launchpad-dev team mailing list archive

[Robert Collins <robert.collins@xxxxxxxxxxxxx>]

<snipped a good write up of how we might answer some questions via
data driven approaches>

Data driven approaches - +1

I think its a great idea to get in place. I've been muttering about
cross-thread contention since I got the TA job, as it happens :).

That said, I don't think GIL contention is an actual effect we're
seeing today : I think things would be much worse if we were
triggering that pathology.

I'd like to put a few constraints on how we do this though : we should
be clear about what experiments we're doing; it shouldn't affect the
continuous rollout project advancement, and we should try to answer
things without affecting production if we can.

On the how-small a thing do we optimise side, we all know where we
are, and there is a lot of work to go.

We *need* though, all of us, to be aware that caring about 100ms is
simply sane *if* we want a web site that is fast and responsive in
much of Asia including NZ, Australia, China and Thailand. Places that
are either physically far away, have shoddy internet or government
filters in the way have the highest constant overhead affecting them :
so its all the more important to remove *our* overheads.

To be clear, I already agree that that wasn't the first thing to fix
*on that page*, and I'm not saying *tune every page to be 200ms
average when you next touch it* - I proposed an incremental slope to
overall faster performance, and the current target is 5 seconds for
99% of requests: that means that when things scale badly, our *high
end* needs to be 5 seconds.

100ms is 2% of that seconds. Anything in a loop that takes 100ms per
iteration will blow past 5 seconds almost trivially.

-Rob