Heterogeneous Hardware Management and Benchmarking via Docker

We do a lot of Benchmarking and functionality testing here at Bitfusion.io to continuously evaluate our application acceleration technology across different CPU architectures, storage and memory configurations, and accelerators such as GPUs and FPGAs. As a startup, we have various high performance systems in-house, but there are always newer and more exciting hardware configurations for us to test and develop on. To compound the problem further, more often than not different cloud and hardware provider offer access to different types of hardware.

For example, Amazon has a nice selection of systems when it comes to Intel Sandybridge and Haswell processors, but their GPU selection of K520 GRID Nvidia GPUs in the g2.2xlarge and g2.8xlarge instances is rather dated by today’s standards. To get access to machines with multiple Nvidia state of the art K80 GPUs we had the need to quickly deploy our software on other cloud providers such as Softlayer where K80s are available. To test on AMD CPUs or to access some of our FPGAs we needed an easy way to do the same on our bare-metal servers in Rackspace. And finally, as mentioned before, we have several local systems which also needed to be part of the test infrastructure – I think you get the idea.

To solve the problem we needed three things: First, an ability to provision systems quickly on demand and bring them back down – a startup is always on the clock and on a budget. Second, we needed an easy way to deploy our benchmarks and testing across any infrastructure transparently and reliably. Third, we needed a way to get performance and monitoring data for our runs and to quickly visualize them – the tricky part here was we were shooting for high granularity data collection without interfering with the actual performance itself.

To solve these problems we adopted the following high-level architecture. To manage the systems inside our office and across the different cloud providers we utilized SaltStack exclusively. SaltStack manages all our machine images, created with Packer, for the various systems and cloud providers. While in general the images are fairly similar, different hardware occasionally needs specific run-time environment and drivers on the host systems, especially when it comes to accelerators. Using SaltStack we then create and manage dynamic pools of machines, where machines can be added or removed at any time from our pools.

All our builds, applications, and benchmarks live within Docker containers, along with various performance monitoring software. We can monitor performance from inside containers as well as from outside containers depending what it is exactly that we are looking for – some tools we utilize here include collectl and sysdig. Any streaming performance data we pipe out to Elasticsearch, taking care to cluster the streaming data periodically such that we only minimally disturb any resource measurements or I/O activity. Once the data is in Elasticsearch the visualization is just a simple exercise in data querying and java scripting. Below is one possible visualizations example of this data from our free Profiler tool which we discuss in more detail toward the end of this post:


Since we have a use-case which differs quite considerably from the usual Docker case of many small web sites or microservices, we have many more Docker container images and they are much larger than usual. In order to manage these we run our own instance of Docker Trusted Registry so that we keep the data traffic local when pushing or pulling container images to and from the registry. As a result of our obsession with performance, we run overlayfs as the Docker graph backend. The overlayfs driver is the component of Docker which manages all the layers of the files in the container images. Whilst there are several options here, overlayfs is the preferred, and fastest, choice going forward. It does have some caveats, we had to move to Linux Kernel 4.2+ and Docker 1.8.1 in order to achieve a stable configuration, and the underlying file system on which overlayfs is layered requires a configuration with a very large number of “inodes”, the data element which records where files are stored on the disk. This means the underlying file system must be specially formatted with a much larger number of inodes than normal. This can be done using a command such as

# mke2fs -N <desired inodes> -t ext4 ….

One major advantage of using Saltstack to manage our docker containers is that the Docker event stream. Events generated whenever a docker container is started, stopped or otherwise changes state can be trivially exposed on the Saltstack event-bus, where Saltstack Reactors can listen and respond to Docker events by triggering Saltstack actions, or sending messages to other systems.

All Docker containers across our pools are managed via an asynchronous event API which we wrote in-house. We wrote this API in house as it allows us to do run-time profiling of entire tool chain flows in diverse fields such as genomics, semiconductors, and machine learning just to name a few. Using our proprietary Boost technology we then automatically optimize these tool pipelines for maximum performance across heterogeneous compute clusters. The high-level diagram below illustrates one possible use case where thin container clients leverage heterogeneous containers to obtain compute acceleration:


One quick note regarding accelerators and device pass through, as of SaltStack version 2015.8.5 Docker device pass-through is not natively supported, this can be solved with a small patch to the SaltStack docker driver which we wrote internally and which can be found here:


For day to day use, all the above is triggered from our Jenkins build servers. However, for those interested in taking a lightweight version of our profiling architecture for a spin, we put a web front-end on this flow which can be found at http://profiler.bitfusionlabs.com – here you can build any type of Docker based application benchmark directly from a browser exposed terminal, and can get quick performance results from various instance types across multiple cloud providers. Feel free to share the benchmarks with your friends and colleagues, or make them public so that others can benefit from them as well.

This article is part of our Docker and the Future of Configuration Management blog roundup running this November.  If you have an opinion or experience on the topic you can contribute as well

Leave a comment

Filed under DevOps

Immutable Delivery

This article proposes a design pattern modeled after “Immutable Infrastructure”, something I call “Immutable Delivery”.  There has been a lot of debate and discussion about the usage of the term “Immutable” lately. Let me clearly say that there is no such thing as an immutable server or immutable infrastructure. I know of no example in my 35 years of working with IT infrastructure of a system or infrastructure that is  completely immutable. A system changes and continues to change the minute it is powered on. Things like logs, dynamic tables  and memory are constantly changing during a system’s lifecycle.

However, I think it’s okay to use the phrases “Immutable Infrastructure” or “Immutable Delivery” in the context of an system or infrastructure delivery pattern. In fact, I propose we think of it as a metaphor for a kind of full stack stateless application delivery pattern.  I’ve had mathematicians actually yell at me after a presentation on Infrastructure as Code and my use of the term idempotency.  When confronted in this sort of battle, I would always retreat with saying “It’s a term used to describe a certain type of operation we do in configuration management”. Henceforth; I suggest the same idea for the use of the phrases “Immutable Infrastructure” and “Immutable Delivery”.

First Things First

Let’s describe what an “Immutable Infrastructure” model might look like. Basically, it is a model where the complete infrastructure is delivered intact, for example as a  set of immutable virtual machines or as a set of immutable Linux containers. The idea is, by design, to never touch or change the running system. In most cases, the running system is the production system; but in some recent examples with containers, this model is also used in integration and testing structures. I like to say no CRUD for applications, middleware configuration files and operating systems. The model is that when something needs to be changed in the infrastructure,  it is done as a new deploy from the most recent versioned artifact (i.e., the complete system). If there needs to be a rollback, the same process is still a redeploy, except when in this case, the artifact is the older version. One caveat in this model is relational databases. It is very difficult, maybe impossible, to have immutable relational databases. However, some companies I have talked to sort of do what I call an “No RUD” for databases. In that, they create new records but do not replace, update or delete existing ones. Mileage always varies for all of this.

Order Matters

Back in 2010, I was given an interesting paper written by Steve Traugott called “Order Matters:  Turing Equivalence in Automated Systems Administration” (2002).  Taugott’s article described systems that were either divergent, convergent or congruent.  At the time, I was working at Chef and desired state configuration and building systems through convergent models was what I was currently evangelizing. At that time, I felt that, the “Order Matters” paper described the differentiation between how Chef and Puppet worked.  The short version here is that Chef used a prescriptive Ruby-based DSL that was executed on the local host in an order specific manner based on how you wrote the DSL code.  Whereas Puppet used a clever dependency graph on the Puppet server to determine some of the ordering at deployment time. In some cases, this made a difference for certain types of organizations (Note 1).  Traugott’s paper does an excellent job describing a thesis on why this “could” be so. What fascinated me was Traugott’s explanation of congruent systems.  At that time, there didn’t seem to be a commodified way to deliver this form of infrastructure, at least from my perspective.  

Building With Legos

A year later Netflix wrote an interesting blog post  called “Building With Legos”.  At the time, there was a lot of outrage (myself included) regarding this post. At first glance, it looked like Netflix was suggesting that they were advocating a model of “Golden Image” delivery infrastructure. Part of my initial reservation was that I had seen this pattern twice in my career with dire consequences. The first time was back in the late 90’s where companies would use a product called “Ghost” to build Windows servers. This was another one of those ideas where it sounded like a good idea at the time until you had, in some cases, thousands of poorly cataloged images and wrong image deploys caused major outages. Fast forward to around 2008 and here again, organizations were starting to make the same mistakes all over again with “cloud” images,  specifically in the form of Amazon AMI’s. I believe that sometime around 2009 the phrase “Image Sprawl” became a popular phrase for doing “cloud” wrong.  In fact, I remember a horrifying story in the early days of cloud, where the Public Broadcast Service (PBS)  accidentally made a proprietary AMI image public and it took them a couple of days to clean up all the viral versions of their private keys.  So at first glance of the Netflix blog post,  you could see how many were thinking they were suggesting a mode of bad configuration management.  However, on a closer read they were much smarter than this.  What they were indeed saying in the blog post was that they were treating the AMI’s like JAR or WAR files in that the AMI images were holistic artifacts built through a continuous integration/continuous delivery (CI/CD) process.  The AMI’s would be pulled at deploy time to be launched into production similar to the way a JAR or WAR would be pulled from an artifact repository.  Only in this example, it  included all the infrastructure (OS, Middleware, application and most of the application configuration files).  I like to use the phrase “Java Lied’ in many of my presentations.  They told us “Write once, run anywhere”.  What they forgot to say is this is true unless you have an incompatible runtime environment.  Netflix at the time of the blog post didn’t refer to that process as “Immutable Infrastructure” and of course it was not completely immutable.  They had to use a set of open source tools to discover the services and converge the deployed services, so hence, their systems were not immutable.  However, their service delivery pattern was indeed a model of an immutable delivery pattern. Some time later, Netflix did start to refer to their architecture as “Immutable Infrastructure”.   

Trash Your Servers and Burn Your Code: Immutable Infrastructure

June 2013 was the first time I had heard the phrase “Immutable Infrastructure”.  Chad Fowler wrote a blog post “Trash Your Servers and Burn Your Code: Immutable Infrastructure and Disposable Components”. In the blog, Fowler proposes his idea born from functional programming techniques that offer immutable data structures. The belief was that if somehow we could deliver the complete infrastructure, for example a server, with all the infrastructure needed  for the application, then in his words, “ it would be easier to reason about and harder to screw up”.  Imagine delivering a server the same way we deliver an application, for example, like a WAR file. Fowler’s main point was that systems grow warts, especially when you are fire fighting.  For example, a system might be well defined through an configuration management  tool, but during an outage, changes may be made on the system directly and then never put back into the configuration management recipe or manifest later.  The list of possible human or system entropy examples goes on.  Fowler also points out that sometimes application code is deployed outside of “normal straight-from-source-control process.” Operating system patching or source repositories sometimes change in flight between testing and production. There are many other examples of possible mismatches in the flow you can find in the wild. All of this could be put in a bucket called “bad configuration management hygiene”; however, just like I have never seen a perfect immutable system, I have also never seen in my 35 years a “perfect system”.  I mean ”system” the way Dr. Deming would describe it, in that, all systems include humans.  

Docker and the Three Ways of Devops

When I first arrived at Docker back in February 2015,  I reviewed a Gartner paper called “Become More Agile and Get Ready for DevOps by Using Docker in Your Continuous Integration Environments” and it set me down a course of thinking. The author, Sean Kenefick, had a strong background in release engineering and wrote an excellent paper of how Gartner would suggest using Docker.  As I read the paper,  the first thing it reminded me of was Steve Traugott’s paper about order and why it mattered and the value of congruent systems.  I decided to write a blog post called “Docker and the Three Ways of Devops”.  During my research, I talked to a number of Docker users that were doing what Fowler described as Immutable Deployments, using Docker images as the immutable artifacts. This process was similar to what Netflix was doing with two major differences. One, the services were container images not virtual machines, and two, they were being delivered immutably from the developer’s perspective. After the container images were compiled and tested the “service” would be pushed to the CI process for service level integration testing. Most of the organizations using this model had already crossed over to a microservices architecture. The flow would go something like this:

  • The developer would test their service as a container, typically on a virtual machine running on their laptop.    
  • They would also load the other primary services in their service-oriented architecture, possibly owned by other teams, into their same virtual machine on their laptop.  
  • They would continue to compile, load and test their service sometimes on their laptop and other times through a sort of first-pass CI server.  
  • When testing was complete, they would typically check in their service as a container (binary) image with a meta file to describe the rest of the pipeline flow (CI/CD).  

All of the people I talked to agreed that the benefit of this process was that the developer was in control not only of the application code they wrote but also of any middleware and basic operating systems behavior (note 2 & 3). The benefits of an immutable delivery process, like the ones described by Netflix with their AMI flow, are that you can increase speed and decrease resources and possible variation with containers. Containers instantiate in around 500 milliseconds whereas virtual machines instantiate well over a minute. In a microservices architecture, many of the containers are around 100 megabytes whereas virtual machines could still be as large as 2 gigabytes. I like to say that containers are the killer app for microservices. With this mode the developer can test all of the other dependent services from their laptop.  Werner Vogels the CTO of Amazon is often quoted as saying, “You build it, you run it”. In Devops we like to say, “developers should wear pagers”. There is a reason why developers like Docker so much. When they build it, own it and get paged in the middle of the night, they know that for the most part the bits that they tested are the same (i.e., congruent ) bits that are running in production.  

In 2014 at Dockercon US, Michael Bryzek, of Gilt, gave a fantastic presentation “Immutable Infrastructure with Docker and EC2”.  In this presentation, he describes a process where the developers check in a set of binary container images with one single meta file.  I have personally transcribed what he says starting at 28:03 in his presentation:

“This is how we run our infrastructure. One of the things that developers have to do is provide the commands to start the Docker container, and that’s it. This is kind of amazing right?  Any EC2 instance that we spin up now, we don’t care if you’re running Node, Ruby, Scala, Java or if you made up your own programming language. It’s absolutely amazing how nice this is.  When we compare this to the way we did this in the past, we had one repository that had all of the different scripts to know how to build all of the different applications at Gilt. We have 1000 Git repos and over 300 different applications. We are 7 years old which means we have like 7 different ways of producing applications. There’s 25 different ways that we build software at Gilt and it’s all captured in a central repo.  That was at conflict with where we are going in terms of teams really owning their software and being able to deploy their own services.”

I have talked to a number of companies over the past year and many of them are moving to an “Immutable Delivery” process driven by microservices implemented by containers. Capital One at the Devops Enterprise Summit in October 2015 (DOES15)  gave a presentation called “Banking on Innovation & DevOps”. They said in that presentation that they are using Docker in production. They have also said that they are delivering software in an immutable delivery pattern. This model is not just for web scale anymore.  

In the end, “Immutable Infrastructure” or what I have coined as “Immutable Delivery”  is just a model with many variants. No large organization uses a single model to manage their infrastructure. Over the next few years, I look forward to working with all sorts of products, old and new, to find the correct balance of service delivery. My only goal is to be an evangelist of a model that Josh Corman, CTO at Sonatype, and I describe as “Immutable Awesomeness”.  This is  a presentation we did at DOES15.  We borrowed many of our ideas from the book “Toyota Supply Chain Management: A Strategic Approach to Toyota’s Renowned System”.  In this book, they describe the 4 V’s.  Increase Variety, Velocity, and Visibility and decrease Variability.  In short whatever works, works…

John Willis
Director of Ecosystem Development, Docker Inc.

This article is part of our Docker and the Future of Configuration Management blog roundup running this November.  If you have an opinion or experience on the topic you can contribute as well


  1. To be clear, Puppet today allows for both models and this particular differentiation, in my opinion, no longer exists. In fact, both products today have relative parity with regards to ordering.
  2. For the nit-pickers, mileage varies on operating system immutability. Containers run on a host operating system and share the kernel.  Bad hygiene on the host will definitely cause “immutable” woes.  
  3. This is, by the way, a great area for co-existence between Infrastructure as Code products like Chef and Puppet and containerization products like Docker


Why Order Matters: Turing Equivalence in Automated Systems Administration


Building with Legos


VM Image Sprawl in Real Life


Trash Your Servers and Burn Your Code: Immutable Infrastructure and Disposable Components


Become More Agile and Get Ready for DevOps by Using Docker in Your Continuous Integration Environments


Docker and the Three Ways of Devops


A conversation with Werner Vogels


Immutable Infrastructure with Docker and EC2”


Banking on Innovation & DevOps


Toyota Supply Chain Management: A Strategic Approach to Toyota’s Renowned System


Immutable Awesomeness



1 Comment

Filed under DevOps

Container Automation: Building a Brickyard

My name is Nathaniel Eliot, and I’ve worked extensively in software deployment over the last several years. I have worked on two automation frameworks around Chef: Ironfan, an open-source cluster management system from Infochimps, and Elzar, a (sadly closed-source) blue-green framework based on Spiceweasel. I currently work at Bazaarvoice, where I’m building out a Flynn.io installation.

Barnyard Blues

There is a catch-phrase in DevOps: “cattle, not pets”. It’s intended to describe the step forward that configuration management (CM, e.g. Chef, Puppet, Ansible, Salt, etc.) tools provide. Instead of building and maintaining systems by hand, DevOps-savvy engineers aim to build them via automated, repeatable systems. This has revolutionized system deployment, and resulted in vast improvements in the stability and manageability of large, complicated systems.

But while cattle are an important part of civilizing your software, they have drawbacks. As anybody who’s worked a farm will tell you, cattle management is hard work. Long-lived systems (which most barnyard-style deployments still are) decay with age, as their surrounding environment changes, and as faults are triggered in software components; upgrades to fix these issues can be complicated and fragile. New hosts for these systems can also suffer from unintended evolution, as external resources referenced by the CM build process change. System builds are often lengthy affairs, and often heavily intertwined, such that singular failures can block updates on unrelated resources.

These issues mean that failures are often addressed by “reaching into the cow”: SSH logins to affected hosts. As the phrasing implies, this should be considered a little gross. Your team’s collective understanding of a system is based on it being build in predictable ways from visible source code: an SSH login undermines that understanding.

Building a Brickyard

The phrase I like for container automation (CA, e.g. Flynn, Mesos+Docker, etc.) is “brickyard, not barnyard”. Bricks are more uniform, quicker to make, and easier to transport than cows: CA provides greater immutability of product, faster cycle time, and easier migration than CM.

Because everything is baked to the base image, the danger of environmental changes altering or breaking your existing architecture is far lower. Instead, those changes break things during the build step, which is decoupled from the deployment itself. If you expand this immutability by providing architecturally identical container hosts, your code is also less vulnerable to “works in dev” issues, where special development configuration is lacking on production machines.

Rapid cycle time is the next great advantage that CA provides, and arguably the largest from a business perspective. By simplifying and automating build and deployment processes, CA encourages developers to commit and test regularly. This improves both development velocity and MTTR (mean time to repair), by providing safe and simple ways to test, deploy, and roll back changes. Ultimately, a brick is less work to produce than a fully functioning cow.

Because CA produces immutable results, those results can easily be transported. The underlying CA tools must be installed in the new environment, but the resulting platform looks the same to the images started on it. This gives you a flexibility in migration and deployment that may be harder to achieve in the CM world.

These benefits are theoretically achievable with configuration management; Ironfan is a good example of many of these principles at work in the CM world. However, they aren’t first class goals of the underlying tools, and so systems that achieve them do so by amalgamating a larger collection of more generic tools. Each of those tools makes choices based on the more generic set of situations it’s in, and the net result is a lot of integration pain and fragility.

Bricks or Burgers

So when should you use CM, and when should you use CA? You can’t eat bricks, and you can’t make skyscrapers from beef; obviously there are trade-offs.

Configuration management works best at smoothing the gaps between the manually deployed world that most of our software was designed in, and the fully automated world we’re inching toward. It can automate pretty much any installation you can do from a command line, handling the wide array of configuration options and install requirements that various legacy software packages expect.

Container automation currently works best for microservices: 12-factor applications that you own the code for. In existing architectures, those often live either in overly spacious (and fallable) single servers, or in messy shared systems that become managerial black-holes. This makes them an easy first target, providing greater stability, management, and isolation than their existing setups.

However, that’s as things stand currently. Civilization may depend on both, and the cattle came first, but ultimately it’s easier to build with bricks. As frameworks like Flynn expand their features (adding volume management, deploy pipelines, etc), and as their users build experience with more ambitious uses, I believe CM is slowly going to be trumped (or absorbed) by the better CA frameworks out there.

This article is part of our Docker and the Future of Configuration Management blog roundup running this November.  If you have an opinion or experience on the topic you can contribute as well

Leave a comment

Filed under DevOps

Templating Config Files In Docker Containers

Configuration management does many things, and it does most of those quite poorly. I could write a long essay on how by solving software problems with more software we’re simply creating more software problems, however I will attempt to resist that urge and instead focus on how Docker and the Docker ecosystem can help make configuration management less sucky.

Ignoring volume mounts (which are an abomination for which I hold @gabrtv wholly responsible for), Docker has two main ways to configure your application: firstly by creating the dockerfile in which you explicitly declare your dependencies and insert any configuration files, and secondly at run time where you pass commands and environment variables to be used inside the container to start your application.

We’re going to ignore the dockerfile here and assume that you have at least a passing familiarity with them;  instead we’re going to focus on how to configure your application at run time.

A true Docker Native app would have a very small config file of which some or all settings could be overridden by environment variables or CLI options that can be set at run time to modify the appropriate configuration option (say, pointing it at a MySQL server at ).

Very few applications are written in this way, and unless you’re starting from scratch or are willing to heavily re-factor your existing applications you’ll find that building and configuring your applications to run in “the docker way” is not always an easy or particularly pleasant thing to have to do. Thankfully there are ways to fake it.

To save writing out a bunch of CLI arguments the cleanest ( in my opinion ) way to pass values into docker containers is via environment variables like so:

 $ docker run -ti --rm -e hello=world busybox env

We can then write an application to read that environment variable and use it as a configuration directive like so:

 echo hello $hello

No prizes for guessing our output, we run a docker container from the image containing this script:

 $ docker run -ti --rm -e hello=world helloworld
 hello world

Now this was a pretty asinine demo, and apart from showing how passing environment variables into a docker container works doesn’t really do anything useful.  Let’s look at a slightly more realistic application.  Take a python app that reads a configuration file and when asked renders a web page using the contents of that configuration file:

note:  these examples are abbreviated sections of the example factorish app.


 import ConfigParser
 import os

 from flask import Flask
 app = Flask(__name__)

 def hello():
 Config = ConfigParser.ConfigParser()
 return 'Luke, I am your {}'.format(
 Config.get("example", "text"))

 if __name__ == '__main__':
 app.run(host='', port=80)


 text: father

Now when we run this application we get the following:

$ docker run -d -p 8080:8080 -e text=mother example
$ curl localhost:8080
 Luke, I am your father

Obviously the application reads from the config file and thus passing in the environment variable `text` is meaningless. We need a way to take that environment variable and embed it in the config file before running the actual `example.py` application.  Chances are the first thing that popped into your head would be to use `sed` or a similar linux tool to rewrite the config file like so:


 sed -i "s/^text:.*$/text: ${text}" example.conf
 exec gunicorn -b app:app

Now we can run it again with run.sh set as the starting command and the config should be rewritten.

$ docker run -d -p 8080:8080 -e text=mother example ./run.sh
$ curl localhost:8080
Luke, I am your mother

This might be fine for a really simple application like this example, however for a complicated app with many configuration options it becomes quite cumbersome and offers plenty of opportunity for human error to slip in. Fortunately there are now several good tools written specifically for templating files in the docker ecosystem,  my favourite being confd by Kelsey Hightower which is a slick tool written in golang that can take key-pairs from various sources ( the simplest being environment variables ) and render templates with them.

Using confd we would write out a template file using the `getv` directive which simply retrieves the value of a key.  You’ll notice that the key itself is lowercase,  this is because confd also supports retrieving key-pairs from tools such as etcd and confd which use this format.  When set to use environment variables it is translated into reading the variable “SERVICES_EXAMPLE_TEXT”.

 text: {{ getenv "/services/example/text" }}

We would accompany this with a metadata file that tells confd how to handle that template:

 src   = "example.conf"
 dest  = "/app/example/example.conf"
 owner = "app"
 group = "app"
 mode  = "0644"
 keys = [
 check_cmd = "/app/bin/check {{ .src }}"
 reload_cmd = "service restart example"

The last piece of this puzzle is a executable command in the form of a shell script that docker will run which will call confd to render the template and then start the python application:


 # read 'text' env var and export it as confd expected value
 # set it to 'father' if it does not exist
 # run confd to render out the config
 confd -onetime -backend env
 # run app
 exec gunicorn -b app:app

Now let’s run it, first without any environment variables:

 $ docker run -d -p 8080:8080 --name example factorish/example
 $ curl localhost:8080
 Luke, I am your father
 $ docker exec example cat /app/example/example.conf
 text: father

As you can see the server is responding using the default value of `father` that we set in the export command above.   Let’s run it again but set the variable in the docker run command:

 $ docker run -d -e SERVICES_EXAMPLE_TEXT=mother -p 8080:8080 --name example factorish/example
 $ curl localhost:8080
 Luke, I am your mother
 $ docker exec example cat /app/example/example.conf
 text: mother

We see that because we set the environment variable it is be available to `confd` which renders it out into the config file.

Now if you go and look at the full example app you’ll see there a bunch of extra stuff going on.   Let’s see some more advanced usage of confd by starting a coreos cluster running etcd in Vagrant. etcd is a distributed key-value store that can be used to externalize application configuration and retrieve it as a service.

 $ git clone https://github.com/factorish/factorish.git
 $ cd factorish
 $ vagrant up

This will take a few minutes as the servers come online and build/run the application.   Once they’re up we can log into one and play with our application:

 $ vagrant ssh core-01
 core@core-01 ~ $  docker ps
 CONTAINER ID        IMAGE               COMMAND             CREATED             STATUS              PORTS                    NAMES
 ee80f89d2565        registry            "docker-registry"   25 seconds ago      Up 25 seconds>5000/tcp   factorish-registry
 c763ed34b182        factorish/example   "/app/bin/boot"     52 seconds ago      Up 51 seconds>8080/tcp   factorish-example
 core@core-01 ~ $ docker logs factorish-example
 ==> ETCD_HOST set.  starting example etcd support.
 2015-11-08T21:35:16Z c763ed34b182 confd[23]: INFO Target config /app/example/example.conf out of sync
 2015-11-08T21:35:16Z c763ed34b182 confd[23]: INFO Target config /etc/service/confd/run out of sync
 2015-11-08T21:35:16Z c763ed34b182 confd[23]: INFO Target config /etc/service/confd/run has been updated
 2015-11-08T21:35:16Z c763ed34b182 confd[23]: INFO Target config /etc/service/example/run out of sync
 2015-11-08T21:35:16Z c763ed34b182 confd[23]: INFO Target config /etc/service/example/run has been updated
 2015-11-08T21:35:16Z c763ed34b182 confd[23]: INFO Target config /etc/service/healthcheck/run out of sync
 2015-11-08T21:35:16Z c763ed34b182 confd[23]: INFO Target config /etc/service/healthcheck/run has been updated
 echo ==> example: waiting for confd to write initial templates...
 2015-11-08T21:35:16Z c763ed34b182 confd[23]: ERROR exit status 1
 Starting example
 *** Booting runit daemon...
 *** Runit started as PID 51
 2015-11-08 21:35:22 [56] [INFO] Starting gunicorn 0.17.2
 2015-11-08 21:35:22 [56] [INFO] Listening at: (56)
 2015-11-08 21:35:22 [56] [INFO] Using worker: sync
 2015-11-08 21:35:22 [67] [INFO] Booting worker with pid: 67
 core@core-01 ~ $ curl localhost:8080
 Luke, I am your father

You can see here that we’ve started the example app,  but notice at the top where it says “starting example etcd support”.  This is because we’ve actually started it with some environment variables that makes it aware that etcd exists. It uses these to configure `confd` to run in the background and watch an etcd key for templated config value.

We can see this and modify the config setting using etcd commands:

 core@core-01 ~ $ etcdctl get /services/example/text
 core@core-01 ~ $ etcdctl set /services/example/text mother
 core@core-01 ~ $ curl localhost:8080
 Luke, I am your mother
 core@core-01 ~ $ exit
 $ vagrant ssh core-02
 core@core-02 ~ $ curl localhost:8080
 Luke, I am your mother

With confd aware of etcd it is able to notice values being changed and react accordingly,  in this case it rewrites the templated config file and then restarts the example application.   If you look at the template’s metadata from earlier you’ll see it is instructed to watch a certain key and rewrite the template if it changes.  It also has two directives `check_cmd` which is used to ensure the created template would be syntactically correct and `reload_cmd` which it runs any time the template is successfully written,  in this case to reload our application.

You’ll also notice that we were able to connect to the other coreos nodes each of which was also running the example application and because etcd was clustered across the three nodes all three applications registered the changed and updated themselves.

So now, not only do we have good clean templating in our container, we also even have the ability to change some of those config settings on the fly by connecting it to etcd.

From this very simple building block we are only a short hop away from being able to automatically configure complicated stacks that react to changes in the infrastructure instantaneously.

Pretty cool huh?

This article is part of our Docker and the Future of Configuration Management blog roundup running this November.  If you have an opinion or experience on the topic you can contribute as well

Leave a comment

Filed under DevOps

DevOps Enterprise Summit Videos Are Up

There’s a crop of great talks from this event, check them out here. If you look really hard you can see my talk too!

Leave a comment

Filed under Conferences, DevOps

Containers, Configuration Management, and The Right Tool for the Right Job

Docker brings an incredibly appealing method of managing applications to the table, but also requires us to rebuild a lot of systems that aren’t broken. In this post I’m going to look at the pros and cons of Docker and its accompanying ecosystem, and take a look at how one might start to leverage the best parts of Docker without rewriting everything.

What is it about Docker that is so exciting? It was a moonshot experiment that struck home. Rather than providing an incremental improvement over existing infrastructure patterns, Docker takes several leaps forward by providing a fresh way of managing our applications and making them accessible to developers at the same time.

Part of Docker’s success relies on providing highly opinionated solutions to the problems that come with containerizing components in your system. While these solutions are invaluable in terms of accessibility and gaining adoption, they are neither the only solutions nor necessarily the best ones in every case. Even if you are sure you want to subscribe to the opinionated “Docker way” or think it’s a worthwhile trade-off, you will still be accepting a new set of problems in exchange for the old ones, but the new set doesn’t come with the benefit of a decade or so of tools and experience to leverage.

In this post I’m going to discuss what makes Docker attractive to me and what I’m not such a fan of. Then I’m going to explore a hybrid approach that seeks to take advantage of the best parts of Docker without locking me in to the not-so-great parts.

P.S. I was hoping to put together a working demo of the architecture I describe below, but the proposed integration is still not possible… so that’s not going to happen. I’ve sat on this post for a while hoping things would change, but they haven’t, and I’ve decided instead to put this out there as is, as a theoretical architecture.

The Good

Some aspects of docker are home runs. In a time where microservices rule and developing software for the cloud is an obvious choice, how can you pass up a tool that makes managing a gazillion apps as cheap and easy as managing your monolith? And it’s a DevOps game changer: In the same way that AWS removed the friction between dev and ops for provisioning a VM, Docker removes the friction of configuring an app’s dependencies and installation. What’s more, local dev of even dozens of applications can be kept lean, and we’re closer than ever to feeling confident in “it works on my laptop.”

In summary:

  • Dense clouds
  • Transactional installs
  • Bundled dependencies
  • Tools for packaging deterministic and repeatable deploys (and rollbacks)
  • Developer workflow is simple and production-like

The Bad

A lot of the design decisions of Docker involve trade-offs, and networking is no exception.

For local dev, where managing many VMs is especially difficult and high availability isn’t important, Docker’s unique method of standing up a virtual bridge interface and allocating new IPs as containers are started is really convenient. But it is less than complete when you start worrying about high availability and exposing your services to external systems. Layering in additional networking or mapping ports to the bridge interface starts to solve this problem, but it also leaves you with a jumble.

Service discovery largely abstracts away this jumble, but at this point we’ve gone through at least three networking transformations to effectively address our services and we haven’t even started to integrate with non-Docker managed services.

Don’t get me wrong, I definitely think service discovery is the way to go. I just think that since Docker has coupled networking so tightly with its implementation, it should have made it more flexible and done more to make inter-host communication work the same as intra-host communication. Additionally, better hooks to integrate with existing software-defined networks would make all the integration work feel less like a yak shave.

Isolation security is also a concern, but it is easier to shrug off because it should be coming soon. For the time being, however, there is a lack of user namespaces in Docker containers, so UID 0 (root) in a container is also UID 0 (root) on the host machine and has all the access that comes with that.

Another concerning thing about Docker is the Docker hub. Although you don’t have to use this service or the images housed there in production, it’s presented in such a way that many people still do. Even with the addition of download signature checks, we are still left with an index of images that aren’t particularly well vetted or necessarily kept up to date. Many are built on top of full OSes that expose a larger attack surface than is necessary, and there is no established technique to ensure the base OS is up to date on its patches. There are great resources for building thin base OSes and ways to make sure they are kept up to date, but this is still more management left unaddressed.

In summary:

  • User namespace security
  • Docker hub security
  • Networking

The Ugly

One of the first things you realize with Docker is that you have to rethink everything. One is drawn in by the prospect of encapsulating their apps in a clean abstraction, but after fooling around with Supervisord for a while, most people start down the slippery slope of rewriting their entire infrastructure to keep their implementation clean.

This is because Docker isn’t very composable when taken as a whole. If you want to talk to something in a docker container, you need an ambassador. If something in a docker container needs to talk to something not managed in docker, you need an ambassador. Sometimes, you even need an ambassador for two apps both running in containers. This isn’t the end of the world, but it’s more work and is a symptom of how the docker containers are really only composable with other Docker containers, not with our systems as a whole.

What this means is that to leverage the parts of docker we want (the transactional installs, bundled dependencies, and simplified local dev), we have to rewrite and rewire a whole lot of other stuff that wasn’t broken or giving us trouble. Even if it was, you’re still forced to tackle it all at once.

If we were being honest about the shipping container analogy, we’d end up with a container ship not just carrying containers but built with containers as well!

A lot of these problems come from the same thing that makes Docker so easy to use: the bundling (and resulting tight coupling) of all components needed to build, manage, and schedule containers.

This becomes especially clear when trying to place Docker on Gabriel Monroy’s Strata of the Container Ecosystem. Although he places Docker in layer 4 (the container engine), aspects of it leak into almost every layer. It’s this sprawl that makes Docker less composable and is why integrating with it often entails a huge amount of work.


  • Incompatibile with config management
  • Not composable with existing infrastructure patterns and tools

If not Docker, Then What?!

I’m not saying we should forget about docker and go back to the tried-and-true way of doing things. I’m not even saying to wait for the area to mature and produce better abstractions. I’m simply suggesting you do what you’ve always done: Choose the right tool for the right job, pragmatically apply the parts of Docker that make sense for you, and remember that the industry isn’t done changing, so you should keep your architecture in a state that allows for iteration.

Other players in this space

Part of Docker’s appeal is its dirt simple bundling of a new approach that removes a lot of pain we’ve been having. But there are other tools out there that solve these same problems.

  • System containers like OpenVZ or LXD provide similar cloud density characteristics
  • rkt is (almost ready to be) a competing application container that promises to implement a more composable architecture
  • Snappy Ubuntu offers an alternative model for transactional installs, bundling dependencies, and isolation
  • Numerous SDN technologies
  • Config management (Puppet, Chef, Ansible) provides deterministic and repeatable deploys
  • Vagrant simplifies local development in production-like environments

I have no doubt that we will look back and see Docker as the catalyst that led to a very different way of treating our software, but it isn’t going to stay the only major player for long, and some of the old rules still apply: Keep your architecture in a place that allows for iteration.

Lxd, cfg mgmt, docker, and the next generation of stateless apps

So what would a cloud architecture that adopts just the good parts of Docker look like?

First off, here are the characteristics that are important to me and that I would like to support:

  • The density and elasticity of containers.
  • Transactional installs and rollbacks for my applications.
  • The ability to develop locally in a near production environment (without near production hardware).
  • Ephemeral systems.
  • Managed systems (I’m not comfortable making them immutable because I trust config management tools more than a home-built “re-roll everything” script to protect me against the next bash vulnerability.).
  • A composable platform that doesn’t rely on the aspects of Docker (like networking) that would make iterating on it difficult.

One way to accomplish this it is to replace our traditional VMs with a system containers like LXD, continue managing infrastructure on those nodes the same way we always have with tools like Puppet, and start installing the service each node maintains in an application container like Docker.

I wish I could put together a demo to illustrate this, but right now running Docker on LXD is infeasible.

With this setup, we would have to change relatively little: We can expose our application on known ports to abstract away nonstandard networking; we only have one app so the namespace security vulnerability isn’t a problem; and our infrastructure only needs incremental updates to support a container instead of a process.

Scaling can be achieved by adding new system container instances, and, while not as fast as spinning a new application container, it’s still something that can be automated. We also don’t have quite the same density, but we’ve retained most of the benefits there as well.

With this as a starting point, we can build out more twelve-factor cloud support based on what’s most important to us: Service discovery, externalized config, software-defined networking, etc.


The tired old debate of “Containers vs. Configuration Management” is rooted in fallacy. We’ve never tried to solve all of our problems with one technology before (You wouldn’t pull down your Java libraries with Puppet, or keep your load balancer config in Maven Central), so why do we have to start now?

Instead, I recommend we do what we always do: Choose the right tool for the right job.

I think there is definitely room for system containers, application containers, and configuration management to coexist. Hopefully more work will be done to make these two great technologies play nicely together.

I (Ben Schwartz) am a software architect at Kasasa by BancVue where lately I’ve been spending most of my time standing on the shoulders of our awesome DevOps culture to play with emergent infrastructure tools and techniques. Sometimes my experimentation makes it to my blog (including the original posting of this article) which can be found at txt.fliglio.com.

This article is part of our Docker and the Future of Configuration Management blog roundup running this November.  If you have an opinion or experience on the topic you can contribute as well

Leave a comment

Filed under DevOps

Docker: Service Management Trumps Configuration Management

Docker – It’s Lighter, Is That Really Why It’s So Awesome?

When docker started to hit so big, I have to admit I initially wondered why.  The first thing people would say when they wanted to talk about its value is “There’s slightly less overhead than virtualization!” Uh… great? But chroot jails etc. have existed for a long time, like even back when I got started on UNIX,and fell out of use for a reason,  and there also hadn’t been a high pressure rush of virtualization and cloud vendors trying to keep up with the demand for “leaner” VMs – there was some work to that end but it clearly wasn’t a huge customer driver. If you cared too much about the overhead, you had the other extreme of just jamming multiple apps onto one box, old school style. Now, you don’t want to do that – I worked in an environment where that was the policy and I developed my architectural doctrine of “sharing is the devil” as a result. While running apps on bare metal is fast and cost effective, the reliability, security, and manageability impediments are significant. But “here’s another option on the spectrum of hardware to VM” doesn’t seem that transformative on its face.

OK, so docker is a process wrapper that hits the middle ground between a larger, slower VM and running unprotected on hardware. But it’s more than that.  Docker also lets you easily create packaged, portable, ready to run applications.

The Problems With Configuration Management Today

The value proposition of docker started to become more clear once the topic of provisioning and deployment came up. Managing systems, applications and application deployments has been at worse a complicated muddle of manual installation, but at best a mix of very complex configuration management systems and baked images (VMs or AMIs). Engineers skilled in chef or puppet are rare. And developers wanted faster turnaround to deploy their applications. I’ve worked in various places where the CM system did app deploys but the developers really, really wanted to bypass that via something like capistrano or direct drop-in to tomcat, and there were always continuous discussions over whether there should be dual tooling, a CM system for system configuration and an app deployment system for app deploys. And if you have two different kinds of tooling controlling  your configuration (especially when, frankly, the apps are the most important part) leads to a lot of conflict and confusion and problems in the long term.

And many folks don’t need the more complex CM functionality. Many modern workloads don’t need a lot of OS and OS software – the enterprise does that, but many new apps are completely self-contained, even to the point of running their own node.js or jetty, meaning that a lot of the complexity of CM is not needed if you’re just going to drop a file onto a  vanilla box and run it. And then there’s the question of orchestration. Most CM systems like to put bits on boxes, but once there’s a running interconnected system, they are more difficult to deal with.  Many discussions about orchestration over the years were frankly rebuffed by the major CM vendors and replied to with “well, then integrate with something (mcollective, rundeck).” In fact, this led to the newer products like ansible and salt arising – they are simpler and more orchestration focused.

But putting bits on boxes is only the first step.  Being able to operate a running system is more important.

Service Management

Back when all of the agile admins were working at National Instruments, we were starting a new cloud project and wanted to automate everything from first principles. We looked at Chef and Puppet but first, we needed Windows support (this was back in 2008, and their Windows support was minimal), and second, we had the realization that a running cloud, REST services type system is composed of various interdependent services, and that we wanted to model that explicitly. We wanted more than configuration management – we wanted service management.

What does it look like when you draw out your systems?  A box and line diagram, right? Here’s part of such a diagram from our systems back then.


Well, not to oversimplify, but when you use something like CloudFormation, you get the yellow boxes (hardware, VMs, cloud instances). When you use something like chef or puppet, you get the white boxes (software on the box). But what about the lines? The point of all those bits is to create services, which are called by customers and/or other services, and being able to address those services and the relationships between them is super important. And trying to change any of the yellow or white boxes without intelligent orchestration to handle the lines – what makes your services actually work – is folly.

In our case, we made the Programmable Infrastructure Environment – modeled the above using XML files and then added a zookeeper-based service registry to handle the connections, so that we could literally replace a database server and have all the other services dependent on it detect that, automatically parse their configurations, restart themselves if necessary, and connect to the new one.

This revolutionized the way we ran systems.  It was very successful and was night and day different from the usual method of provisioning, but more importantly, controlling production systems in the face of both planned and unplanned changes. It allowed us to instantiate truly identical environments, conduct complex deployments without downtime, and collaborate easily between developers and operations staff on a single model in source control that dictated all parts of the system, from system to third party software to custom applications.

That, in conjunction with ephemeral cloud systems, also made our need for CM a lot simpler – not like a university lab where you want it to always be converging to whatever new yum update is out, but creating specifically tooled parts for one use and making new ones and throwing the old ones away as needed. Since we worked at National Instruments, this struck us as on the same spectrum the difference from hand-created hardware boards to FPGAs to custom chips – the latter is faster and cheaper and basically you throw it away for a new one when you need a change, though those others are necessary steps along the path to creating a solid chip.

We kept wondering when the service management approach would catch on.  Ubuntu’s Juju works in this way, but stayed limited to Ubuntu for a long time (though it got better recently!) and hasn’t gotten much reach as a result.

Once docker came out – lo and behold, we started to see that pattern again!

Docker and Service Management

Dockerfiles are simple CM systems that pull some packages, install some software, and open some ports. Here’s an example of a dockerfile for haproxy:

# Haproxy Dockerfile
# https://github.com/dockerfile/haproxy
# Pull base image.
FROM dockerfile/ubuntu
# Install Haproxy.
sed -i ‘s/^# \(.*-backports\s\)/\1/g’ /etc/apt/sources.list && \
apt-get update && \
apt-get install -y haproxy=1.5.3-1~ubuntu14.04.1 && \
sed -i ‘s/^ENABLED=.*/ENABLED=1/’ /etc/default/haproxy && \
rm -rf /var/lib/apt/lists/*
# Add files.
ADD haproxy.cfg /etc/haproxy/haproxy.cfg
ADD start.bash /haproxy-start
# Define mountable directories.
VOLUME [“/haproxy-override”]
# Define working directory.
WORKDIR /etc/haproxy
# Define default command.
CMD [“bash”, “/haproxy-start”]
# Expose ports.

Pretty simple right? And you can then copy that container (like an AMI or VM image) instead of re-configuring every time. Now, there are arguments against using pre-baked images – see Golden Image or Foil Ball. But  at scale, what’s the value of conducting the same operations 1000 times in parallel, except for contributing to the heat death of the universe? And potentially failing from overwhelming the same maven or artifactory server or whatever when massive scaling is required?  There’s a reason Netflix went to an AMI “baking” model rather than relying on config management to reprovision every node from scratch. And with docker containers each container doesn’t have a mess of complex packages to handle dependencies for; they tend to be lean and mean.

But the pressure of the dynamic nature of these microservices has meant that actual service dependencies have to be modeled. Bits of software like etcd and docker compose are tightly integrated into the container ecosytem to empower this. With tools like this you can define a multi-service environment and then register and programmatically control those services when they run.

Here’s a docker compose file:

build: . 
 - "5000:5000"  
 - .:/code  
 - redis 
 image: redis

It maps the web server’s port 5000 to the host port 5000 and creates a link to the “redis” service.  This seems like a small thing but it’s the “lines” on your box and lines diagram and opens up your entire running system to programmatic control.  Pure CM just lets you change the software and the rest is largely done by inference, not explicit modeling. (I’m sure you could build something of the sort in Chef data bags or whatnot, but that’s close to saying “you could code it yourself” really.)

This approach was useful even in the VM and cloud world, but the need just wasn’t acute enough for it to emerge.  It’s like CM in general – it existed before VMs and cloud but it was always an “if we have time” afterthought – the scale of these new technologies pushed it into being a first order consideration, and then even people not using dynamic technology prioritized it. I believe service management of this sort is the same way – it didn’t “catch on” because people were not conceptually ready for it, but now that containers is forcing the issue, people will start to use this approach and understand its benefit.

CM vs. App Deployment?

In addition, managing your entire infrastructure like a build pipeline is easier and more aligned with how you manage the important part, the applications and services.  It’s a lot harder to do a good job of testing your releases when changes are coming from different places – in a production system, you really don’t want to set the servers out there and roll changes to them in one way and then roll changes to your applications in a different way.  New code and new package versions best roll through the same pipeline and get the same tests applied to them. While it is possible to do this with normal CM, docker lends itself to this by default. However, it doesn’t bother to address this on the core operating system, which is an area of concern and a place where well thought out CM integration is important.


The future of configuration management is in being able to manage your services and their dependencies directly, and not by inference. The more these are self-contained, the more the job of CM is simpler just as now that we’ve moved into the cloud the need for fiddling with hardware is simpler. Time spent messing with kernel configs and installing software has dropped sharply as we have started to abstract systems at a higher level and use more small, reusable bits. Similarly, complex config management is something that many people are looking at and saying “why do I need this any more?”  I think there are cases where you need it, but you should start instead with modeling your services and managing those as the first order of concern, backing it with just enough tooling for your use case.

Just like the cloud forced the issue with CM and it finally became a standard practice instead of an “advanced topic,” my prediction is that containers will force the issue with service management and cause it to become more of a first class concern for CM tools back even in cloud/VM/hardware environments.

This article is part of our Docker and the Future of Configuration Management blog roundup running this November.  If you have an opinion or experience on the topic you can contribute as well


Filed under DevOps