I see frequent timeouts on this bot -- can the timeout be increased?
Thank you for the service offered!
Hi Davide,
Increasing the timeout is not ideal. We need to fix the problem in a
different way.
Our other ARM bots are using CMake 3.2 + Ninja 1.5, which gets rid of
one source of time outs (slow testing), so that may be one way out.
For now, please ignore any timeout on bots, as they happen not just
with this one. 
cheers,
--renato
What's the problem with increasing the timeout? Asking people to ignore
buildbot mails does not seem right. If the buildbot is flaky I believe
the buildbot owner should ensure it shuts up until the problems have
been resolved and the buildbot has a low false positive rate again.
Best,
Tobias
I see frequent timeouts on this bot -- can the timeout be increased?
Hi Davide,
Increasing the timeout is not ideal. We need to fix the problem in a
different way.Our other ARM bots are using CMake 3.2 + Ninja 1.5, which gets rid of
one source of time outs (slow testing), so that may be one way out.For now, please ignore any timeout on bots, as they happen not just
with this one.What's the problem with increasing the timeout? Asking people to ignore
buildbot mails does not seem right. If the buildbot is flaky I believe
the buildbot owner should ensure it shuts up until the problems have
been resolved and the buildbot has a low false positive rate again.
Yes, please. I would go one step further and ask that authors of flaky
tests with non-deterministic failures should revert those tests until
the problems have been resolved as well. The number of false positives
currently makes it very hard to know what the state of the product is.
~Aaron
That's the point I make about solving the real issue, not increase the timeout.
CMake + Ninja has fixed virtually all our flakiness on all other ARM
bots, so I think we should give it a try first.
--renato
What time-line do you have in mind for this fix? If you are in charge
and can make this happen within a day, giving cmake + ninja a chance seems
OK.
However, if the owner of the buildbot is not known or the fix can not come
soon, I am in favor of disabling the noise and (re)enabling it when someone
found time to address the problem and verify the solution. The cost of
buildbot noise is very high, both in terms of developer time spent, but
more importantly due to people starting to ignore them when monitoring them
becomes costly.
Best,
Tobias
What time-line do you have in mind for this fix? If you are in charge
and can make this happen within a day, giving cmake + ninja a chance seems
OK.
It's not my bot. All my bots are CMake+Ninja based and are stable enough.
However, if the owner of the buildbot is not known or the fix can not come
soon, I am in favor of disabling the noise and (re)enabling it when someone
found time to address the problem and verify the solution.
That's up to Galina. We haven't had any action against unstable bots
so far, and this is not the only one. There are lots of Windows and
sanitizer bots that break randomly and provide little information, are
we going to disable them all? How about the perf bots that still fail
occasionally and we haven't managed to fix the root cause, are we
going to disable then, too?
You're asking to reduce considerably the quality of testing on some
areas so that you can reduce the time spent looking at spurious
failures. I don't agree with that in principle. There were other
threads focusing on how to make them less spurious, more stable, less
noisy, and some work is being done on the GreenDragon bot structure.
But killing everything that looks suspicious now will reduce our
ability to validate LLVM on the range of configurations that we do
today, and that, for me, is a lot worse than a few minutes' worth of
some engineers.
The cost of
buildbot noise is very high, both in terms of developer time spent, but
more importantly due to people starting to ignore them when monitoring them
becomes costly.
I think you're overestimating the cost.
When I get bot emails, I click on the link and if it was timeout, I
always ignore it. If I can't make heads or tails (like the sanitizer
ones), I ignore it temporarily, then look again next day.
My assumption is that the bot owner will make me aware if the reason
is not obvious, as I do with my bots. I always wait for people to
realise, and fix. But if they can't, either because the bot was
already broken, or because the breakage isn't clear, I let people know
where to search for the information in the bot itself. This is my
responsibility as a bot owner.
I appreciate the benefit of having green / red bots, but you also have
to appreciate that hardware is not perfect, and they will invariably
fail once in a while. I had some Polly bots failing randomly and it
took me only a couple of seconds to infer so. I'm not asking to remove
them, even those that fail more than pass throughout the year. I
assume that, if they're still there, it provides *some* value to
someone.
cheers,
--renato
That’s not how I understand his point. In my opinion, he is asking to increase the quality of testing. You just happen to disagree on his solution
The situation does not seem that black and white to me here. In the end, it seems to me that is is about a threshold: if a bot is crashing 90% of the time, does it really contributes to increase the quality of testing or on the opposite it is just adding noise? Same question with 20%, 40%, 60%, … We may all have a different answer, but I’m pretty sure we could reach an agreement on what seems appropriate
Another way of considering in general the impact of a bot on the quality is: “how many legit failures were found by this bot in the last x years that weren’t covered by another bot”.
Because sometimes you may just having a HW lab stress rack, without providing any increased coverage for the software.
Cheers,
>
>> What time-line do you have in mind for this fix? If you are in charge
>> and can make this happen within a day, giving cmake + ninja a chance
seems
>> OK.
>
> It's not my bot. All my bots are CMake+Ninja based and are stable enough.
>
>
>> However, if the owner of the buildbot is not known or the fix can not
come
>> soon, I am in favor of disabling the noise and (re)enabling it when
someone
>> found time to address the problem and verify the solution.
>
> That's up to Galina. We haven't had any action against unstable bots
> so far, and this is not the only one. There are lots of Windows and
> sanitizer bots that break randomly and provide little information, are
> we going to disable them all? How about the perf bots that still fail
> occasionally and we haven't managed to fix the root cause, are we
> going to disable then, too?
>
> You're asking to reduce considerably the quality of testing on some
> areas so that you can reduce the time spent looking at spurious
> failures. I don't agree with that in principle.That’s not how I understand his point. In my opinion, he is asking to
increase the quality of testing. You just happen to disagree on his
solutionThe situation does not seem that black and white to me here. In the end,
it seems to me that is is about a threshold: if a bot is crashing 90% of
the time, does it really contributes to increase the quality of testing or
on the opposite it is just adding noise? Same question with 20%, 40%, 60%,
… We may all have a different answer, but I’m pretty sure we could reach
an agreement on what seems appropriateAnother way of considering in general the impact of a bot on the quality
is: “how many legit failures were found by this bot in the last x years
that weren’t covered by another bot”.
Even that doesn't really capture it - if the bot has enough false
positives, or spends long periods being red, even those legit failures will
be lost in the noise & the cost to the whole project (not only in ignoring
that bot, but in reducing confidence in the bots in general (which is
pretty low generally because of this kind of situation)) may outweigh the
value of those bugs being found.
If a bot is of low enough quality that most engineers ignore it due to
false positives, long periods of broken-ness, then it makes sense to me to
remove it from the main buildbot view and from sending email. The owner can
monitor the bot and, once they triage a failure, manually reach out to
those who might be to blame.
(oh, and add long cycle times to the list of issues - people do have a
tendency to ignore bots that come back with giant blame lists & no obvious
determination as to who's patch caused the problem, if any)
- David
The situation does not seem that black and white to me here. In the end, it seems to me that is is about a threshold: if a bot is crashing 90% of the time, does it really contributes to increase the quality of testing or on the opposite it is just adding noise?
That question is not alone, and per se, it's meaningless. Your next
question, however, is the key.
Another way of considering in general the impact of a bot on the quality is: “how many legit failures were found by this bot in the last x years that weren’t covered by another bot”.
In this criteria, which was my point, those bots haven't added much
after I added some faster ones. So, if we want to shut them down,
let's do so because they don't add value, not because they are
unstable.
However, that is the *only* bot running on an A9. As an example, this
year, I spent 2 whole weeks during the release to bisect and fix an
issue because I disabled, for two months, one bot that I thought was
already covered by another.
That headache was real. I've wasted two whole weeks, maybe more or my
time. I've wasted time from other people waiting to do the release
validation. I have delayed the release and all that it entangles. All
because I thought that bot was noisy. It's ratio was about 20 passes
to 1 failure.
The A9 bots are more than that, so on my monitor[1], I currently
ignore their results. I still keep them there to see what's going on,
and when my bots fail, I look at that, too, to see if the problem is
the same. Sometimes, they do provide useful insight on the other bot's
breakages.
So, for me, disabling the A9 bots would be a loss. But as I said
before, that's up to Galina, as she's the bot owner. If she's ok with
finally putting them to rest, I'll respect the community's decision
and remove it from my monitor. But we can't transform this in to a
which hunt. It's not about thresholds, it's about cost and value,
which may be different for you than it is for me. We have to consider
the whole community, not just our opinions.
For every broken bot that someone want to get rid of, I propose
consult the bot owner first, and then a vote on llvm-dev@ / cfe-dev@
if the owner is not in agreement. After all, you can always have an
internal buildmaster (like I have) for unstable bots.
cheers,
--renato
[1] http://llvm.tcwglab.linaro.org/monitor/
@Galina: It seems this bot is now almost permanently running into a compile-time
timeout. Maybe you can fix this by either increasing the timeout or by
switching to a cmake/ninja based build as suggested by Renato.
What time-line do you have in mind for this fix? If you are in charge
and can make this happen within a day, giving cmake + ninja a chance seems
OK.It's not my bot. All my bots are CMake+Ninja based and are stable enough.
I should have looked it up myself. I did not want to finger-point, but
ensure we understand who will address this issue. I just looked up who
owns these builders and to my understanding it is Galina herself.
I CC her such that she can take action.
I also have the feeling I was generally to harsh in my mail, as I seem
to have triggered a rather defensive reply. Sorry for this.
Regarding the discussion about disabling/enabling buildbots. I agree with
Mehdi there is no black and white. For this bot, it seems important to
address this issue as it seems to start failing very regularly now.
Regarding my own bots: In case you see flaky polly buildbots or any other
of my bots sending emails without reason, please send me a short ping
such that I can fix the issue. None of my LNT bots send emails as they
run too long before starting to report.
Best,
Tobias
Yes, but remember, not all hardware is as fast as a multi-core Xeon
server. Build times can't always be controlled.
But I agree with you on all accounts. The bot owner should bear the
responsibility of his/her own unstable bots. If it brings less value
than it adds cost to the community, it should be moved to a separate
buildmaster that doesn't email people around, but can be accessed, so
the owner can point breakages to devs.
cheers,
--renato
What time-line do you have in mind for this fix? If you are in charge
and can make this happen within a day, giving cmake + ninja a chance seems
OK.It's not my bot. All my bots are CMake+Ninja based and are stable enough.
However, if the owner of the buildbot is not known or the fix can not come
soon, I am in favor of disabling the noise and (re)enabling it when someone
found time to address the problem and verify the solution.That's up to Galina. We haven't had any action against unstable bots
so far, and this is not the only one. There are lots of Windows and
sanitizer bots that break randomly and provide little information, are
we going to disable them all? How about the perf bots that still fail
occasionally and we haven't managed to fix the root cause, are we
going to disable then, too?
If the bot fails regularly (say false positive rate 1 in 10 runs), then yes, it should be disabled until the owner fixes it. It's perfectly okay for it to be put into a "known unstable" list and for the bot owner to report failures after they've been confirmed.
To say this differently, we will revert a *change* which is problematic. Why shouldn't we "revert" a bot?
You're asking to reduce considerably the quality of testing on some
areas so that you can reduce the time spent looking at spurious
failures. I don't agree with that in principle. There were other
threads focusing on how to make them less spurious, more stable, less
noisy, and some work is being done on the GreenDragon bot structure.
But killing everything that looks suspicious now will reduce our
ability to validate LLVM on the range of configurations that we do
today, and that, for me, is a lot worse than a few minutes' worth of
some engineers.The cost of
buildbot noise is very high, both in terms of developer time spent, but
more importantly due to people starting to ignore them when monitoring them
becomes costly.I think you're overestimating the cost.
When I get bot emails, I click on the link and if it was timeout, I
always ignore it. If I can't make heads or tails (like the sanitizer
ones), I ignore it temporarily, then look again next day.
I disagree strongly here. The cost of having flaky bots is quite high. When I make a commit, I'm committing to be responsive to problems it introduces over the next few hours. Every one of those false positives is a 5-10 minute high priority interruption to what I'm actually working on. In practice, that greatly diminishes my effectiveness.
As an illustrative example, I submitted some documentation changes earlier this week and got 5 unique build failure notices. In this case, I ignored them, but if that had been a small code change, that would have cost me at least an hour of productivity.
My assumption is that the bot owner will make me aware if the reason
is not obvious, as I do with my bots. I always wait for people to
realise, and fix. But if they can't, either because the bot was
already broken, or because the breakage isn't clear, I let people know
where to search for the information in the bot itself. This is my
responsibility as a bot owner.
First, thanks for being a responsible bot owner.
If all bot owners were doing this, having a unstable list which doesn't actively notify would be completely workable. If not all bot owners are doing this, I can't say I really care about the status of those bots.
> (oh, and add long cycle times to the list of issues - people do have a
> tendency to ignore bots that come back with giant blame lists & no
obvious
> determination as to who's patch caused the problem, if any)Yes, but remember, not all hardware is as fast as a multi-core Xeon
server. Build times can't always be controlled.
Small blame lists can still be acquired by having more hardware. Certainly
not always possible/in the budget for those who want to verify these things.
But I agree with you on all accounts. The bot owner should bear the
responsibility of his/her own unstable bots. If it brings less value
than it adds cost to the community, it should be moved to a separate
buildmaster that doesn't email people around, but can be accessed, so
the owner can point breakages to devs.
More than just not emailing, it'd be great to have a generally-green
dashboard to look at. For now it's hard to get a sense of what's 'really'
broken'.
But yeah, all of that stuff.
How I fixed my bots:
1. Remove cmake and ninja from your system. They are too old.
2. Download cmake stable sources (3.2+), untar, bootstrap, make, make install
3. Checkout ninja from github, bootstrap, copy "ninja" to /usr/local/bin
4. Install ccache from packages, add ccache to path
5. Change the builder to ClangCMakeBuilder like all ARM and AArch64 bots now in.
6. Restart.
The Ninja+CMake combo has a feature that makes sure you print
everything without buffering, so the time out works exactly as
intended: if any single process takes more than that time, it's a bug.
I'm copying Gabor, as AFAIK, his bot is not based on the new
CMake+Ninja fix, but on an old polling script we had, which makes
timeouts useless.
Let's try this one first, and only consider any more drastic solution after.
cheers,
--renato
To say this differently, we will revert a *change* which is problematic.
Why shouldn't we "revert" a bot?
I don't disagree, just don't want to do that lightly. Most certainly
not before we have comments from the bot owner.
As an illustrative example, I submitted some documentation changes earlier
this week and got 5 unique build failure notices. In this case, I ignored
them, but if that had been a small code change, that would have cost me at
least an hour of productivity.
I have to say, I never spent more than a few minutes looking up
failing bots. If there's nothing that I can find in 30 seconds of
looking at the bot screen, I rely on the bot owners to ping me, revert
my patches, let me know what's wrong.
I'll make your words, mine:
If all bot owners were doing this, having a unstable list which doesn't
actively notify would be completely workable. If not all bot owners are
doing this, I can't say I really care about the status of those bots.
cheers,
--renato
(oh, and add long cycle times to the list of issues - people do have a
tendency to ignore bots that come back with giant blame lists & no obvious
determination as to who's patch caused the problem, if any)Yes, but remember, not all hardware is as fast as a multi-core Xeon
server. Build times can't always be controlled.
No, but there are many known partial solutions. We can explore some of them if build time is the only issue.
Some options (I don't know which if any have already been explored)
- Kick off builds on slow bots at revisions which have passed a fast bot
- Overlap builds on different pieces of hardware so that revision ranges are smaller (i.e. "that change was in the previous build and didn't fail, so it can't be that")
- Build incrementally with occasional clean builds at already known good revisions (or for failure recovery)
- Separate build and test. Cross build on a faster machine then transfer the binaries to a slower test machine. Run only a build on the slow machine. (i.e. decrease latency of build+test to latency of build OR test)
- Consider using an emulator on a faster machine to get initial results. Only kick off builds on hardware for changes that pass the emulator. (This is a particular variant of (1) above.)
More unstable hardware is more unstable.
cheers,
--renato
> @Galina: It seems this bot is now almost permanently running into a
> compile-time
> timeout. Maybe you can fix this by either increasing the timeout or by
> switching to a cmake/ninja based build as suggested by Renato.How I fixed my bots:
1. Remove cmake and ninja from your system. They are too old.
2. Download cmake stable sources (3.2+), untar, bootstrap, make, make
install
3. Checkout ninja from github, bootstrap, copy "ninja" to /usr/local/bin
4. Install ccache from packages, add ccache to path
5. Change the builder to ClangCMakeBuilder like all ARM and AArch64 bots
now in.
6. Restart.The Ninja+CMake combo has a feature that makes sure you print
everything without buffering, so the time out works exactly as
intended: if any single process takes more than that time, it's a bug.I'm copying Gabor, as AFAIK, his bot is not based on the new
CMake+Ninja fix, but on an old polling script we had, which makes
timeouts useless.Let's try this one first, and only consider any more drastic solution
after.
*shrug* I haven't looked at whatever specific bots are under discussion,
but I really wouldn't mind/would like if the bots had a more "revert to
green" feel to them just like we have for commits: take a bot offline,
fix/iterate/improve it, see if it comes good, then bring it back to the
mainline.
> Small blame lists can still be acquired by having more hardware.
Certainly
> not always possible/in the budget for those who want to verify these
things.More unstable hardware is more unstable.
I was referring specifically to the issue of long cycle times producing
long blame lists. That can be reduced by having more bots so that blame
lists are smaller.
Even if the hardware is just as unstable this is actually better, it's not
more unstable as such. It means that when it does flake out, fewer people
are distracted/informed of this. That's an improvement in a small degree.