtk3) and stated that 6.2 will be last line with 32bit builds provided
by TDF (of course distros continue to provide buidls, and 32bit
support in sourcecode is not removed).

So unless there's an outcry of users there won't be master buidls by
that box anymore - (well even if there is there won't be, but rather
32bit version then would be cross-compiled from a 64bit host)

ciao
Christian

_______________________________________________
LibreOffice mailing list
[hidden email]
https://lists.freedesktop.org/mailman/listinfo/libreoffice
Jan-Marek Glogowski Jan-Marek Glogowski
Reply | Threaded
Open this post in threaded view
|

Re: Need for new x86 Jenkin's Bot

Hi everyone,

Am 04.06.19 um 17:01 schrieb Luke Benes:
> The decision was to demote release builds, but NOT to remove x86
> compatibility. If you take this as an excuse to also end x86 CI testing bots,
> you are effectively killing x86 through bit rot. This is not a hypothetical.
> As you know, @87 just recently caught a regression introduced by an easyHack
> to convert use of sal_uLong to better integer types. Without CI testing,
> these bugs will pile up effectively removing x86 compatibility.

IMHO we should "just" switch one of the Jenkins Linux builds to 32bit, not have
an other one. Preferable the gcc release one, as that has smaller object size,
if that is a concern for 32bit (linking etc). Should be fine just to cross-build
on 64bit in a 32bit chroot.

I remember there were some problems because of a missing 32bit Red Hat Developer
Toolset needed for the build, which was AFAIK the real reason to get rid of the
32bit builds - aehmm. Not sure that is now available.

And I / Rene was bitten a few times by broken 32bit builds, last was the
QVariant constructor ambiguity, so generally I'm all for it, if possible.

Jan-Marek
_______________________________________________
LibreOffice mailing list
[hidden email]
https://lists.freedesktop.org/mailman/listinfo/libreoffice
Michael Stahl-3 Michael Stahl-3
Reply | Threaded
Open this post in threaded view
|

Re: Need for new x86 Jenkin's Bot

On 04.06.19 19:23, Jan-Marek Glogowski wrote:

> Hi everyone,
>
> Am 04.06.19 um 17:01 schrieb Luke Benes:
>> The decision was to demote release builds, but NOT to remove x86
>> compatibility. If you take this as an excuse to also end x86 CI testing bots,
>> you are effectively killing x86 through bit rot. This is not a hypothetical.
>> As you know, @87 just recently caught a regression introduced by an easyHack
>> to convert use of sal_uLong to better integer types. Without CI testing,
>> these bugs will pile up effectively removing x86 compatibility.
>
> IMHO we should "just" switch one of the Jenkins Linux builds to 32bit, not have
> an other one. Preferable the gcc release one, as that has smaller object size,
> if that is a concern for 32bit (linking etc). Should be fine just to cross-build
> on 64bit in a 32bit chroot.
>
> I remember there were some problems because of a missing 32bit Red Hat Developer
> Toolset needed for the build, which was AFAIK the real reason to get rid of the
> 32bit builds - aehmm. Not sure that is now available.

just to reiterate again, we have a docker container setup that builds on
the same CentOS 7 + DevToolset 7 baseline that the 64-bit builds use;
the toolchain is 64-bit, which helps with the OOM issues that used to
plague the 32-bit build.

https://fosdem.org/2019/schedule/event/lightning_talks/attachments/slides/3334/export/events/attachments/lightning_talks/slides/3334/7_Thorsten.odp
_______________________________________________
LibreOffice mailing list
[hidden email]
https://lists.freedesktop.org/mailman/listinfo/libreoffice
Christian Lohmaier-2 Christian Lohmaier-2
Reply | Threaded
Open this post in threaded view
|

Re: Need for new x86 Jenkin's Bot

In reply to this post by slacka
On Tue, Jun 4, 2019 at 5:01 PM Luke Benes <[hidden email]> wrote:
>
> The decision was to demote release builds, but NOT to remove x86 compatibility. If you take this as an excuse to also end x86 CI testing bots, you are effectively killing x86 through bit rot.

That is completely missing the point. The tinderbox *IS* using
baseline. And since the baseline is not recent enough to build 32bit
anymore, there is *no way* to have it still produce builds. Otherwise
32bit binaries could still be provided.

> This is not a hypothetical. As you know, @87 just recently caught a regression […]

Anyone is free to setup a tinderbox and send results - no registration
or similar needed for that,
https://wiki.documentfoundation.org/Development/Tinderbox/Setup

> If we chose to go down the path of killing x86, we should discuss it. While I agree that overhead of x86 release build may not be worth it, the advantages of maintaining a working x86 build are clear.  My suggestion would be to upgrade @87 to CentOS 7 or Debian 9.0 before bit rot makes this a difficult task. Currently master has no problem building on 32-bit Fedora 30.

My personal interest is having baseline builds verified. So there is
no point if the build succeeds in a current distro, but cannot run on
the end-users systems because of glibc or other requirements.

But again: anyone can setup a tinderbox...

ciao
Christian
_______________________________________________
LibreOffice mailing list
[hidden email]
https://lists.freedesktop.org/mailman/listinfo/libreoffice
slacka slacka
Reply | Threaded
Open this post in threaded view
|

Re: Need for new x86 Jenkin's Bot

Christian,

Our concern is that by scuttling @87, we will lose important diversity in TDF's testing infrastructure. While upgrading @87 does address this, it is not what we are asking for.

Why have you chosen to not upgrade @87 like you have the others? Master will build on x86 Fedora 30, Cent OS 7, and should on Cent OS 6 with the devtoolset-7 found here: https://copr-be.cloud.fedoraproject.org/results/mlampe/devtoolset-7/epel-6-i386/


I can help you with this or setting up a new Jenkins bot. However, I cannot provide the hardware nor network resources for this purpose. Providing resources to keep our codebase healthy is at the heart of TDF’s mission.  If @87 can not longer be used for baseline binaries, it should be used for this purpose.

-Luke


_______________________________________________
LibreOffice mailing list
[hidden email]
https://lists.freedesktop.org/mailman/listinfo/libreoffice