Development issues part 2: Bug tracking

There comes a time, though, when even the most die-hard free software proponent wishes that things would just work. As our software finds its way into more situations where failures are unwelcome (at best), the level of tolerance for bugs is falling. The desire for fewer flaws, however, runs counter to the desire for increasingly capable (and thus more complex) software. Somehow we have to find ways to simultaneously grow our systems and reduce the total number of bugs. To this end, a few projects have been having some interesting discussions on the tracking and fixing of bugs.

As has been discussed in this companion article, Eric Raymond has been busily stirring up trouble on the Emacs development list. His point, deemed reasonable by your editor, is that Emacs must adopt a number of relatively modern development practices if it is to have any hope of remaining relevant at all. One of his key points is that Emacs needs to have a real bug tracking system. Says Eric:

While some of Eric's suggestions appear to be non-starters - imagine trying to get Richard Stallman to hang out on an IRC channel - the bug tracker suggestion might just go somewhere. Certainly it could only be an improvement for a project of that size to have some sort of idea of what the current list of outstanding bugs looks like. It might even help bring about another Emacs release before the end of the decade.

Bug trackers are not a magical solution to the bug problem, though; in fact, they can create some problems of their own. The Fedora project, which does have a bug tracker, is currently trying to figure out what to do with the contents of that tracker. It seems that said tracker contains over 13,000 bugs, almost 10,000 of which apply to Fedora 7 and later.

A bug database of this size is simply overwhelming to anybody who tries to do something about it. As a result, Fedora users are filing bugs, only to see nothing happen in response. Not even a "thanks for your report" message. This situation is discouraging for everybody involved, causing Fedora users to give up on reporting bugs and developers to fear looking at the tracker.

In the Fedora case, there appears to be a near-consensus that the biggest problem is in triaging bug entries. This is not a job which can be automated; somebody has to go through bug submissions, weed out the duplicates, identify those which are really "features," figure out which developer should be notified, etc. Tying bug entries to those found in upstream trackers would be a highly useful bonus. Without this sort of effort, the bug tracker quickly fills with low-quality entries which help nobody.

For the most part, nobody is doing this job for Fedora now. Red Hat is not paying for a staff member to triage bugs, and the wider community has not filled this gap. In the short term, any sort of solution looks like it will have to come from the community, so the Fedora folks are wondering what can be done to encourage more participation. Simply asking for help is the obvious first step, as is making sure that the process is easy. Then they may consider the tactics adopted by other large projects - Mozilla's policy of expressing its appreciation by sending a T-shirt, for example.

As an aside, one of the more useful bits of information to come from this discussion was the existence of this family of URLs:

    http://bugz.fedoraproject.org/<package-name>

Fill in the name, and the result is an immediate list of open bugs for the given package. Thus, for example, a visit to bugz.fedoraproject.org/gcc yields a list of compiler bugs. This result can be had directly from bugzilla, of course, but this interface is faster and easier.

The Fedora developers have discussed a number of related issues, such as whether the Fedora bug database should be separated from the RHEL system and what can be done to make Red Hat better appreciate the value of doing more of its quality assurance work in the Fedora repository. But the core problem is just getting human attention applied to the bug reports. Digging through bug databases is a relatively unglamorous job; it is not an easy path toward rock-star hacker status. But it is an important and relatively easy way to help make free software better.

Just in time to serve as an example of how well bug management can work, the GNOME project has posted its annual bugzilla statistics. It seems that over 110,000 GNOME bugs were filed in 2007, almost 109,000 of them were closed. The top bug-closers for the year were:

14254	Andre Klapper
9800	Tom Parker
7047	Susana Pereira
6882	Bruno Boaventura
6649	Pedro Villavicencio

It is worth pondering for a moment on the amount of energy required to close over 14,000 bugs in a year - that's almost 40 per day, every day, without a break. This kind of energy does exist within our community, and some projects are putting it to very good use.

While it is easy to get a contrary impression, the kernel does, in fact, have a bug tracker; there is also, in the form of Natalie Protasevich, somebody who handles the care and feeding of that tracker. But, as a recent episode shows, that still is not always sufficient to actually get the bugs fixed.

On November 13, 2007, a bug in the SCSI subsystem was reported to the linux-kernel mailing list. It was put into the tracker as bug 9370 on the same day. Some developers looked at it over the next few days, but, even though a specific commit which appeared to cause the bug had been identified, no solution was forthcoming. Discussion eventually died out. At least until January 2, when Ingo Molnar decided to stir the pot by posting a patch to revert the seemingly guilty commit. At that point the discussion picked up and a reliable way of reproducing the bug was found. The commit which was said to have caused the problem was, in fact, not guilty; it had just caused an older bug to come to light. The discussion did not stop there, though.

A number of charges went back and forth which do not require discussion here. But one core point is this: as long as the bug report sat in the tracker, nothing much appeared to be happening with it - though, it seems, the SCSI developers had not forgotten it and were trying to figure out what was really going on. But once the problem came back to the linux-kernel list in the form of a brute-force solution, the root cause was found in short order. The key here was bringing the problem to the attention of a wider group of people; the crucial recipe for reproducing the problem came from a developer who had not been looking at the problem previously.

In the kernel context, at least, giving wide exposure to a bug often helps immensely in getting that bug fixed. That is especially true for the sort of hard-to-reproduce bugs which tend to come up in kernel programming. So, while bug trackers are a useful tool for ensuring that problems do not fall through the cracks, it seems that one of the most potent anti-bug tools we have - discussing the problem via a widely-distributed email list - is the same tool we have been using for decades.

The RH bugzilla is slow, and part of the problem is the Query mask where there is one entry
for each package. This gives a HUGE html file for browsers to parse. This would really make me
want to use the search form less and less (if I actually used FC more, where I would run into
bugs that block my task), and let some developer see if it is a duplicate.

"So, while bug trackers are a useful tool for ensuring that problems do not fall through the
cracks, it seems that one of the most potent anti-bug tools we have - discussing the problem
via a widely-distributed email list - is the same tool we have been using for decades."

The bug tracker of Ruby is a mailing list (ruby-talk).
But when you use a ml like a bugtracker, with more than 100 bugs every day, it's clearly too
big for anyone to keep up the pace. And so is this mailing list. At the end, it was really
used like a bugtracker and discussion about development were on other mailing lists. 
IMHO, the problem is not the mean, bugtracker, mailing list, irc or whatever, it's mainly a
lack of ressources. Everyone in the world can post a bug report. But a small tiny part of
"everyone" is able to read this report and say if it's really a bug or not. And a more smaller
tiny part of them know how to fix it.

Yeah, a bug tracker is no use if the bugs don't get fixed.

I stopped using Mepis and won't go back because nobody cared about a bug that was causing me
grief.

I've submitted a bug or two to Debian and KDE and the lack of any response makes me unlikely
to bother again, except for a real show stopper.

I have an even better example of bugs in the kernel bug database 
languishing. Bug #2645 - msync() does not update the st_mtime and 
st_ctime fields (http://bugzilla.kernel.org/show_bug.cgi?id=2645).

A relatively clear cut bug that is easily reproduced on any machine, and 
can even cause dataloss (because mtime isn't updated, most incremental 
backup systems will ignore the file, as it doesn't seem to have changed). 
The bug was reported in 2004, and only just now something is starting to 
happen in this regard, eventhough several patches for this problem have 
appeared on linux-kernel.

Quite underwhelming.

"All bugs are shallow, given enough eyes."  As you point out with the SCSI bug, it isn't
assigning a bug number, or identifying a problem, but getting more eyes interested in the
problem, that fixes things.  F/OSS relies on people scratching itches, and bug repair relies
on making enough people itch.

It's reasonable to claim that the interdependence of data in the kernel is bigger than in
gnome.
(data = code,knowledge required etc)
You need to distribute to scale, for better scaling when distributing you need to minimize
data interdependence. In gnome there are many seperate projects with little or none dependence
on each other - at least compared to the kernel.

Buzzwords, guessing... not facts.. etc...:)
Also - the kernel has more code?

I think one common problem with bug tracking systems is that they don't provide a good
workflow for reporters from the point of having a problem to the point of getting other people
involved. It is pretty much impossible for a user with a problem to identify whether the
problem has been reported already by using the query capabilities built into most bug
trackers, and, if a user doesn't find their bug already in the system, they have to re-enter
all of the information that they used to search with. Furthermore, the user has to do
something different to find out that something is a common user error. And the user has to do
yet another thing in the case where the problem has been reported by somebody else, but the
report doesn't cover all of the aspects of the new situation. So there's a high chance of user
error in this process, and poor handling of failure cases.

On the other side, one of the most useful things in resolving a bug is a second report which
clarifies or corrects the description of what's actually going on, and this is something that
bug tracking systems and triage actively filter out at pretty much all stages.

Well, another issue is how you deal with things that are fixed. Until everybody is using a
version of the software which has an appropriate error message, which is effectively never,
there will be people who get the old behavior and can't find out from the software that a
newer version is available and fixed their issue.

And, of course, fixing the user documentation doesn't help, because users never read
documentation until something goes wrong, at which point they're faced with the choice of
whether to go to the documentation or to the bug tracker, and they don't necessarily do the
right one. Ideally, there should be some common troubleshooting starting point where the
answer can be: read this section of the manual; or use a version newer than X; or wait for the
next release; or add your account of the issue to this bug; or make a new bug with this
information.

One approach to the bug tracker problem is to make the bug reporter "responsible" for their
bug report - i.e. for looking for other relevant information, including duplicates in the bug
tracker, duplicates in other bug trackers, possible workarounds and fixes and people who might
know how to help with it.  Combined with a policy of automatically closing bugs which are
inactive for too long (with the option of the reporter re-opeining them).  Granted this will
not solve all problems, but it may help a bit.

It can also help to give the people who report bugs (especially the non-coding types) the
feeling that they are making valuable contributions, which in turn can make them more
enthusiastic about the software :)

Not a 'keeping track' issue.

We have this strange belief that if only the bug was written down 
somewhere it has a better chance of being fixed. The only thing writing 
it down does is, um, writing it down.

KDE has a huge bug database filled with a huge number of bugs. Some 
valid, many duplicates, many wishes, a large number too old and possibly 
fixed but who knows.

The best way to ruin anyone's desire to help and contribute to free 
software (imho) is to assign them to bug triage. My joints hurt thinking 
about it. There are rare people with the mix of personality traits that 
love bug triage and are very good at it. A bronze bust must be made of 
each one. No sane person would do it for nothing :) There are probably 
fewer people with these gifts than there are software developers.

So we have a situation where it is easier to harness the manpower to 
write a bug database than it is to maintain the content within.

Someone told me of an executive that would clear his desk on friday 
evening into the garbage can. His theory was that if something was very 
important it would show up again the next week. I think bugs fall into 
that category. And we wouldn't have to maintain the fiction of people 
caring about reported bugs.

Providing a way of communicating with the developer at a data rate that 
they can handle would seem more productive.

Derek