[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Bits (Nybbles?) from the Vancouver release team meeting

david@schmitt.edv-bus.at (David Schmitt) writes:

> On Monday 14 March 2005 11:10, Rene Engelhard wrote:
>> pcc is barely at 98%. I don't think that barrier should be that high. We
>> *should* at last release with the tree most important archs: i386, amd64,
>> powerpc.
> Please, 98% is not high. It is just a call to porters to get their act 
> together.

When I was actively running the ia64 and hppa buildd systems, which I am not
doing today, it was my observation that completion percentage was "noisy" and
depended a lot on what was happening in the upload queue on a given day.  So,
while I agree that 98% seems like a reasonable goal, there are some things I
think it is worth keeping in mind if you try to look at the raw numbers.

The number of uploads per day varies a lot.  Some days it's fairly quiet, some
days it's not... and a really productive bug-squash weekend can run the upload
rate *way* up and depress the completion percentages a bit until everyone
catches up.

If there's a buildability problem in an upload that's at the base of a build 
dependency tree, the completion percentage can droop quite a bit until that 
issue gets resolved, then it may bounce back up more or less immediately when 
it gets fixed.  If you happen to look at the numbers in the middle of one of
these events, you'll get unexpected results.

A transient infrastructure problem at a "sensitive" time of day can cause a
droop in completion percentage because of email being backed up, etc.  Easily
enough to cross any "magic threshold" depending on which day you look.

The admins of the various buildds have different patterns of behavior in terms
of how often they review and sign uploads.  If they're on it all the time day
and night, the visible completion percentage is likely to be higher than if
they do this once a day... and the time of day relative to when katie runs
can make a difference, too!

My point is that I would expect any rational application of this sort of 
criteria to look at more than a single instant in time...  I don't think it
is very useful to fret over the "noise" on the signal day to day.  Some
buildd admins treat this as a real game, and work hard to be at the top of 
completion graphs every day... some don't.  

The real question on the day of release is what the build percentage of 
'testing' is for each architecture, and that's a pretty easy place to drive 
the numbers near or to 100% if we think it's important enough!


Reply to: