hg convert from SVN repo getting stuck

[Malte Helmert]

On 26.04.2014 19:36, Malte Helmert wrote:
> Dear group,
> 
> I'm having some problems with hg convert:
[...]
> Any suggestions?

Hi again,

while the above email was stuck in moderation (or gmane?), I've done
some tests to reproduce the same behaviour outside of hg convert. This
time, I used the most recent hg from the stable branch:

$ ./hg version
Mercurial Distributed SCM (version 3.0-rc+28-d36440d84328)
[...]

I produced smaller versions of the file that was modified in the
changeset that caused hg convert to get stuck. For this I truncated the
file at $SIZE bytes for various values of SIZE. Then I created a
pristine repository and committed first the truncated before-changeset
version of the file and then the truncated after-changeset version:

$ hg init testrepo
$ cd testrepo
$ cp ../before file.txt
$ hg add file.txt
$ hg commit -m "first commit"
$ cp ../after file.txt
$ hg diff file.txt | wc -l
$ hg commit -m "second commit"

For various values of SIZE, here is how long the "hg diff" and second
commit took, along with the size of the diff:

SIZE=1M: diff 0.29s (22895 lines), commit 0.20s
SIZE=2M: diff 0.81s (47439 lines), commit 0.61s
SIZE=3M: diff 1.76s (72830 lines), commit 1.47s
SIZE=4M: diff 7.43s (97965 lines), commit 6.99s
SIZE=5M: diff 4.43s (122899 lines), commit 3.91s
SIZE=6M: diff 88.36s (147787 lines), commit 90.56s
SIZE=7M: diff 10.51s (172759 lines), commit 9.74s
SIZE=8M: diff 202.15s (198395 lines), commit 202.04s
SIZE=9M: diff 25.33s (223536 lines), commit 24.55s
SIZE=10M: diff 162.89s (248567 lines), commit 159.23s
SIZE=12M: diff 132.42s (299271 lines), commit 130.41s
SIZE=14M: diff 53.40s (350046 lines), commit 51.73s
SIZE=16M: diff 2822.02s (400085 lines), commit 2802.49s
SIZE=18M: diff 16687.10s (450685 lines), commit 16856.23s
SIZE=20M: diff 20673.23s (501958 lines), commit 20639.25s
SIZE=22M: diff 25507.32s (552989 lines), commit 25758.22s
SIZE=24M: diff 30875.81s (603709 lines), commit 31276.82s
...
(Larger sizes did not terminate yet.)

So there is some nasty scaling here. As a reference point, for the full
file (35 MB), "diff -u" (GNU diff) takes 0.9 seconds.

So I seem to have hit a bad case for Mercurial's diff algorithm. Would
there be interest in uncovering the reason for this, and possibly
modifying the diff algorithm to address this?

Malte