Long running builds can monopolise the build farm

My immediate reaction to this is that auto builds have the same right to the PPA build resources as anything else does.

However, I'd like to invite comments from others before I go any further with this.

I'm not saying that autobuilds should be restricted from using PPA build resources, or even that they should be forced to build at a lower priority. I'm just saying that they should be given the option to do so if they so desire. If I were running an autobuild, I'd want to be able to run it at a lower priority for the same reason that I often nice CPU-intensive background tasks.

The daily builds that exist now (thinking of the Mozilla and Chromium ones, in particular) take forever to build. They also build across multiple series, and three architectures. That's a lot of buildd time. With the current scheduling, that can for several hours monopolize the reduced buildd farm that we often see. While it is a useful use of buildd time, it makes the PPA service less attractive for others when there are waiting times of many hours.

But, of course, this couldn't be an easy problem to solve. How should the builds be scheduled to avoid this monopoly?

Most PPA builds take in the order of a few minutes, and code already exists to estimate the build duration. Maybe the builddmaster could refuse to fill up the buildds with builds expected to take more than, say, 10 minutes. By always leaving some fraction of buildds performing short tasks, the effect of the daily DoS could be decreased substantially.

I think this situation can be ameliorated using the fix from bug 382804. It's not automatic, but we could ask the owners of the PPAs who do daily builds if they would be happy to see their scores adjusted.

Let me know what you think.

That wouldn't help too much on its own -- as soon as the queue is empty (apart from the problematic builds) for just a moment, the queued long-running low-priority builds will start, inhibiting the start of any subsequent builds.

On Tuesday 30 June 2009 10:59:12 William Grant wrote:
> That wouldn't help too much on its own -- as soon as the queue is empty
> (apart from the problematic builds) for just a moment, the queued long-
> running low-priority builds will start, inhibiting the start of any
> subsequent builds.

Well, they need to build *sometime*, is there a better time to start them than
when the queue is empty?

There is no good time to let long-running builds start on all of the buildds.

What about limiting that a PPA can only use one buildd per arch at a time? That way an upload of a long-running build for several series won't occupy all available buildds for an architecture at the same time but build them one after another.
IMHO it's fair to expect that someone (or a team) uploading a huge amount of package in short time has to wait a little bit to get them all build. If the packages are build fast that shouldn't introduce a huge extra delay (depending on the queue of course) and long-builds won't take all available buildds.
Of course a small number of teams doing daily-builds will still be able clog the buildds but they all have the same rights to use them as any other PPA user.

An other option would be that a certain amount of buildds is reserved for the "general" use and not used by the daily upload builds. That way "normal" PPA user have a chance to get their package build in a timely manner.

Which option is more preferable depends also on the usage rate of the buildds. Are there any statistics about it (what part of a day are they building or waiting for jobs)?

I'm going to triage this, but will also mention that:
 * I /think/ you only noticed this when there were some builders temporarily taken out of the rotation which made the queue get bigger
 * We're working on making the existing builders work for both PPA and Ubuntu builds
 * We're working on making the builders build for any architecture

The last two points there will help quite a lot, as will adding more builders as the PPA service increases in popularity.

I think Michael's suggestion of "limiting that a PPA can only use one buildd per arch at a time" is sounding quite good now.

Just some initial planning/investigation: currently the BuilddManager.scan() iterates each builder, and effectively for each idle builder it calls builder.findBuildCandidate() - and if it finds *one* - dispatches it.

So, afaics, the main issue for implementing the above is that we have all our builders picking single builds off the one primary build queue, where as it would be handy to have instead individual queues for each builder and dispatch a set of builds from the primary queue to the individual builder queue.

Eventually I think it would be great to utilize a messenging system (RabbitMQ - where are you?) to do the above, but in the mean time, another possible solution might be to have a new build-state in between NEEDSBUILD and BUILDING - perhaps BUILDER_ASSIGNED. During builder.findBuildCandidate() - in addition to returning the next build - we could set the status of all the other builds of same priority for the same SPR.creator (? hrm... uploader?), assign the builder and set the BuildStatus.BUILDER_ASSIGNED. Subsequent calls to builder.findBuildCandidate() would always find and dispatch the assigned builds first before checking the general queue again?

Actually, as a slight improvement, the builder would first check if there are any *higher* priority builds in the build queue (and there are a number of options for what it could do if one/some are found, fairest probably being to reset existing assigned builds to NEEDSBUILD).

This behaviour in IBuilder.findBuildCandidate() would have to be optional, as I assume we would *not* want this behaviour for urgent builds (eg. security). There are a number of ways we could make it optional - ppa setting, or a priority threshold etc.

Thoughts?

Sorry - I had my head fixed on ensuring that all the builds are assigned/built by the same builder - but that's not necessary at all - so as suggested by Julian, we can simply ensure during findBuildCandidate() that we skip potential builds iff there is already a build for the same architecture+archive currently building on any builder.

The build start time estimation code will have to take the new assignment algorithm into account as well.

Very good point! But that should be a separate bug as it's more important to get this one fixed in its own right first.

Fixed in devel r9685 <http://bazaar.launchpad.net/~launchpad-pqm/launchpad/devel/revision/9685>

I've outlined a few ways this initial implementation could be improved/tweaked in bug 450124.

Is it possible to restrict which builders the excessive daily builds can use? I believe Chromium takes about 3 hours to build each of the 4 i386 builds every single day which contributes to a ridiculous backlog for everyone else.

I don't think restricting by builder will help. Let's see how this fix works out and then we can decide if further action is needed.