Re: [Gnu-arch-users] patch: automatic cacherev and smarter get

[Mikhael Goikhman]

On 23 Nov 2005 18:59:08 +0000, Andrew Suffield wrote:
> 
> On Wed, Nov 23, 2005 at 06:16:11PM +0100, Alfred M. Szmidt wrote:
> >    > Caveats:
> >    > * the number 50 is hard coded; it really should be a value that
> >    > the user can customize.
> > 
> >    Before this is viable, it *must* be possible to turn this behaviour
> >    off. On large trees, cacherevs are massively wasteful of space;
> >    revision libraries are often a more efficient solution.

One may have a reason to increase "50", but people who keep long 1000
revision branches without a single cacherev should really be deported.

> > It depends.  If you have 50 small patches that fix typos or so, then
> > on a large tree this is wastefull.
> 
> Or even large changes that only touch half the files - the hardlinking
> of revision libraries is *really* good most of the time, and cacherevs
> have nothing similar.

This is only theoretically correct. In reality, cacherevs are compressed,
so they are usually more disk friendly then a hard-linked revlib.

Besides, a revlib consumes a _huge_ amount of inodes that constitutes a
real problem in many cases (compared to one tarball per 50 revisions).

> We had this thread years ago, and to summarise:
> 
>  - you want revision library entries for people doing any kind of real
>    work with tla

Disagree. Revision libraries although nice to have are not needed for
real work with tla. Pristines and periodical cacherevs are often enough.
It's true tla has some unoptimal archive operations, but this is fixable.

>  - cacherevs are only really interesting to people who want to use
>    'get'

Cacherevs are helpful with merges not less than with 'get'.

In short, a much more appropriate summary would be: greedy revlibs on
every developer account is quite expensive, while cacherevs are not.

Regards,
Mikhael.