[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH 0/3] fix build failures from incorrectly skipped container bu
Re: [PATCH 0/3] fix build failures from incorrectly skipped container build jobs
Tue, 9 Feb 2021 07:55:11 +0100
Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101 Thunderbird/78.7.0
On 09/02/2021 07.01, Stefan Weil wrote:
Am 08.02.21 um 19:12 schrieb Daniel P. Berrangé:
On Mon, Feb 08, 2021 at 07:08:39PM +0100, Philippe Mathieu-Daudé wrote:
On 2/8/21 6:22 PM, Daniel P. Berrangé wrote:
On Mon, Feb 08, 2021 at 04:33:36PM +0000, Daniel P. Berrangé wrote:
This series fixes a problem with our gitlab CI rules that cause
container builds to be skipped. See the commit description in the
first patch for the details on this problem.
The overall result of this series though is a small increase in overall
- When container jobs are skipped: approx 1hr 5 mins
- When container jobs are run, cached by docker: approx 1hr 15 mins
- When container jobs are run, not cached by docker: approx 1hr 30 mins
With this series applied the first scenario no longer exists, so
all piplines are either 1hr 15 or 1hr 30 depending on whether the
container phase is skipped.
I mean to say the biggest problem I see is the cross-win64-system
job. This consumes 1 hour 5 minutes all on its own. It is at least
15 minutes longer that every other job AFAICT. So no matter how
well we parallelize stuff, 1 hr 5 is a hard lower limit on pipeline
duration right now.
We might want to consider how to split the win64 job or cut down
what it does in some way ?
When the win64 build was added (on Debian), it was to mostly to cover
the WHPX. Later we moved mingw jobs to Fedora. I just checked and
WHPX is no more built, and nobody complained, so it is not relevant
I don't mind much what you do with the Gitlab win64 job, as this config
is better covered on Cirrus. I'd like to keep the win32 job because it
has been useful to catch 32-bit issues.
I'm not suggesting we remove it. Most developers won't setup Cirrus CI,
so will only run GitLab CI jobs. IME it is good to have both win32
and win64 coverage because things do break differently on them - especially
if you use bad printf format strings that are not independant of host
I would not say that something is not relevant just because nobody
complains. Nobody would miss any CI if everything were always fine. So
people would miss WHPX CI as soon as there are changes (which happen
infrequently) and something breaks. WHPX should be covered by the w64 build,
and as many as possible other features where I currently see a "NO" in the
configure output as well.
Yes, I agree, we should add WHPX there again. Could somebody please check
whether the headers are already available in the latest Fedora? Then we
might simply switch the container to use the latest version of Fedora instead.
Otherwise we should install the headers manually there, like we did in
commit d3dd34a1e5e134 for the now-removed Debian container.
Nevertheless I don't think that each CI job must run frequently. Each run
not only costs time, but also energy, and contributes to our climate change.
I think that for the win32 and win64 jobs it would be sufficient to run them
weekly, maybe even alternating if that is possible.
Maybe it would be sufficient to run those jobs only on tags (so that they
get checked for pull requests) and on the master and staging branch?
[PATCH 2/3] gitlab: add fine grained job deps for all build jobs, Daniel P . Berrangé, 2021/02/08
[PATCH 3/3] gitlab: fix inconsistent indentation, Daniel P . Berrangé, 2021/02/08
Re: [PATCH 0/3] fix build failures from incorrectly skipped container build jobs, Daniel P . Berrangé, 2021/02/08
Re: [PATCH 0/3] fix build failures from incorrectly skipped container build jobs, Philippe Mathieu-Daudé, 2021/02/16
- Re: [PATCH 1/3] gitlab: always build container images, (continued)