In spite of the numerous deploy strategies, I wanted to deploy to a hosting provider via SSH. Here are some explanations on why and how to do it.Continue reading
It is possible to deploy via SSH from Travis CI by combining the
travis encrypt-file and
Why using Travis CI for this?
I will keep it short but in essence:
- I want reproducible builds and to deploy only tested artifacts;
- I want a clean state for each build;
- I don't want to worry about my flaky Internet connection bandwidth;
- I kind of fancy merging a PR on GitHub mobile web interface and see a website being rebuilt and deployed without further manual operation.
I personally use this process:
- to deploy this blog;
- to deploy a 3K+ pages static website with a 20 minutes build time and content indexation with Algolia;
- to assemble an Asciidoc based book into a website and publisher-ready LibreOffice document.
Build and deploy workflow
Here is what happens during the build:
- Travis installs your project dependencies;
- Travis runs the tests;
- Travis builds the artifacts (HTML pages or other kinds of files you want to move somewhere else);
- Travis runs the deploy strategy in order to transfer its files on disk to a remote machine.
The deploy strategy can be:
- scp all the things! Nice but limited in term of features, especially if you need to prevent certain files to be transferred (it is all or nothing);
- rsync The recommended way to transfer assets over the wire, mostly because of the various local to remote file comparison mechanisms – by using content fingerprinting, date changes, file deletion etc;
- solid archive upload Can be handy if you don't trust the wire and fancy untarring a PGP encrypted archive, assessing a tarball checksum after a successful transfer etc;
- a combination of these and others Which is perfect if you want to push static files as well as to index their content in a remote database with secured write credentials.
With any of the strategies, you can upload to the remote folder, or to a temporary folder and swap them out once the transfer is complete. Which can be useful to prevent UI or service disruptions – because some assets calls are not in sync anymore. (eg: an updated page refering to a CSS file which has not been transfered yet).
Travis and sensitive informations
You can benefit from custom SSH keys on Travis CI if you are a paid user.
Connecting from a Travis build box (or any CI system really) to a remote host implies to have the private SSH key on the CI box (
*_rsa files) and its associated public SSH key on the remote host end (
But you certainly don't want your private key to be available in your Git repo or to be seen in the CI build logs. It is as secure as hosting tax heaven data on a Drupal website.
travis ruby client supports file encryption since its version 1.7 to help us a bit.
Setup SSH encryption key
So let's recap what we need to do to securely prepare our SSH tunnel between our remote host and Travis:
- Generate a dedicated SSH key (it is easier to isolate and to revoke);
- Encrypt the private key to make it readable only by Travis CI (so as we can commit safely too!);
- Copy the public key onto the remote SSH host;
- Cleanup after unnecessary files;
- Stage the modified files into Git.
Which translates into:
Setup SSH decryption in Travis job
We still need to setup a few things before we are able to enact anything:
- Preventing the SSH interactive prompt when connecting to a host for the first time (which means, for each build);
- Decrypting the encrypted SSH private key;
- Adding the key to the current
ssh-agentto make any SSH-based command agnostic to the private key location.
The result of the previous
travis encrypt-file call will generate a content similar to this in our
Notice: we do extract the
deploy_rsa in Travis
/tmp folder to avoid deploying the decrypted key by any mean.
Travis build lifecycle steps are thoroughly documented.
I mainly use
before_deploy to setup the SSH agent for two reasons:
- it happens after the
scriptstage: we then avoid any possibility for third party scripts to leak active keys (npm worm, remember?);
- the build will fail if the deploy setup is incorrect.
It is now time to deploy!
Travis deploy script
My favourite way of deploying is to use the
deploy script lifecycle event.
<dir> is the folder you want to recursively upload remotely. Best is to assemble and copy all the files you want to copy in this folder, like:
$TRAVIS_BUILD_DIR/_distif your generator is Jekyll;
$TRAVIS_BUILD_DIR/publicif your generator is hexo.
skip_cleanup is kind of mandatory otherwise Travis resets the git directory state.
In my case I want to deploy only when new modifications land
master branch but you might want to deploy against different conditions too.
You can even run several deploy commands, to push content and to index it in a search engine:
A last warning though: if your deploy fails for some reason, is incomplete or if the host network connection shuts down, you are screwed.
Why? Because the build status will be green and unless you inspect the logs you will not notice it. Indeed, deploy scripts do not affect the build status. Use
after_success lifecycle event if it does not suit you — it feels a bit more hackish though.
Again, you could:
- push to a temporary directory:
scp ... user@ssh:/tmp/$TRAVIS_BUILD_ID;
- swap the deploy folder with the production one:
rm -rf /tmp/old.build && mv path/to/files /tmp/old.build && mv /tmp/$TRAVIS_BUILD_ID path/to/files.
It is up to you and the volume and ratio of changed files you have to send over the wire.
I hope this article helped you understand the workflow of a custom deployment with Travis CI. And make you realise you have now all the tools needed to avoid deploying from your machine.
This is also a great way to document the deployment process and avoid the bottleneck of relying on a person machine setup to push new changes online.
You can now deploy several times a day without much more effort than merging a pull request. Isn't that sweet?
Now, if you do want to prevent pushing crap (like badly formatted content etc.), you are good to implement defensive tools like linters, functional tests and markup validators at