Test driven Infrastructure development
Tomas [email protected]@bobtfish
Today, I’m going to talk about the promised land!And by ‘repeatable’, I mean I need to be able to spin up an arbitrary set of servers for any environment I want, whenever I want - so _all_ the configuration of all the instances has to be dynamic!
• High availability!
Today, I’m going to talk about the promised land!And by ‘repeatable’, I mean I need to be able to spin up an arbitrary set of servers for any environment I want, whenever I want - so _all_ the configuration of all the instances has to be dynamic!
• High availability!
• Automated testing of all infrastructure changes
Today, I’m going to talk about the promised land!And by ‘repeatable’, I mean I need to be able to spin up an arbitrary set of servers for any environment I want, whenever I want - so _all_ the configuration of all the instances has to be dynamic!
• High availability!
• Automated testing of all infrastructure changes
• Entirely repeatable application environments
Today, I’m going to talk about the promised land!And by ‘repeatable’, I mean I need to be able to spin up an arbitrary set of servers for any environment I want, whenever I want - so _all_ the configuration of all the instances has to be dynamic!
• High availability!
• Automated testing of all infrastructure changes
• Entirely repeatable application environments
• High confidence in changes
Today, I’m going to talk about the promised land!And by ‘repeatable’, I mean I need to be able to spin up an arbitrary set of servers for any environment I want, whenever I want - so _all_ the configuration of all the instances has to be dynamic!
• High availability!
• Automated testing of all infrastructure changes
• Entirely repeatable application environments
• High confidence in changes
• Continuous integration and deployment for infrastructure
Today, I’m going to talk about the promised land!And by ‘repeatable’, I mean I need to be able to spin up an arbitrary set of servers for any environment I want, whenever I want - so _all_ the configuration of all the instances has to be dynamic!
So who the hell am I?
Dev
Infrastructure automation nut!Ex-backend web developer, Ex-security, currently fixing puppet at Yelp!
Dev / Ops
State of repeatability and testing in infrastructures is generally shocking.Leads to systems/operations teams being adverse to change and conservative - slows the business down!Why isn’t your infrastructure an agile software project?
Dev / Ops•Developer viewpoint
State of repeatability and testing in infrastructures is generally shocking.Leads to systems/operations teams being adverse to change and conservative - slows the business down!Why isn’t your infrastructure an agile software project?
Dev / Ops•Developer viewpoint
•Grass IS greener
State of repeatability and testing in infrastructures is generally shocking.Leads to systems/operations teams being adverse to change and conservative - slows the business down!Why isn’t your infrastructure an agile software project?
Dev / Ops•Developer viewpoint
•Grass IS greener
State of repeatability and testing in infrastructures is generally shocking.Leads to systems/operations teams being adverse to change and conservative - slows the business down!Why isn’t your infrastructure an agile software project?
Dev / Ops•Developer viewpoint
•Grass IS greener
•Think of your infra as an agile software project...
State of repeatability and testing in infrastructures is generally shocking.Leads to systems/operations teams being adverse to change and conservative - slows the business down!Why isn’t your infrastructure an agile software project?
Dev / Ops•Developer viewpoint
•Grass IS greener
•Think of your infra as an agile software project...
•What workflow do I want?
State of repeatability and testing in infrastructures is generally shocking.Leads to systems/operations teams being adverse to change and conservative - slows the business down!Why isn’t your infrastructure an agile software project?
The state of the art
Going to talk about how I think the generally accepted way of doing some things is fundamentally broken!But lets start with a simple description of the issues I’m worrying about.
CM = state machine
Each change puppet makes (or attempts to make) is a state transition. Each circle represents the configuration state of the server on disc + services running etc..
Non deterministic
This is the key observation here - you don’t know which way puppet’s gonna jump :)In this case - it doesn’t matter, as the two operations are orthogonal.
Convergent!
Convergence is when each run of puppet takes you nearer to 0 changes, but the next run makes additional changes..The classic way to screw this up is to miss a dependency in your code.
Convergent!
Of course, this doesn’t happen - the first step goes BANG, then mysql gets installed, creates /etc/mysql.The second puppet run _then_ sets the config up..
err: /Stage[main]//File[/etc/mysql/my.cnf]/ensure: change from absent to file failed: Could not set 'file on ensure: No such file or directory - /etc/mysql/my.cnf.puppettmp_3706 at /home/tdoran/test.pp:4
Aaand in your puppet logs, you get.
Purple text of rage!
err: /Stage[main]//File[/etc/mysql/my.cnf]/ensure: change from absent to file failed: Could not set 'file on ensure: No such file or directory - /etc/mysql/my.cnf.puppettmp_3706 at /home/tdoran/test.pp:4
THE PURPLE TEXT OF RAGE
Convergent!
(Shamelessly stolen from https://www.usenix.org/legacy/publications/library/proceedings/lisa02/tech/full_papers/traugott/traugott.pdf)
Aaand your machine is convergent - i.e. it gets towards the desired state in a number of steps..
•before
• require
• subscribe
•notify
As I noted, this all happens as you missed a dependency. This is the easy case, where puppet can detect hat and tell you! It’s also entirely possible to be totally silent.It is though totally possible to write your puppet code well enough to need EXACTLY 1 puppet run to fully provision a server!
Fixable!
•before
• require
• subscribe
•notify
As I noted, this all happens as you missed a dependency. This is the easy case, where puppet can detect hat and tell you! It’s also entirely possible to be totally silent.It is though totally possible to write your puppet code well enough to need EXACTLY 1 puppet run to fully provision a server!
Fixable!
•before
• require
• subscribe
•notify
What about an entire
infrastructure?
The $64,000 question is....
A whole stack
Lets start simple, but semi realistic.Gonna ignore databases.Gonna ignore monitoring.Gonna ignore the n[eo]twork.
Exported resources
Each layer of systems can publish data to the systems which depend on it. (I.e. webs register, proxies find the webs + register themselves, lbs then find the proxy).Given you know the dependencies - you can get consistent runs by ordering them.
Exported resources
• Inter machine dependencies
Each layer of systems can publish data to the systems which depend on it. (I.e. webs register, proxies find the webs + register themselves, lbs then find the proxy).Given you know the dependencies - you can get consistent runs by ordering them.
Exported resources
• Inter machine dependencies
• Unidirectional!
Each layer of systems can publish data to the systems which depend on it. (I.e. webs register, proxies find the webs + register themselves, lbs then find the proxy).Given you know the dependencies - you can get consistent runs by ordering them.
Exported resources
• Inter machine dependencies
• Unidirectional!
• Known graph - webs, proxies, lbs
Each layer of systems can publish data to the systems which depend on it. (I.e. webs register, proxies find the webs + register themselves, lbs then find the proxy).Given you know the dependencies - you can get consistent runs by ordering them.
Exported resources
• Inter machine dependencies
• Unidirectional!
• Known graph - webs, proxies, lbs
• Puppetroll (github.com/youdevise/puppetroll)
Each layer of systems can publish data to the systems which depend on it. (I.e. webs register, proxies find the webs + register themselves, lbs then find the proxy).Given you know the dependencies - you can get consistent runs by ordering them.
Exported resources
(Shameless ripoff of http://xkcd.com/1171/ )
Ordering dependent. Hard to test (in isolation). Slooow (have to run in order)
Co-dependence
And if we really are talking about entire infrastructures...Then maybe we need some of these.
Co-dependence
:(You _know_ that if everything is dynamically configured that you’re gonna have to do multiple puppet runs per server...Do we _really_ want to keep running puppet till it stops changing things?
The solution - an external model
Use your software model to generate a set of machines for an environment.And generate config for puppet to apply to each system to configure itAdd super secret special sauce (lots and lots of mcollective!)
The solution - an external model
• Represent system as a set of ruby classes
Use your software model to generate a set of machines for an environment.And generate config for puppet to apply to each system to configure itAdd super secret special sauce (lots and lots of mcollective!)
The solution - an external model
• Represent system as a set of ruby classes
• DSL for describing environments
Use your software model to generate a set of machines for an environment.And generate config for puppet to apply to each system to configure itAdd super secret special sauce (lots and lots of mcollective!)
The solution - an external model
• Represent system as a set of ruby classes
• DSL for describing environments
• Dependencies
Use your software model to generate a set of machines for an environment.And generate config for puppet to apply to each system to configure itAdd super secret special sauce (lots and lots of mcollective!)
The solution - an external model
• Represent system as a set of ruby classes
• DSL for describing environments
• Dependencies
• Domain knowledge
Use your software model to generate a set of machines for an environment.And generate config for puppet to apply to each system to configure itAdd super secret special sauce (lots and lots of mcollective!)
This is a simplified / minimal example jenkins environment - just 4 machines (2 web apps, 2 load balancers)
ENC data!
Our external node classifier generates this for each of the 4 machines, which translates to puppet code run on the server.Note how every server gets all of it’s dependenciesThere’s a companion data structure sent to the agent which actually provisons the virtual machines
Call tree looks something like this: Model all the nodes, allocate all their IPs. Make calls to KVM servers to provision machines.. VMs start, boot, run puppet, send cert to puppetmaster, --waitforcert.Central provisioning asks ‘do we have a cert’, waits - signs it. Looks up DNS and ENC to compile catalogue. Catalog shipped to node, runs puppet. Provisioning uses MCO to determine when puppet finished.. When all nodes up, check nrpe all green on all nodes, then run end to end app tests!
Automate all the things
Suddenly, I have massive power.I can write a small script to bring up a whole production like environment, run tests against it, tear it down. I can do this against the latest puppet changes, and only promote them to run on production servers when the tests pass!
BDD infrastructure
Behavior driven development - given I have a high level model of the systems comprising an infrastructure, I can then write equally high level tests to assert the behavior of that infrastructiure
BDD infrastructure• Given
For example...
BDD infrastructure• Given – the Service has finished being
provisioned
BDD infrastructure• Given – the Service has finished being
provisioned
• And
BDD infrastructure• Given – the Service has finished being
provisioned
• And – all monitoring related to the service is passing
BDD infrastructure• Given – the Service has finished being
provisioned
• And – all monitoring related to the service is passing
• When
BDD infrastructure• Given – the Service has finished being
provisioned
• And – all monitoring related to the service is passing
• When – when we destroy a single member of the service
BDD infrastructure• Given – the Service has finished being
provisioned
• And – all monitoring related to the service is passing
• When – when we destroy a single member of the service
• Then
BDD infrastructure• Given – the Service has finished being
provisioned
• And – all monitoring related to the service is passing
• When – when we destroy a single member of the service
• Then – we expect all monitoring at the service level to be passing
BDD infrastructure• Given – the Service has finished being
provisioned
• And – all monitoring related to the service is passing
• When – when we destroy a single member of the service
• Then – we expect all monitoring at the service level to be passing
• And
BDD infrastructure• Given – the Service has finished being
provisioned
• And – all monitoring related to the service is passing
• When – when we destroy a single member of the service
• Then – we expect all monitoring at the service level to be passing
• And – we expect all monitoring at the single machine level to be failing
Yes, I am suggesting regression testing your load balancer setup...
Is this for real?
Is this for real?
• Yes!
Is this for real?
• Yes!
• We actually built this, the core parts are on github
Is this for real?
• Yes!
• We actually built this, the core parts are on github
• Deployed real applications to production at TIM Group
• High availability!
• Automated testing of all infrastructure changes
• Entirely repeatable application environments
• High confidence in changes
• Continuous integration and deployment for infrastructure
This is my promised land!
Questions?
• https://devblog.timgroup.com/2013/06/14/exported-resources-considered-harmful/
• https://devblog.timgroup.com/2013/06/26/scenario-testing-infrastructures/
• https://github.com/youdevise/provisioning-tools
• https://github.com/youdevise/stackbuilder