Greetings from London
John Arbash Meinel
john at arbash-meinel.com
Mon May 21 10:44:50 BST 2007
Matthew D. Fuller wrote:
> On Mon, May 21, 2007 at 09:01:40AM +0000 I heard the voice of
> Lars Wirzenius, and lo! it spake thus:
>> On su, 2007-05-20 at 17:03 -0400, Aaron Bentley wrote:
>>> Actually, I believe John Meinel is testing Bazaar with the FreeBSD
>>> tree. No idea whether it's comparable in size, but I'd think
>>> roughly so.
>> The Debian etch source tree is 46 gigabytes, with approximately
>> 3122649 filesystem objects, of which 2868448 are plain files, a
>> total of about one billion lines of text (according to wc), or about
>> 351 lines per file, on average.
> Yah, that's much bigger. Ports is only around 100k (maybe 150k if you
> count files that existed in the past); it's big, but not ginormous.
> It's the length of the history (~140k revs?) that drives bzr up a
Well, I was able to convert Moz, which was 175k revisions. So we've
gotten a lot better.
What really drives us up the wall is that it is 150k files+dirs * 175k
revisions. So inventory.knit gets pounded.
It would be interesting to try converting it again. The last time I
tried we didn't have cvsps-import, and we didn't have the dynamic
inventory deltas (fixed at 26 rather than variable up to 200).
And plugins can set that to whatever they want. One of my experimental
cvsps-import branches sets it to 500. Which actually fits the FreeBSD
use case pretty well. Basically squish the hell out of history because
you aren't going to use it, but make accessing current revisions possible.
Also, with Branch6 having 1000's of branches around doesn't hurt as much
as it used to, either. (A 175k revision-history branch is about 9MB just
for .bzr/branch/revision-history. Getting rid of that file makes each
branch cost about 200 bytes, though spread across several files)
More information about the bazaar