Prototype "improved_chk_index"

[Martin Pool]

2009/10/29 Ian Clatworthy <ian.clatworthy at canonical.com>:
> John Arbash Meinel wrote:
>
>> However, the proof of concept came together a lot faster than I expected
>> (makes me wish I had pushed for it as part of 2.0 :). And the
>> performance results are pretty good, so I figured I'd share the results
>> so far.
>
> Sounds really promising. *Very* well done.
>
>> I may not get back to this for a while. Mostly because we've stated "no
>> new formats" for 2.1. So while it might be able to be a dev format, it
>> would not get into active use for > 6 months.
>
> That depends. If features like nested trees appear in the development
> format prior to 2.1, it may get more use than we realise. I don't think
> delaying landing stuff into the development format is a good idea.
> Historically, it's taken some time to bed down performance-related
> changes, partly because it takes a long time to QA the changes on lots
> of operations across lots of data sets. The game "whack-a-mole" springs
> to mind.
>
> If you have changes suitable for the development format, I'd prefer to
> see them land when they are ready - with a preference for early in the
> cycle over later. If nothing else, it helps me when benchmarking to have
> the development format reflecting our latest and greatest. If we still
> suck on certain data sets, at least I can detect that and communicate it
> rather than hoping pending changes will improve things.

I think landing code into something that's not commonly used is
probably better than letting it languish in a separate branch.

I think we need to do some work on making the process of format change
easier before we contemplate either recommending that upgrade, or
having interesting features only available in a nondefault format.
The more features that we put into development formats before doing
this work, the more unshipped inventory we will accumulate.  Also,
we'll have features that we want to tell people about, but which they
get a poor experience and poor interoperability in actually trying to
use.

-- 
Martin <http://launchpad.net/~mbp/>