This tutorial is about using travis-ci to run packaging scripts to update datapackages. The purpose is to achieve continuous processing : deliver updated data every time something changes, either the source data or the processing code.
The whole tutorial exposes a simplified version of s-and-p-500-companies which scrapes the list of companies from Wikipedia to creates a datapackage.
- Prerequisite
- How to update the datapackage from the Makefile ?
- Run the project with travis-ci
- More control over your script
- How does it affect usage of my tools ?
- Discipline
- Conclusion
This tutorial assumes you already have a git project that contains a datapackage. This project has a script
directory in which the Makefile retrieve remote data to create the core of the datapackage in the data directory.
Also, you need to have tests to validate the data produced before you publish it. See if you don't have tests, stop reading this and jump to this really good article.
So the scripts/Makefile in your project should look like our example's :
all: valid.txt
List_of_S%26P_500_companies.html:
curl -o List_of_S%26P_500_companies.html "https://en.wikipedia.org/wiki/List_of_S%26P_500_companies"
../data:
mkdir ../data
../data/sp500-companies.csv: ../data List_of_S%26P_500_companies.html scrape-wikipedia.py
python scrape-wikipedia.py
valid.txt: ../data/sp500-companies.csv ../datapackage.json test_data.py
python test_data.py
echo "Datapackage is valid" > valid.txt
.PHONY: all
Where scrape-wikipedia.py extracts the table from the wikipedia page into a csv file and test-data.py tests the data.
First, you'll need a travis account !
If you haven't added the project to travis-ci yet, go to your travis profile and switch the project on. Then you can add a .travis file :
language: python
python:
- 2.7
install:
- pip install --upgrade -r scripts/requirements.txt
script:
- cd scripts
- make
Now every time you commit a change in the repository, the data is validated by travis. You'll be notified if the data is invalid.
How to commit file sp500-companies.csv and push it back to the repository once it has been processed ?
To push data to a git repository on github, you need to be authenticated, and the git way is to use ssh keys. So we'll create a specific key set for the project :
ssh-keygen -t rsa -b 4096 -C "Ssh key for repository ex-continuous-processing" -f ~/.ssh/ex-continuous-processing
This command creates two files :
- the public key
ex-continuous-processing.pubis the file you can distribute so that servers allow your connections - the private key
ex-continuous-processingis the file you need to keep secret on your computer that lets you connect on a server
The easiest way is to be able to push to github is to add the public key to the repository : In the repository's Settings > Deploy Keys, we'll add a key by pasting the content of ex-continuous-processing.pub.
Now we can use the private key to write on the repository.
If we needed to work with several repositories, we could also have created a specific github account, like datasets-update-bot.
Before we set travis-ci, we'll configure our local computer to be able to work on the project and test it.
We need to configure our ssh client to use the key we've created with github. Add these lines to your ~/.ssh/config :
Host github.com
HostName github.com
User git
IdentityFile ~/.ssh/ex-continuous-processing
We should be able to communicate with github :
$ ssh [email protected]
> Hi lexman/ex-continuous-processing! You've successfully authenticated, but GitHub does not provide shell access.
> Connection to github.com closed.
The odds are that your local repository has not been cloned through ssh. That's why we'll add a remote
called publish to our git repository in order to be able to push the changes by ssh.
git remote add publish [email protected]:lexman/ex-continuous-processing.git
This means that we can share the changes on github with the command git push publish. Let's test it :
touch test-write
git add test-write
git commit -m "Testing write from local workstation"
git push publish
This should work without asking password. Now let's roll-back everything to normal :
git rm test-write
git commit -m "Test has succeeded"
git push publish
Now that our workstation is able to write to github, we can automate publication with this new rule at the end of the Makefile :
pushed.txt: valid.txt
git add ../data/sp500-companies.csv
git commit -m "[data][skip ci] automatic update" || exit 0
git push publish
echo "Update has been pushed if there was a change" > pushed.txt
The first line pushed.txt: valid.txt means that validation of data must have succeeded before we try publishing. The last
line creates file pushed.txt so that running make again won't run this code again, unless valid.txt has changed (which means
something has changed in the data).
The git part of this code needs to be explained :
- If there has been no change in
../data/sp500-companies.csv,git addwill succeed but won't add any change to the repository. In this case,|| exit 0in thegit commitline prevents the command from failing... Andmakecan go on withgit pushthat will succeed even if there is nothing to publish. - If there has been some change in file
../data/sp500-companies.csv, the code is pretty straightforward...
We'll also change the main rule of the Makefile, because now the purpose of the project is to publish :
all: pushed.txt
Let's see what it does (it there was no change in the wikipedia page) :
$ make
> git add ../data/sp500-companies.csv
> git commit -m "[data][skip ci] automatic update" || exit 0
> # On branch master
> # Your branch is up-to-date with 'origin/master'.
> #
> # Untracked files:
> # List_of_S%26P_500_companies.html
> # valid.txt
> #
> nothing added to commit but untracked files present
> git push publish
> Warning: Permanently added the RSA host key for IP address '192.30.252.131' to the list of known hosts.
> Everything up-to-date
> echo "Update has been pushed if there was a change" > pushed.txt
Now we'd need to run from travis rather than our workstation.
At the moment, if we push the code above to github, travis-ci will fail because it has not been configured yet. We'll have to enter the private key in the project's settings so that travis can write to github.
Open your private key with your favourite text editor and copy the inner part without the header (-----BEGIN RSA PRIVATE KEY-----) nor
the footer (-----END RSA PRIVATE KEY-----) and paste into quotes it into travis settings as an environment variable called SSH_PUSH_KEY. Your
browser will paste it into one line. Before you add the variable, make sure the Display value in build log button is set to off, so that
your key remains secret all the time.
Now we'll have to recompose the key in the build, by adding a before_script section in the .travis.yml file :
before_script:
# Make the the ssh key available to git
- echo "-----BEGIN RSA PRIVATE KEY-----" >> ~/.ssh/id_rsa
- echo "$SSH_PUSH_KEY" | sed 's/ /\n/g' >> ~/.ssh/id_rsa
- echo "-----END RSA PRIVATE KEY-----" >> ~/.ssh/id_rsa
- chmod 600 ~/.ssh/id_rsa
We also have to add the publish remote to the repository in the before_script section, as we have done in local :
# Add remote for ssh access
- git remote add publish [email protected]:$TRAVIS_REPO_SLUG.git
Note that we used the travis $TRAVIS_REPO_SLUG variable to name the project's repository, so that future forks won't have to
edit the .travis.yml file.
We also have to set a few more things in git, to avoid useless warning messages and set our identity for commits. The the overall
install section should look like :
install:
- pip install --upgrade -r scripts/requirements.txt
# Avoid warnings from git :
- git config --global push.default simple
- git config --global user.email "[email protected]"
- git config --global user.name "Your commit bot name"
After all this configuration, Travis will be able to write to our repository...
Travis has a specific step for deploying a build that is well suited for our task :
deploy:
skip_cleanup: true
provider: script
script: make pushed.txt
on:
all_branches: true
provider: script means we'll run a shell script, and the script is make pushed.txt. The other option skip_cleanup: true
is to prevent travis from reverting the changes we made in the build before publish. all_branches: true should not be difficult
to understand.
Unfortunately this is not enough yet to work, because travis checks out the repository in
detached head state so we've got
to get back on the branch we're building to be able to commit. We can fix it in script:,
but it is easier to add a before_deploy section :
before_deploy:
# Git is in detached HEAD, so we get back on the branch we are supposed to be
- git checkout -B $TRAVIS_BRANCH
After we push this last change, we can see our travis build green ! At last !
We can even see two green builds ! Because when the first one succeeded, it pushed the data... Which triggered a new build !
We have to add [Skip ci] to the commit message to prevent endless builds. So in the Makefile, the git commit
instruction will be changed to :
git commit -m "[data][skip ci] automatic update" || exit 0
When we know the source data has changed, ie the wikipedia page about SP500 in our example, we don't need to check out the project and run it on our computer any more. Now that we have a working travis project, we can just go our project and click Restart build to update the data !
You can configure your project to run every day, in order to follow updates on the source data. As explained in the travis doc, you need to mail the travis support to enable this option.
Also, you'll have to remove the [skip ci] tag from the commit message, because it would prevent the build from being triggered... So
you'll have to bear the double build.
Now this is the end of your journey : once it's done, your project auto updates every day and on every commit ! Congratulations !
Travis will update the data on the new branch ! Thanks to on: all_branches: true and git checkout -B $TRAVIS_BRANCH on before_deploy...
Making the data will still succeed, but the pushing part will fail. The owner of the new project will have to configure a new set of keys : the public key in the github account, and the private key in the travis settings of the project.
Travis will make the data and will run the tests, so you'll know if the data is valid. But hopefully it will skip the deploy phase. So you'll know how a a pull request behave, except if the deploy phase was changed.
Restarting a build which is not the last of a branch will fail. But what did you expect ? Only the last commit is allowed to publish on a branch, because it would be a mess if you could publish out-of-date data... Pushing will display this error message :
To [email protected]:lexman/ex-continuous-processing.git
! [rejected] master -> master (non-fast-forward)
error: failed to push some refs to '[email protected]:lexman/ex-continuous-processing.git'
hint: Updates were rejected because the tip of your current branch is behind
hint: its remote counterpart. Integrate the remote changes (e.g.
hint: 'git pull ...') before pushing again.
hint: See the 'Note about fast-forwards' in 'git push --help' for details.
make: *** [pushed.txt] Error 1
Once continuous processing has been achieved, data will update automatically, but it's not effortless.
Maintaining this system is demanding because experience shows builds do break. In our example the structure of the wikipedia page is likely to change when people edit it so the scraping script should be fixed. Also travis will be updated and default python could migrate to version 3.
When build breaks, experience with continuous integration shows that it should be repaired quickly or it's never done. Also all the value of bringing online new data, is void if existing ones are broken. That's why it's important to be able to work locally to solve broken builds.
You don't want to share some code or some data that does not work. That's why we configured our project to work locally. You should always run the project before you push, to detect failures. And avoid the terrible pattern fix -> push-to-travis -> find-a-new-bug-in-the-logs -> redo : it's very time consuming and harmful for your co-workers and people using your data.
And remember, before you fix a bug, you should add a test to prevent future regressions !
With a Makefile, some tests, travis-ci and a bit of maintenance we can achieve continuous processing to provide up-to-date quality data : anyone can contribute through the well know process of pull-request and validation and publication of datasets will be automatic. It's very well suited for small data (less then 300 MB) and when the whole processing is short (less 10 minutes), That's why it is perfect for core datapackges !
The impatient will implement this tutorial from the shorter cheat sheet !