Skip to content

Rolling out patches and changes, often and fast

Fefe had a short pointer to an article Patching is Hard. It is, but you can make it a lot easier by doing a few things right.  I did s small writeup (in German) to explain this, which Fefe posted.

I do have an older talk on this, titled “8 rollouts a day” (more like 30 these days). There are slides and a recording. The Devops talk “Go away or I will replace you with a little shell script” addresses it, too, but from a different angle (slides, recording).

Here is the english version of the writeup:

The Patching is Hard article says:

Patching is hard? Yes—and every major tech player, no matter how sophisticated they are, has had catastrophic failures when they tried to change something. Google once bricked Chromebooks with an update. A Facebook configuration change took the site offline for 2.5 hours. Microsoft ruined network configuration and partially bricked some computers; even their newest patch isn’t trouble-free. An iOS update from Apple bricked some iPad Pros. Even Amazon knocked AWS off the air.

Where I am working we made the same observation. 100% uptime is impossible, but management needs metrics and controls to understand if we are still on the right track.

We do have an outage budget. That means we are measuring the actual current business and compare with the predicted business. If a failed rollout creates a loss of (potential) income, we can tell how much that is and deduct that from the outage budget.

We are trying to make the budget as precisely as possible. That does not mean we are trying to create outages on purpose, but having less outage than the budget allows may mean that we are not moving fast enough and have become complacent in what we are doing.

Rolling out often is useful, because then change sets are small and easy to check, much easier than large and hard to understand changes. So we do need to roll out often in order to maintain speed.

Rollout problems tend to manifest in subsystems where we do not roll out often enough. Usually secondary changes, changes in dependencies and libraries, accumulate and then the rollout fails. The solution is to roll out even if the code did not change (that is, we roll out for the sake of keeping dependencies current).

This is a very tangible form of representing technical debt: the larger the diff between production and trunk is, the more likely moving trunk to production is going to result in a failure.

To make patching safe you have to do it often.

To be able to patch often, you need to make patching and rolling out an operative process and not an upgrade/migration project.

To make rollouts and operative thing it is useful to make testing in production safe and survivable.

To be able to test in production, the following things are useful:

  • have at most two versions of a change to a certain system. Do not do three phase or multiphase rollouts. Finish one change before you do the next one.
  • have overcapacity. A change might make it necessary to provide more machinery or more computer power or more memory for a short time. You need to have that or be able to provision this on short notice. Being efficient means having no reserves, no elasticity.
  • separation of code distribution and code activation. That means feature flags and experiments in your code. The same rollout must be able to switch at runtime between old and new behavior.
  • changes to persistent data structures requires that you update both versions of the structure concurrently in order for old and new code to be able to co-exist.
  • you need to know what is going on. Have a good monitoring, central log collection, good search on all of that. There is a thing called monitoring lag, the distance in time between a thing happening and the time you will see this in your monitoring. Measure that lag and have it shown on each screen. Alert on monitoring lag.
  • have a culture of failure that focuses on improvement and learnings. Check out the search term “blameless postmortem” and read up on this.

It’s not actually rocket science, but it will make things a lot better around you.

Stört euch bitte nicht an dem Denglisch, das ist in der Ops-Abteilung von internationalen Firmen durchaus normal 🙂

Indeed. The office is in Amsterdam, and the language at work is English. When I am thinking about work, I am doing it in English. Writing about work in German means that I will use a lot of english terms for the things I am trying to describe, because I’d have to search for appropriate German terms and concepts first and that would interrupt the flow. Sorry about that.

Published inComputer ScienceErklärbärWork


  1. Regarding Denglisch: it’s not just that one would have to search for a term. Very often, the term does not exist and then you end up with compound noun monsters that remind everyone of Mittlere Datentechnik.

    A bit like writing SF in German 🙂

  2. Markus

    Imagine Microsoft would do this with Windows. Then they had to monitor if the patches worked fine. Windows would have to call home to let MS know about any issues.

  3. kiu

    Der erste “Recording” link ist identisch zum 2ten “slides” link…

    • kris kris

      I fixed that. Thanks!

  4. Regarding data structure duplication. Do I understand it correctly that you have two seperate – lets say MySQL – clusters and every change on either the old or the new version of your app results in two writes/updates, one to either cluster? Where do you handle that duplication?

Leave a Reply

Your email address will not be published. Required fields are marked *