Continuous Delivery

Starting and Scaling DevOps in the Enterprise – review

As many of you know, I’m a huge fan of the work Gary Gruver has done – in particular his book “Leading the Transformation” on his experiences at HP trying to transform a very traditional enterprise. (See my earlier mention of his book on this blog, here.) His newest work is out – Starting and Scaling DevOps in the Enterprise. I am recommending it very highly to all my customers that are following DevOps! I think its unique – by far the best I’ve read so far when it comes to putting together specific metrics and the questions you’ll need to know in setting your priorities.

Gary notes that there are three types of work in an enterprise:

  1. New work – Creating new features or integrating/building new applications
    1. new work can’t be optimized (too much in flux)
    2. Best you can hope for here is to improve the feedback loop so you’re not wasting time polishing features that are not needed (50%+ in most orgs!)
  2. Triage – finding the source of defects and resolving
    1. Here DevOps can help by improving level of automation. Smaller batch sizes means fewer changes to sort through when bugs crop up.
  3. Repetitive – provisioning environments, building, testing, configuring the database or firewall, etc.
    1. More frequent runs, smaller batches, feedback loop improved. All the DevOps magic really happens in #2 and #3 above as these are the most repetitive tasks.

Notice of the three types above – the issues could be in one of five places:

  1. Development
    1. Common pain point here is Waterfall planning – i.e. requirements inventory and a bloated, aging inventory)
  2. Building Test Environments
    1. Procurement hassles across server, storage, networking, firewall. Lengthy handoffs between each of these teams and differing priorities.
    2. Horror story – 250 days for one company to attempt to host a “Hello World” app. It took them just 2 hours on AWS!
  3. Testing and Fixing Defects – typically QA led
    1. Issues here with repeatability of results (i.e. false positives caused by the test harness, environment, or deployment process)
    2. Often the greatest pain point, due to reliance on manual tests causing lengthy multi-week test cycles, and the time it takes to fix the defects discovered.
  4. Production Deployment – large, cross org effort led by Ops
  5. Monitoring and Operations

The points above are why you can’t just copy the rituals from one org to another. For any given company, your pain points could be different.


So, how do we identify the exact issue with YOUR specific company?

  1. Development (i.e. Requirements)
    1. Metrics:
      1. What % of time is spent in planning and documenting requirements?
      2. How many man-hours of development work are currently in the inventory for all applications?
      3. What % of delivered features are being used by customers and fit the expected results?
    2. An important note here – organizations often commit 100% of dev resources to address work each sprint. This is terrible as a practice and means that the development teams are too busy meeting preset commitments to respond to changes in the marketplace or discoveries during development. The need here is for education – to tell the business to be reasonable in what they expect, and how to shape requirements so they are actual minimum functionality needed to support their business decisions. (Avoid requirements bloat due to overzealous business analysts/PM’s for example!)

  1. Provisioning environments
    1. Metrics:
      1. How much time does it take to provision environments (on avg)
      2. How many environments are requested per month/sprint
      3. % of time these environments require manual fixing before they are complete
      4. % of defects associated with non-code – i.e. environments, deployments, data layer, etc.
    2. The solution here for provisioning pinch points is infrastructure as code. Here there is no shortcut other than developers and IT/operations working together to build a working set of scripts to recreate environments and maintaining them jointly. This helps with triage as changes to environments now show up clearly in source control, and prevents DEV-QA-STG-PROD anomalies as it limits variances between environments.
    3. It’s critical here for Dev and Ops to use the same tool to identify and fix issues. Otherwise strong us vs them backlash and friction.
    4. This requires the organization to have a strong investment in tooling and think about their approach – esp with simulators/emulators for companies doing embedded development.

  1. Testing
    1. Metrics
      1. What is the time it takes to run a full set of tests?
      2. How repeatable are these? (i.e. what’s the % of false errors)
      3. What % of defects are found with testing (either manual, automated, or unit testing)
      4. What is the time it takes to approve a release?
      5. What’s the frequency of releases?
    2. In many organizations this is the most frequent bottleneck – the absurd amount of time it takes to complete a round of tests with a reasonable expectation the release will work as designed. These tests must run in hours, not days.
    3. You must choose a well-designed automation framework.
    4. Development is going to have to change their practices so the code they write is testable. And they’ll need to commit to making build stability a top priority – bugs are equal in priority (if not higher than) tasks/new features.
    5. This is the logical place to start for most organizations. Don’t just write a bunch of automated tests – instead just a few automated Build Acceptance Tests that will provide a base level of stability. Watch these carefully.
      1. If the tests reveal mostly issues with the testing harness, tweak the framework.
      2. If the tests are finding mostly infrastructure anomalies, you’ll need to create a set of post-deployment tests to check on the environments BEFORE you run your gated coding acceptance test. (i.e. fix the issues you have with provisioning, above).
      3. If you’re finding coding issues or anomalies – congrats, you’re in the sweet spot now!
    6. Horror story here – one company boasted of thousands of automated tests. However, these were found to not be stable, maintainable, and had to be junked.
    7. Improve and augment over time these BATs so your trunk quality gradually moves closer to release in terms of near-produciton quality.
      1. Issue – what about that “hot” project needed by the business (which generally arrives with a very low level of quality due to high pressure?
        1. Here the code absolutely should be folded into the release, but not exposed to the customer until it fits the new definition of done: “All the stories are signed off, automated testing in place and passing, and no known open defects.”

  1. Release to Production
    1. If a test cycle takes 6 weeks to run, and management approval takes one day – improving this part just isn’t worth it. But if you’re trying to do multiple test cycles a week and this is the bottleneck, absolutely address this with managers that are lagging in their approval or otherwise not trusting the gated testing you’re doing.
    2. Metrics
      1. Time and effort to release to production
      2. Number of issues found categorized by source (code, environment, deployment process, data, etc)
      3. Number of issues total found in production
      4. MTTR – mean time to restore service
      5. # of green builds a day
      6. Time to recover from a red build
      7. % of features requiring rework before acceptance
      8. Amt of effort to integrate code from the developers into a buildable release
    3. For #1-4 – Two areas that can help here are feature toggling (which you’ll be using anyway), and canary releases where key pieces of new functionality are turned on for a subset of users to “test in production.”
    4. For #5-6 – here Continuous Integration is the healer. This is where you avoid branching by versioning your services (and even the database – see Refactoring Databases book by Scott)
    5. For #7-8 – If you’re facing a lot of static here likely a scrum/agile coach will help significantly.


So – how to win, once you’ve identified the pain points? You begin by partitioning the issue:

  • Break off pieces that are tightly coupled versus not developed/tested/deployed as a unit. (i.e. HR or Purchasing processes)
  • Segment these into business critical and non-business critical.
  • Split these into tightly coupled monoliths with common code sharing requirements vs microservices (small, independent teams a la Amazon). The reality is – in most enterprises there’s very valid reasons why these applications were built the way they are, You can’t ignore this complexity, much as we’d like to say “microservices everywhere!”

I really admire Gary’s very pragmatic approach as it doesn’t try to accomplish large, difficult things all at once but it focuses on winnable wars at a company’s true pain points. Instead of trying to force large, tightly coupled organizations to work likely loosely coupled orgs – you need to understand the complex systems and determine together how to release code more frequently without sacrificing quality. Convince these teams of DevOps principles.



This is the fourth of a series on DevOps. The first focused on the three ways explored in the Phoenix Project, and I stuck in some thoughts from the Five Dysfunctions of a Team by Lencioni. The second discussed the lessons taught by GM’s failure in adopting Toyota’s Lean processes with their NUMMI plant. The third went through some great lessons I’ve learned from “Visible Ops” by Gene Kim.

“The single largest improvement an IT organization can benefit from is implementing repeatable system builds. This can’t be done without first managing change and having an accurate inventory. When you convert a person-centric and heavily manual process to a quick and repeatable mechanism, the reaction is always positive. Even a partially automated release/build process greatly improves the ability for individuals to be freed from firefighting and focus on their areas of real value. And by making it more efficient to rebuild than repair, you also get much faster systems downtime and significantly reduced downtime.” (Joe Judge, Adero)



So I am putting together a presentation for PADNUG tomorrow on DevOps. I’ve reworked this presentation like three times, and I’ve never been very happy with it. Let’s just say Steve Jobs would have rolled his eyes at something like this:

Look at that crap above. I mean, there’s information here – but way too MUCH information. There’s no way any audience is going to absorb this. I’ll lose them halfway through the second bullet point.

So, I was struggling with this a few weeks ago, trying to come up with a better idea. And I was watching my kids play Monopoly. And I started to think – since there’s no recipe for DevOps, and you can choose your own course, and some amount of it is up to chance or your individual circumstances – well, isn’t that a game? (And isn’t that a more fun way of learning than using an endless stream of bullet points?)

So, DevOpoly was born!

Let’s take a look at this in blocks shall we?

  • MTTR – Mean Time to Repair. This indicates how robust you are, how quickly you can respond and react to an issue.
  • Stakeholder Signoff – this is after you inventory your applications – instituting any change management policy and change window will require the business to provide signoff.
  • Inventory Apps – listing applications, servers, systems and services in tiers. This is a prereq for getting your problem children identified and frozen, see below.
  • CAB Weekly Meetings – I used to think these were a complete and total waste of time. In fact several books I have claim that they don’t measurably reduce defects and slow down development – bureaucracy at its worst. But, Gene Kim swears by it – and he thinks it’s a base level requirement for change management culture.

  • Versioned Patches – Putting any software patches into source control
  • Security Auditing – having controls that are visible, verifiable, regularly reported
  • Configuration Management – Infrastructure as Code, a key part of implementing repeatable system builds, using software like Puppet, Chef, Octopus etc.
  • Golden Build – The end goal and the building block of a release library, a set of ‘golden builds’ that are verifiable and QA’d. The length of time that these builds stay stable is another metric helpful in determining reliability of your apps.

  • Feed to Trouble Ticket – Creating a system where any changes – authorized or unauthorized – show up in trouble ticket for first responders to access. % Success rate in first response in diagnosis is a key metric for DevOps.
  • Dashboarding – creating visibility around these metrics (see stage 3 of the Phoenix Project post) is the only way you’ll know if you’re making progress – and securing management support.
  • Form RM Team – This is part of the process in moving more staff away from firefighting and early in the release process. Mature, capable orgs have more personnel assigned to protect quality early on versus catching defects late.


  • MTBF – Mean Time Between Failures. As configuration management knocks out snowflake servers and fragile artifacts are frozen, this number should go up.
  • Automated Release – creating a release management pipeline of dev bits from DEV-QA-STG-PROD, with as much automated signoff as possible using automated tests, is a great step forward.
  • Gated Builds – See above, but having functional/integration testing and unit tests run on checkin is key to prevent failures.
  • Continuous Integration – bound up with testing and the RM cycle – having any dev changes get checked in and validated and merged safely with other development changes. (And, remember, CI means the barest amount of release branching possible. It’s a tough balance.)

  • Eliminate Access – Actually I don’t know many devs (besides the true cowboys) that really WANT access to production. But, removing access to all but change managers is a key step. And when you’re done with that…
  • Electrify the Fence – Have change policy known and discipline the (inevitable) slow learners. Not fire them. Maybe have a few “disappear” in suspicious accidents, to warn the others!
  • Monitor Changes – Use some software (like Tripwire maybe?) to monitor any and all changes to the servers.
  • Server to Admin Ratio – Typically this is a 15:1 ratio – but for high performing orgs with an excellent level of change management, 100:1 or greater is the norm.

  • Document Policy – Writing out the change management policy is a key to electrifying the fence and preventing the org from slipping back into bad habits.
  • Rebuild Not Repair – With a great release library of golden builds and a minimal amount of unique configs and templates, infrastructure is commonly rebuilt – not patched and limping along.

  • Find Fragile Artifacts – Once you’ve done your systems inventory, you can document the systems that have the lowest uptime, the highest impact to the business when its down, and the most expensive infrastructure.
  • Enforce Change Window – Set a change window for each set of your applications, and freeze any and all changes outside of that window. It must be documented and stakeholders must provide signoff.
  • Soft Freeze Fragile Systems – These fragile artifacts have to be frozen, one by one, until the environments can be safely replicated and maintained. This soft freeze can’t last long until the systems are part of configuration management/IAC.

  • Accountability – #1 of the two failure points in any change. True commitment and accountability from each person involved.
  • Firefighting Tax – Less than 5% of time spent in firefighting is a great metric to aim for. Most organizations are at about 40%.
  • Management Buy-In – DevOps can be started as a grassroots effort, but for it to be successful- it must have solid buy-in from the top. Past a pilot effort, you must secure management approval by publicizing your dashboards and key metrics.

Anyway, this was fun. I have some cards on the way for both the Gene Kim Chest – yes, not Jez Humble, but I’m thinking about it – and Chance. Lots of chance in the whole DevOps world.

(I tried this back in August with Life but it never worked by the way.)