Bug 1320977 - performance tweaks

[dylanwh]

In #672 performance regressed a bit, so let's win that back here.
The theme of this change is to load bugzilla classes earlier. It seems to save about 40ms per API call.

Before:

After

[dylanwh]

@purelogiq can you review this?
Maybe verify the profiling numbers?
(see https://www.youtube.com/watch?v=1hrdVxI0uFM for how, or ask me in IRC)

[purelogiq]

Nice

[purelogiq]

Interesting use of mojo in a not-yet-mojo app :)

[purelogiq]

@dylanwh Will verify the profiling numbers Monday, missed your comment there.

[dylanwh]

cool, I'll wait to merge. if it's not fast enough-er I can find another slow part.

[purelogiq]

@dylanwh First off, awesome stuff with the flamegraphs and profiling. When I compared the profile on master vs this branch I saw that webservice_user_get is faster in this PR but the overall request is the same or even slower. Can you verify that I'm doing this right?

Vid on google drive: https://drive.google.com/file/d/194cHcyvlZpyP5JvXUr1eyY42tCmp0DlA/view

[dylanwh]

Are you restarting httpd between tests?

But actually, looking at the video this is mostly expected -- the difference disappears after the first request because the optimization here is just loading some modules earlier.

However "first requests" are quite common -- because it's per worker-process and there are hundreds of those, and they get restarted somewhat frequently.

[dylanwh]

Looks like the remaining balance of time is in using Safe->reval(). I wanted to fix that in #321 but it's a little tricky.

[dylanwh]

@purelogiq I don't want to check now because I really want to get this test case posted, but tomorrow can you try benchmarking $cpt->reval($serialized) replaced with just eval($serialized)?

[purelogiq]

@dylanwh Retested and still getting the same results :/

Changing $cpt->reval($serialized) in DB/Schema.pm as below and removing the Safe import:
(preface I have no clue what I'm doing in this area of the code :). No idea if/why I need a 'my' inside eval. No clue of the implications of changing this, use at your own risk!)

my $stuff= eval("my $serialized") || die "Unable to restore cached schema: " . $@;
$thawed_hash = $stuff;

results in a pretty good speed up:

BEFORE:

AFTER:

[purelogiq]

@dylanwh I'm still good on r+ing this because I've been testing post-warmup and the pre-warmup speed improvement like you mentioned maybe be worth it.

[dylanwh]

Yup, I'll take it, and prioritize #321 later.