Error rate miscalculated, excluding machines with no errors today

[Matthew Paul Thomas]

Currently the "Average number of crashes" each day is the total number of errors reported that day, divided by the number of machines that reported at least one error that day.

This is wrong, because it excludes machines that didn't report any errors today, even if they reported errors yesterday and probably still have Ubuntu installed. It means the average could never be lower than 1, no matter how reliable Ubuntu was.

A better approach would be to divide the number of errors each day by the number of machines that reported any errors in the previous 90 days. The result will still be biased up slightly, because it excludes machines that previously reported errors but none in the past 90 days. But it will also be biased down slightly, because it includes machines that were destroyed or had Ubuntu removed from them sometime in the past 90 days. These two biases should roughly cancel each other out.

[Originally pointed out by Steve Langasek in <https://lists.ubuntu.com/archives/ubuntu-release/2012-August/001688.html>.]

[Evan (ev)]

lp:~ev/oops-repository/whoopsie-daisy r44 starts recording unique system identifiers for each distribution release. It previously only recorded them for all releases combined. It is possible to back-populate this for previous days, but the query for that needs to be further tuned before we can attempt to run it again.

[Evan (ev)]

The aforementioned query is now running (it's currently at 2012/06/07, taking about 20 minutes per day).