Yak shaving with Vagrant, Travis-CI and AWS

Tldr; Don’t use the vagrant package from your distribution if you intent to build plugins.

I’ve just finished setting up CI pipeline for a personal project. The project has an ansible playbook that I want to exercise every time there’s a commit or a PR. While completing the task I shaved a yak and narrowly avoided shaving a whole herd. I planned to use Vagrant in my Travis-CI pipeline to start an instance in AWS, run the playbook, look at the result and terminate the instance. Vagrant, Travis-CI and AWS are pretty common tools, so I was surprised at the wrangling involved before I ended up with a solution. I thought I’d document my findings to minimise the chance that others will have the same experience.

Finding #1: Travis’ default build agent has an old Vagrant

The default build agent is based on Ubuntu 12.04 LTS Server which ships with Ansible 1.0 in its apt repo. The vagrant-aws plugin requires Vagrant 1.2. Fortunately they have a Ubuntu 14.04 LTS Server beta which has a newer Vagrant in the apt repo.

Finding #2: The AWS-Vagrant plugin won’t build with a newer Vagrant because of missing libraries and tools

$ vagrant plugin install vagrant-aws
Installing the 'vagrant-aws' plugin. This can take a few minutes...
/usr/lib/ruby/1.9.1/rubygems/installer.rb:562:in `rescue in block in build_extensions': ERROR: Failed to build gem native extension. (Gem::Installer::ExtensionBuildError)

        /usr/bin/ruby1.9.1 extconf.rb
/usr/lib/ruby/1.9.1/rubygems/custom_require.rb:36:in `require': cannot load such file -- mkmf (LoadError)
    from /usr/lib/ruby/1.9.1/rubygems/custom_require.rb:36:in `require'
    from extconf.rb:4:in `<main>'

Searching for the mkmf LoadError in the context of Debian and Ubuntu gives recommendations to install a few devel packages including ruby-dev (some will say a specific version of ruby dev). It seems that AWS-Vagrant is a Ruby gem, and gems are built on-box (at least it seems so - I’m a Ruby rookie), so libraries and the ruby compiler are needed. These aren’t installed on the build agent.

$ sudo apt-get install build-essential libxslt-dev libxml2-dev zlib1g-dev ruby-dev

Finding #3: The AWS-Vagrant plugin needs Ruby version 2.0+

$ vagrant plugin install vagrant-aws
Installing the 'vagrant-aws' plugin. This can take a few minutes...
/usr/lib/ruby/1.9.1/rubygems/installer.rb:388:in `ensure_required_ruby_version_met': json requires Ruby version ~> 2.0. (Gem::InstallError)
    from /usr/lib/ruby/1.9.1/rubygems/installer.rb:156:in `install'
    from /usr/lib/ruby/1.9.1/rubygems/dependency_installer.rb:297:in `block in install'
    from /usr/lib/ruby/1.9.1/rubygems/dependency_installer.rb:270:in `each'
    from /usr/lib/ruby/1.9.1/rubygems/dependency_installer.rb:270:in `each_with_index'

vagrant plugin install wants to use Ruby 1.9.1 to perform the gem build, but that version is too old. Fortunately the build agent has Ruby 2.3.

Finding #4: Ubuntu Vagrant can’t load plugins built with Ruby 2.3

$ gem install --verbose vagrant-aws
Successfully installed vagrant-aws-0.7.2
41 gems installed
$ vagrant plugin install /home.travis/.rvm/gems/ruby-2.3.1/cache/vagrant-aws-0.7.2.gem
/usr/lib/ruby/1.9.1/rubygems/format.rb:32:in `from_file_by_path': Cannot load gem at [/home.travis/.rvm/gems/ruby-2.3.1/cache/vagrant-aws-0.7.2.gem] in /home/travis/build/edwinsteele/biblebox-pi (Gem::Exception)
    from /usr/share/vagrant/plugins/commands/plugin/action/install_gem.rb:36:in `call'
    from /usr/lib/ruby/vendor_ruby/vagrant/action/warden.rb:34:in `call'
    from /usr/share/vagrant/plugins/commands/plugin/action/bundler_check.rb:20:in `call'
    from /usr/lib/ruby/vendor_ruby/vagrant/action/warden.rb:34:in `call'
    from /usr/lib/ruby/vendor_ruby/vagrant/action/builder.rb:116:in `call'

Now I’m running out of ideas and I’m considering less conventional means like building on the Travis MacOS environment where I can use Homebrew or moving to another CI hosting provider entirely. Fortunately I stumbled on the answer…

The Answer

From the Vagrant docs

Beware of system package managers! Some operating system distributions include a vagrant package in their upstream package repos. Please do not install Vagrant in this manner. Typically these packages are missing dependencies or include very outdated versions of Vagrant. If you install via your system’s package manager, it is very likely that you will experience issues. Please use the official installers on the downloads page.

Yeah, I experienced issues. Once I followed the advice it was smooth. Perhaps I should have looked at the official docs sooner!

$ wget -O /tmp/vagrant.deb https://releases.hashicorp.com/vagrant/1.8.7/vagrant_1.8.7_x86_64.deb
$ sudo dpkg -i /tmp/vagrant.deb
$ vagrant plugin install vagrant-aws

And now I have a CI pipeline running on AWS after each commit. Nice.

Site Security Improvements

I did some security work on this site recently. I was able to get some nice wins without a great investment of time, in part due to the great resources that are available. Here are the areas of work, the resources that I used, and the outcomes:

Content Security Policy (CSP)

A CSP constrains the actions that a web page can take or the actions that can be performed upon it. It allows one to apply the principle of least privilege to a page and site. A CSP allows one to specify constraints like “Only Load CSS from these sources”, “Don’t allow this site to be embedded in frames” and “Don’t allow inline JavaScript”. I developed a CSP after reading a the HTML5rocks CSP tutorial and Scott Helme’s CSP intro. I validated my policy using Google’s CSP evaluator and Mozilla’s Observatory tool. In order to apply best-practices, which include disabling inline JavaScript and CSS, I needed to make a simple changes to the site. I’ve been conscious to minimise JavaScript and CSS as I’ve developed this site, and it was great to see how that choice made the application of best-practices a simple task.

Miscellaneous security headers

I implemented X-XSS-Protection, X-Content-Type-Options and X-Frame-Options and while the effect of these headers overlaps a little with CSP, providing them is still a good idea because of inconsistent CSP implementations and benefits unrelated to CSP. I learnt about them from Scott Helme’s Response Headers page and Mozilla’s web security guidelines. I validated my setup with the SecurityHeaders validation tool and Mozilla’s Observatory tool.


I already had a reasonable SSL setup but while looking at the Mozilla web security guidelines, I didn’t consider how my list of cipher choices would need regular updating (I’d last reviewed them 2 years ago!). Mozilla are good enough to provide nginx config snippets for to help with good cipher selection, and their config snippet included an HTTP Strict Transport Security (HSTS) directive. I’d considered HSTS before, but found the SSL certificate renewal process to be complex enough that I was unsure I wouldn’t accidentally take my site offline around renewal time. Having recently switched my site certificates over to the (awesome) Let’s Encrypt renewal process, I felt comfortable activating HSTS at the same time. I validated my setup with the Qualys SSL Report


It took about 4 hours to make the changes, and after the changes were applied, this site 1 moved from an A to an A+ on the Qualys SSL Report. The Mozilla Observatory tool gives the site an A+ and the SecurityHeaders.io validator gives it an A. My nginx config is available on GitHub.

  1. Actually, I use Cloudflare as a CDN, so I ran the tests against my origin server.  

Travelling through LAX just became a lot easier

I’ve transited Los Angeles Airport for many years on my way to conferences and family visits and on a recent trip I was thrilled to find that its no longer necessary to go through airport security when moving between international to domestic flights. While the signposting is pretty bad, one can now move between Tom Bradley International and Terminal 4 (and thus the rest of the domestic terminals) without exiting the terminal and subjecting oneself, and ones family, to security screening and potential delays.


How I use the Apple Watch

I bought an Apple Watch a little while ago. I really like it and I find it useful, but I use it differently to most people and I almost didn’t buy one because I wasn’t sure that it could fit with my need. Here are a few of my “workarounds” and a few findings that someone might find useful.

A non-standard user?

  1. I don’t have an iPhone.
  2. I don’t want to charge the Apple Watch at night; I want to wear it to bed.

I’d love to report that the Apple Watch can connect to an iPad (I only have an iPad), but it doesn’t. You still need an iPhone to configure it, and to serve as a conduit for data and App installs, but you don’t need it to be with you all the time for the watch to be useful. In my case it doesn’t even matter that the watch is hooked into my wife’s Apple ID (she has the iPhone) because all the important data is still accessible1.

What I planned to use it for

  • silent alarms during the day, and to wake me up in the morning
  • the time
  • fitness tracker (mainly step count)

If you’re thinking that I could have used a Fitbit, you’d be right. I’ve had a few of them and the build quality on my bands was poor and the software gets in my way 2, so I wanted something else.

What I also use it for

  • checking the weather when dressing the kids for bed and when getting clothes out for the next day
  • countdown timers
  • driving directions when I’m out with my family

I don’t use it for anything else.

Battery Life

The big surprise is that battery life isn’t a problem, even though the standard usage pattern involves charging at night and I have it on my wrist at night.

My observations are that recording activity drains the battery a little, the attempts by the phone to connect to the iPhone drain the battery, and being in contact with the iPhone drains it quite a bit more.

Battery drain rates: being a couch potato in aeroplane mode - 2% per hour normal activity levels in aeroplane mode - 2.5% per hour normal activity levels (not in aeroplane mode) when the iPhone is not around - 3.5% per hour normal activity levels (not in aeroplane mode) when the iPhone is around - 5% per hour

I have a 42mm watch. Bluetooth is off, brightness is at the lowest setting and haptics are at the strongest setting). I have one third party complication Pocket Weather AU.


I’ve observed the watch charging at about 1% per minute, just like David Smith, whose great idea allowed me to even consider an Apple Watch.

I don’t really need the watch while I’m preparing for my day, so I put it on charge within a few minutes of getting up so it’s charging while I do my morning routine (shower, coffee, eat, read) so it has about 75 minutes to charge.

I have the watch out of aeroplane mode in the evening (~4 hours) i.e. 4 x 3% = 12% and the rest of the non-charge time in aeroplane mode (~19 hours) i.e. 19 x 2.5% = 47.5%. This means I need about an hour of charge time to be constantly topped off. That has always been enough and consequently I’m not worried about battery life.

So, it works for me

Initially, I thought the watch would be of limited use without connectivity but it’s not been a problem. Most of the features that require connectivity are for notifications and I’m happy to forgo them - I like not being interrupted, to be honest, and where connectivity is necessary for data updates, I’ve found a once-a-day refresh to be sufficient. So, I’m really happy with the watch. It’s been a step-up from the Fitbit in terms of functionality and quality so it’s been a good purchase and I’m looking forward to what’s coming in watchOS 3.

  1. iCloud shared calendars for the win 

  2. Why can’t I set an alarm when I’m not connected to the internet? Why can’t graphs render without an Internet connection? 

Making Ansible, Doas and OpenBSD play nicely

A little while ago, OpenBSD replaced sudo in the base system with doas. This transition follows a typical path in the OpenBSD community where a large, difficult to audit daemon or key application is replaced by an auditable and securable alternate implementation. At the same time, features are pruned if they add significant complexity. I don’t need all the features of sudo and I like the simplicity and security of doas so I wanted to switch. I use Ansible to provision and manage my small collection of servers and when version 2.0 added support for privilege escalation via doas (via the become keyword) I started to make my switch. The main use-case in my Ansible playbooks is where I run as root and become my non-admin user so that permissions are set correctly by default, and in this case doas behaves differently…

# /usr/bin/env | grep -E '^(HOME|USER)='
# sudo -u esteele /usr/bin/env | grep -E '^(HOME|USER)='
# doas -u esteele /usr/bin/env | grep -E '^(HOME|USER)=' 

The key difference is which environment variables are propagated from the original session. I find doas to be a little unintuitive in this respect, particularly the inability in the new session to write to the directory specified by $HOME and this behaviour causes problems with the Ansible git and the pip tasks. The git task expects $USER to correspond to the effective user id and the pip task expects $HOME to be the home directory of the effective user id. The unexpected doas environment meant the tasks were trying to write files, as my non-admin user, under /root which was unsurprisingly unsuccessful. Fortunately Ansible has a way to specify environment variables for a task and even though it’d be preferable for the become mechanism to take care of this, it’s an effective workaround (albeit a slightly dirty one).

The example below, from my web host playbook, ties the environment specification and the doas become in a block

# doas preserves USER and HOME, which is different to sudo, however by                                                                                                    
#  specifying them as environment variables we can proceed.                                                                                                               
# USER is required for git, and HOME is required for pip                                                                                                                  
- block:                                                                                                                                                                  
  - name: Checkout wordspeak repo                                                                                                                                         
    git: repo=ssh://git@github.com/edwinsteele/wordspeak.org.git                                                                                                          

  - name: Setup nikola virtualenv                                                                                                                                         
    pip: requirements=/home/esteele/Code/wordspeak.org/requirements.txt                                                                                                   

  become: True                                                                                                                                                            
  become_method: doas                                                                                                                                                     
  become_user: esteele                                                                                                                                                    
    USER: esteele                                                                                                                                                         
    HOME: /home/esteele           

And having tied this together, I can remove the sudo package!

    - name: Remove sudo
      openbsd_pkg: name=sudo-- state=absent
Site by Edwin Steele | Uncopyright. No rights reserved | Why give it away?