GitLab CI and conda

I setup GitLab to host several projects at work and I have been quite pleased with it. I read that setting GitLab CI for test and deployment was easy so I decided to try it to automatically run the test suite and the sphinx documentation.

I found the official documentation to be quite good to setup a runner so I won't go into details here. I chose the Docker executor.

Here is my first .gitlab-ci.yml test:

image: python:3.4

before_script:
  - pip install -r requirements.txt

tests:
  stage: test
  script:
    - python -m unittest discover -v

Success, it works! Nice. But... 8 minutes 33 seconds build time for a test suite that runs in less than 1 second... that's a bit long.

Let's try using some caching to avoid having to download all the pip requirements every time. After googling, I found this post explaining that the cache path must be inside the build directory:

image: python:3.4

before_script:
  - export PIP_CACHE_DIR="pip-cache"
  - pip install -r requirements.txt

cache:
  paths:
    - pip-cache

tests:
  stage: test
  script:
    - python -m unittest discover -v

With the pip cache, the build time went down to about 6 minutes. A bit better, but far from acceptable.

Of course I knew the problem was not the download, but the installation of the pip requirements. I use pandas which explains why it takes a while to compile.

So how do you install pandas easily? With conda of course! There are even some nice docker images created by Continuum Analytics ready to be used.

So let's try again:

image: continuumio/miniconda3:latest

before_script:
  - conda env create -f environment.yml
  - source activate koopa

tests:
  stage: test
  script:
    - python -m unittest discover -v

Build time: 2 minutes 55 seconds. Nice but we need some cache to avoid downloading all the packages everytime. The first problem is that the cache path has to be in the build directory. Conda packages are saved in /opt/conda/pkgs by default. A solution is to replace that directory with a link to a local directory. It works but the problem is that Gitlab makes a compressed archive to save and restore the cache which takes quite some time in this case...

How to get a fast cache? Let's use a docker volume! I modified my /etc/gitlab-runner/config.toml to add two volumes:

[runners.docker]
  tls_verify = false
  image = "continuumio/miniconda3:latest"
  privileged = false
  disable_cache = false
  volumes = ["/cache", "/opt/cache/conda/pkgs:/opt/conda/pkgs:rw", "/opt/cache/pip:/opt/cache/pip:rw"]

One volume for conda packages and one for pip. My new .gitlab-ci.yml:

image: continuumio/miniconda3:latest

before_script:
  - export PIP_CACHE_DIR="/opt/cache/pip"
  - conda env create -f environment.yml
  - source activate koopa

tests:
  stage: test
  script:
    - python -m unittest discover -v

The build time is about 10 seconds!

Just a few days after my tests, GitLab announced GitLab Container Registry. I already thought about building my own docker image and this new feature would make it even easier than before. But I would have to remember to update my image if I change my requirements. Which I don't have to think about with the current solution.

Switching from git-bigfile to git-lfs

In 2012, I was looking for a way to store big files in git. git-annex was already around, but I found it a bit too complex for my use case. I discovered git-media from Scott Chacon and it looked like what I was looking for. It was in Ruby which made it not super easy to install on some machines at work. I thought it was a good exercise to port it to Python. That's how git-bigfile was born. It was simple and was doing the job.

Last year, I was thinking about giving it some love: port it to Python 3, add some unittests... That's about when I switched from Gogs to Gitlab and read that Gitlab was about to support git-lfs.

Being developed by GitHub and with Gitlab support, git-lfs was an obvious option to replace git-bigfile.

Here is how to switch a project using git-bigfile to git-lfs:

  1. Make a list of all files tracked by git-bigfile:

    $ git bigfile status | awk '/pushed/ {print $NF}' > /tmp/list
    
  2. Edit .gitattributes to replace the filter. Replace filter=bigfile -crlf with filter=lfs diff=lfs merge=lfs -text:

    $ cat .gitattributes
    *.tar.bz2 filter=lfs diff=lfs merge=lfs -text
    *.iso filter=lfs diff=lfs merge=lfs -text
    *.img filter=lfs diff=lfs merge=lfs -text
    
  3. Remove all big files from the staging area and add them back with git-lfs:

    $ git rm --cached $(cat /tmp/list)
    $ git add .
    $ git commit -m "Switch to git-lfs"
    
  4. Check that the files were added using git-lfs. You should see something like that:

    $ git show HEAD
    diff --git a/CentOS_6.4/images/install.img
    b/CentOS_6.4/images/install.img
    index 227ea55..a9cc6a8 100644
    --- a/CentOS_6.4/images/install.img
    +++ b/CentOS_6.4/images/install.img
    @@ -1 +1,3 @@
    -5d243948497ceb9f07b033da62498e52269f4b83
    +version https://git-lfs.github.com/spec/v1
    +oid
    sha256:6fcaac620b82e38e2092a6353ca766a3b01fba7f3fd6a0397c57e979aa293db0
    +size 133255168
    
  5. Remove git-bigfile cache directory:

    $ rm -rf .git/bigfile
    

Note: to push files larger than 2.1GB to your gitlab server, wait for this fix. Hopefully it will be in 8.4.3.

crontab and date

The other day, I wanted to add a script to the crontab and to redirect the output to a file including the current date. Easy. I have used the date command many times in bash script like that:

current_date=$(date +"%Y%m%dT%H%M")

So I added the following to my crontab:

0 1 * * * /usr/local/bin/foo > /tmp/foo.$(date +%Y%m%dT%H%M).log 2>&1

And... it didn't work...

I quickly identified that the script was working properly when run from the crontab (it's easy to get a script working from the prompt, not running from the crontab due to incorrect PATH). The problem was the redirection but I couldn't see why.

I googled a bit but didn't find anything...

I finally looked at the man pages:

$  man 5 crontab

     ...
     The  ``sixth''  field  (the  rest of the line) specifies the command to be run.  The entire command portion of the line, up to a
     newline or % character...

Here it was of course! % is a special character. It needs to be escaped:

0 1 * * * /usr/local/bin/foo > /tmp/foo.$(date +\%Y\%m\%dT\%H\%M).log 2>&1

Lesson to remember: check the man pages before to google!

Compile and install Kodi on iPad without jailbreak

With iOS 9 and Xcode 7 it's finally possible to compile and deploy apps on your iPhone/iPad with a free Apple developer account (no paid membership required).

I compiled XBMC/Kodi many times on my mac but had never signed an app with Xcode before and it took me some time to get it right. So here are my notes:

First thanks to memphiz for the iOS9 support!

I compiled from his ios9_workaround branch, but it has been merged to master since:

$ git clone https://github.com/xbmc/xbmc.git Kodi
$ cd Kodi
$ git remote add memphiz https://github.com/Memphiz/xbmc.git
$ git fetch memphiz
$ git checkout -b ios9_workaround memphiz/ios9_workaround

Follow the instructions from the README.ios file:

$ git submodule update --init addons/skin.re-touched
$ cd tools/depends
$ ./bootstrap
$ ./configure --host=arm-apple-darwin
$ make -j4
$ make -j4 -C target/binary-addons
$ cd ../..
$ make -j4 -C tools/depends/target/xbmc
$ make clean
$ make -j4 xcode_depends

Start Xcode and open the Kodi project. Open the Preferences, and add your Apple ID if not already done:

/images/add_account.png

Select the Kodi-iOS target:

/images/kodi_ios_target.png

Change the bundle identifier to something unique and click on Fix Issue to create a provisioning profile.

/images/bundle_identifier.png

Connect your device to your mac and select it:

/images/device.png

Click on Run to compile and install Kodi on your device!