Re: Emacs Bazaar repository

[David Kastrup <dak@gnu.org>]

Nicholas Allen <allen@ableton.com> writes:

> Bazaar has the UI spot on I think - it just needs to perform better. I
> have been using Bazaar since 0.8 and I can tell you the performance
> improvements they have made so far are huge. They are still working on
> performance and I'm sure they will fix this issue in the not too
> distant future as they have a very good track record of doing so....

Reminds me of a job I once did in geodesy.  The previous people working
on the job (partly doctors of informatics) had done several iterations
of optimizations on the code, halving its running time in the process,
then squeezing off some more.  Problem was that the data sets to be
processed were expected to more than triple, and the stuff was already
running a day or so.

I tried working myself into the code (I always have a hard time
understanding code of others, and one of the reasons I was put on the
job was that the code was getting more fragile) and after about two
weeks gave up.  I just don't have the mental capacity to find myself
comfortable around in a mess I did not create myself.  So I just spent
two days on thinking about an algorithm and optimizations and stuff and
implemented it from scratch.  Took probably another two weeks of coding
and testing where it wasn't dumping core or running into loops or
similar.  And then I got stuck debugging it.  The stuff just terminated
after a few minutes of run time, and I was running out of ideas what was
wrong.

After about a day of placing breakpoints and looking at the data
structures I was running out of ideas and turned to looking at the
generated output.

It was correct.  Since that time I have a healthy dose of scepticism
about people (and teams) tinkering with their algorithms and "coming
through".

The real metric is not how much you improve things, but where they
should be in the first place.

I am not implying that this is happening here.  But if we are several
orders of magnitude behind the competition, the proven ability of
squeezing off some runtime at the cost of legibility (and thus also the
viability of further optimizations without destabilization) is not a
useful metric.