How to quickly find all git repos under a directory

Question

The following bash script is slow when scanning for .git directories because it looks at every directory. If I have a collection of large repositories it takes a long time f

Answer 1

I've taken the time to copy-paste the script in your question, compare it to the script with your own answer. Here some interesting results:

Please note that:

• I've disabled the git pull by prefixing them with a echo
• I've removed also the color things
• I've removed also the .ignore file testing in the bash solution.
• And removed the unecessary > /dev/null here and there.
• removed pwd calls in both.
• added -prune which is obviously lacking in the find example
• used "while" instead of "for" which was also counter productive in the find example
• considerably untangled the second example to get to the point.
• added a test on the bash solution to NOT follow sym link to avoid cycles and behave as the find solution.
• added shopt to allow * to expand to dotted directory names also to match find solution's functionality.

Thus, we are comparing, the find based solution:

#!/bin/bash

find . -name .git -type d -prune | while read d; do
   cd $d/..
   echo "$PWD >" git pull
   cd $OLDPWD
done

With the bash shell builting solution:

#!/bin/bash

shopt -s dotglob

update() {
    for d in "$@"; do
        test -d "$d" -a \! -L "$d" || continue
        cd "$d"
        if [ -d ".git" ]; then
            echo "$PWD >" git pull
        else
            update *
        fi
        cd ..
    done
}

update *

Note: builtins (function and the for) are immune to MAX_ARGS OS limit for launching processes. So the * won't break even on very large directories.

Technical differences between solutions:

The find based solution uses C function to crawl repository, it:

• has to load a new process for the find command.
• will avoid ".git" content but will crawl workdir of git repositories, and loose some times in those (and eventually find more matching elements).
• will have to chdir through several depth of sub-dir for each match and go back.
• will have to chdir once in the find command and once in the bash part.

The bash based solution uses builtin (so near-C implementation, but interpreted) to crawl repository, note that it:

• will use only one process.
• will avoid git workdir subdirectory.
• will only perform chdir one level at a time.
• will only perform chdir once for looking and performing the command.

Actual speed results between solutions:

I have a working development collection of git repository on which I launched the scripts:

• find solution: ~0.080s (bash chdir takes ~0.010s)
• bash solution: ~0.017s

I have to admit that I wasn't prepared to see such a win from bash builtins. It became more apparent and normal after doing the analysis of what's going on. To add insult to injuries, if you change the shell from /bin/bash to /bin/sh (you must comment out the shopt line, and be prepared that it won't parse dotted directories), you'll fall to ~0.008s . Beat that !

Note that you can be more clever with the find solution by using:

find . -type d \( -exec /usr/bin/test -d "{}/.git" -a "{}" != "." \; -print -prune \
       -o -name .git -prune \)

which will effectively remove crawling all sub-repository in a found git repository, at the price of spawning a process for each directory crawled. The final find solution I came with was around ~0.030s, which is more than twice faster than the previous find version, but remains 2 times slower than the bash solution.

Note that /usr/bin/test is important to avoid search in $PATH which costs time, and I needed -o -name .git -prune and -a "{}" != "." because my main repository was itself a git subrepository.

As a conclusion, I won't be using the bash builtin solution because it has too much corner cases for me (and my first test hit one of the limitation). But it was important for me to explain why it could be (much) faster in some cases, but find solution seems much more robust and consistent to me.