git.tremily.us Git - git.git/log

Jakub Narebski [Thu, 11 Dec 2008 00:33:29 +0000 (01:33 +0100)]

gitweb: cache $parent_commit info in git_blame()

Luben Tuikov changed 'lineno' link from leading to commit which gave
current version of given block of lines, to leading to parent of this
commit in 244a70e (Blame "linenr" link jumps to previous state at
"orig_lineno").  This made possible data mining using 'blame' view.

The current implementation calls rev-parse once per each blamed line
to find parent revision of blamed commit, even when the same commit
appears more than once, which is inefficient.

This patch mitigates this issue by caching $parent_commit info in
%metainfo, which makes gitweb call rev-parse only once per each
unique commit in the output from "git blame".

In the tables below you can see simple benchmark comparing gitweb
performance before and after this patch

File               | L[1] | C[2] || Time0[3] | Before[4] | After[4]
====================================================================
blob.h             |   18 |    4 || 0m1.727s |  0m2.545s |  0m2.474s
GIT-VERSION-GEN    |   42 |   13 || 0m2.165s |  0m2.448s |  0m2.071s
README             |   46 |    6 || 0m1.593s |  0m2.727s |  0m2.242s
revision.c         | 1923 |  121 || 0m2.357s | 0m30.365s |  0m7.028s
gitweb/gitweb.perl | 6291 |  428 || 0m8.080s | 1m37.244s | 0m20.627s

File               | L/C  | Before/After
=========================================
blob.h             |  4.5 |         1.03
GIT-VERSION-GEN    |  3.2 |         1.18
README             |  7.7 |         1.22
revision.c         | 15.9 |         4.32
gitweb/gitweb.perl | 14.7 |         4.71

As you can see the greater ratio of lines in file to unique commits
in blame output, the greater gain from the new implementation.

  Legend:

  [1] Number of lines:
      $ wc -l <file>
  [2] Number of unique commits in the blame output:
      $ git blame -p <file> | grep author-time | wc -l
  [3] Time for running "git blame -p" (user time, single run):
      $ time git blame -p <file> >/dev/null
  [4] Time to run gitweb as Perl script from command line:
      $ gitweb-run.sh "p=.git;a=blame;f=<file>" > /dev/null 2>&1

The gitweb-run.sh script includes slightly modified (with adjusted
pathnames) code from gitweb_run() function from the test script
t/t9500-gitweb-standalone-no-errors.sh; gitweb config file
gitweb_config.perl contents (again up to adjusting pathnames; in
particular $projectroot variable should point to top directory of git
repository) can be found in the same place.

Discussion
~~~~~~~~~~

A possible future improvement would be to open a bidi pipe to
"git cat-file --batch-check", (like in Git::Repo in gitweb caching by
Lea Wiemann), feed $long_rev^ to it, and parse its output, which is
in the following form:

  926b07e694599d86cec668475071b32147c95034 commit 637

This would mean one call to git-cat-file for the whole 'blame' view,
instead of one call to git-rev-parse per each unique commit in blame
output.

Yet another solution would be to change use of validate_refname() to
validate_revision() when checking script parameters (CGI query or
path_info), with validate_revision being something like the following:

  sub validate_revision {
        my $rev = shift;
        return validate_refname(strip_rev_suffixes($rev));
  }

so we don't need to calculate $long_rev^, but can pass "$long_rev^" as
'hb' parameter.

This solution has the advantage that it can be easily adapted to future
incremental blame output.

Signed-off-by: Jakub Narebski <jnareb@gmail.com>
Acked-by: Luben Tuikov <ltuikov@yahoo.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>