[Success!] Re: [MERGE][1.6.1] 1.6 fetch regression
John Arbash Meinel
john at arbash-meinel.com
Fri Aug 29 15:56:34 BST 2008
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
David Ingamells wrote:
> What are the plans for the 1.6.1 branch? Possible release dates?
>
> I'd like to move our team up to a release of 1.6 with John Arbash
> Meinel's branch performance fixes, but since this will be used by a
> large-ish team I will only upgrade if there is a release that has
> undergone the standard regression testing - no criticism of John's great
> work implied ;^). At worst I'll only need to wait for 1.7 ...
Oh, and I should mention. *every* commit to bzr.dev goes through the full 13+k
unit test regression suite.
This wasn't strictly a regression, just a change of code path which triggered
a latent issue in the other code. It only happened under the specific
permutation of Robert's new fetching algorithm allowing us to fetch
more-at-a-time, and the win32 patch causing us to read only 64kB from a
socket. (So we were fetching 10+MB in one request, but only reading and
buffering it 64kB at a time, causing us to reallocate the buffer far too often.)
It also only shows up on branching from-scratch because you are downloading
more of the repository, only really shows up for large projects, and only if
you are using bzr+ssh. (You need >1MB downloaded at a time before the effects
start to become severe, it just happens to be an O(N^2) problem with small
coefficients. So for most use it never was an issue, until you try to download
100+ MB at one time, and the quadratic term becomes a real issue. For 100MB we
would re-allocate ~1,600 times. So it is N*(N+1)/2 *64kB = 76GB of "churn".)
Anyway, you are welcome to pick your route. I hope to have 1.6.1 final before
1.7rc1 is released.
John
=:->
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org
iD8DBQFIuA4iJdeBCYSNAAMRAmjjAJ9nRBjvgZ2oCy4hqWlR05QreuOtVwCfbN6R
tg5350ld+XySq0BGr/We3cw=
=BVK0
-----END PGP SIGNATURE-----
More information about the bazaar
mailing list