[PATCH] diff: use a threshold on similarity index before using word-diff (issue5965)

[Yuya Nishihara]

On Fri, 24 Aug 2018 10:05:14 +0200, Boris FELD wrote:
> On 23/08/2018 14:48, Yuya Nishihara wrote:
> > On Wed, 22 Aug 2018 21:35:31 +0900, Yuya Nishihara wrote:
> >> On Tue, 21 Aug 2018 17:11:51 +0200, Denis Laxalde wrote:
> >>> Yuya Nishihara a écrit :
> >>>> On Tue, 21 Aug 2018 14:10:33 +0200, Denis Laxalde wrote:
> >>>>> # HG changeset patch
> >>>>> # User Denis Laxalde <denis.laxalde at logilab.fr>
> >>>>> # Date 1534853203 -7200
> >>>>> #      Tue Aug 21 14:06:43 2018 +0200
> >>>>> # Node ID c43df6ff42d26163d19e99e15a3cf3094020d822
> >>>>> # Parent  c62184c6299c09d2e8e7be340f9aee138229cb86
> >>>>> # Available At http://hg.logilab.org/users/dlaxalde/hg
> >>>>> #              hg pull http://hg.logilab.org/users/dlaxalde/hg -r c43df6ff42d2
> >>>>> # EXP-Topic issue5965
> >>>>> diff: use a threshold on similarity index before using word-diff (issue5965)
> >>>>>
> >>>>> The threshold is chosen quite arbitrarily with a value of 0.5. It does
> >>>>> not change the results of test-diff-color.t whereas higher values (e.g.
> >>>>> 0.6) would. Looking at what this produces on some changesets in recent
> >>>>> history (e.g. 037debbf869c or 7acec9408e1c), this significantly improves
> >>>>> diff readability.
> >>>>>
> >>>>> Similarity index is computed using difflib.SequenceMatcher's ratio()
> >>>>> method; this is documented as being "expensive", but other faster methods
> >>>>> (that compute an upper bound value) do not give good results.
> >>>>> Nevertheless, since we compute this ratio on each hunk which are usually
> >>>>> small, this might not be problematic in most cases. Also, as we'd
> >>>>> short-circuit computation of inline colors for those hunks that are not
> >>>>> similar enough, this "expensive" ratio computation might also be
> >>>>> compensated.
> >>>> Can you test this against a large BLOB-ish diff (such as machine-generated
> >>>> 10k-line JSON, a binary in Intel HEX format, etc.)? Last time I faced that,
> >>>> the original difflib-based algorithm was painfully slow (~100s-ish to yield
> >>>> one hunk), which made me think the word-diff should never be turned on by
> >>>> default.
> >>> I've set up a test repo with some JSON at
> >>> https://bitbucket.org/dlax/hg-worddiff-tests. As far as I can tell,
> >>> there's no significant difference when diffing the last changeset;
> >> Thanks, but it looks cheaper to compute than the stuff I had at work. I'll
> >> try to collect some number if I get a chance.
> >    $ hg diff -c REV --color=always --config diff.word-diff=true --time > /dev/null
> >    (orig) 1.250sec
> >    (new)  1259.490sec
> >
> > It's an ASCII-fied FPGA image (called tabular text file), containing ~320k
> > decimal numbers plus commas (so ~1000k words in our word-diff.) And there
> > are some large hunks as it is a diff of two similar BLOBs split into chunks
> > per N bytes.
> These files seems to have interesting characteristics that would be 
> useful for performance testing.
> 
> Would it be possible to get examples or redacted versions of such files?

(CC the list again)

It's a closed binary, but you can see how slow the difflib is by using random
data.

  $ dd if=/dev/urandom bs=1k count=100 | hexdump -v -e '16/1 "%3u," "\n"'

IIRC, the computation cost of difflib is more sensitive to input data than
Mercurial's bdiff.