[PATCH 5 of 6 V2] obscache: instantiate the cache and keep it warm

[Jun Wu]

Excerpts from Augie Fackler's message of 2017-05-24 13:41:53 -0400:
> I'm extremely sympathetic to wanting to avoid new deps on revlog, but in
> this case we're talking about a 500ms win on 'hg id' in the Mercurial
> repository, which is significant savings.

I have already demonstrated the index works. They are tagged as "RFC" but
actually quite complete and I can polish them up in 2 days. And my latest
analysis indicates it could be made to be just 20ms difference from
obscache.

I'd note that obscache is also incomplete in this series - the way it
handles strip (could be triggered by histedit/rebase --abort) is bad for
huge repos (the next command takes 7 seconds).

When talking about performance. Please take that 7 seconds after strip (or
histedit --abort) into consideration and not only just read operations.

> Realistically, any non-revlog storage engine is going to have to work out
> similar cache invalidation hooks anyway, isn't it?

True. But the index stuff does not depend on changelog. Its single source of
truth is obsstore. So it's easier.