This Blog Has an Error Budget Policy

This Blog Has an Error Budget Policy

In my last blog post, I publicized an SLO for this blog. I also mentioned that, in the future, I’d couple the SLO with an error budget and error budget policy. Well, the future is today, because this post will define error budgets and error budget policies and their benefits, before proposing a specific error budget and error budget policy to accompany our previously defined SLO.

What are Error Budget and Error Budget Policies?

via GIPHY

Understanding error budgets and error budget policies is a vital first step in enumerating their pros and cons, and deciding whether we should apply these concepts to this blog.

An error budget is the number of errors we can have, and still be within our SLO. Its easiest to understand with an example. Suppose one component of our SLO is that 99% of requests to our web server are successful, where we define success as a non 5XX status code. Additionally, we are tracking our SLO over a four week rolling window and in the previous four weeks we had 1,000 requests. Our error budget is 10 failed requests over the upcoming four week window.

Suppose we deploy a bad container image, which leads to 5 unsuccessful requests. We just used 50% of our error budget. If, within the same four week window, we experience 10 more unsuccessful requests, we will exhaust our error budget by 50%. When you exhaust an error budget, its time to enact the error budget policy.

An error budget policy enumerates the activity a team takes when they’ve exhausted their error budget for a particular service. It is not intended to punish the team violating the SLO, but rather provide structural support for investing in stability as opposed to new features.

An error budget policy can specify a number of actions, and I some common components below:

  • The team does not deploy any new versions of the service, modulo security fixes and fixes directly addressing the SLO failures, until the service no longer exhausts its error budget.
  • The team devotes X amount of time to working on reliability until the service is back within SLO.
  • The team caps time spend working on new features at Y until the service is back within SLO.

Over time, an error budget regenerates as the errors which exhausted the budget fall outside of the rolling time window. Once the error budget has regenerated, meaning our service its back within its SLO, we can return to development as usual. The constraints enumerated by the error budget policy no longer apply.

Why are Error Budgets and Error Budget Policies Useful?

via GIPHY

From the previous section, we understand what error budgets and error budget policies are. But, are they something we want to apply to our SLOs in general, and the SLO for this blog in particular?

I’d argue the answer is a resounding yes. Both error budgets and error budget policies make an SLOs a more effective tool.

Error budgets add a much needed granularity to SLO measurements. On the surface, an SLO is a binary concept: the service is either within the SLO or it is not. If I’m an engineer trying to understand my application’s current stability, this binary is not sufficient. If all I know is that my SLO is passing, my service could either have zero errors or be one error away from being out of SLO. This lack of clarity makes it difficult to make decisions around when to perform risky changes. An error budget solves this issue. At any time, we can determine what percentage of the error the service has exhausted, and use this measurement to guide the operations we undertake.

Additionally, error budgets allow us to quantify the impact of a negative event, even if that event did not cause us to fail our SLO. For example, suppose we can have 10 errors in a four week window and still meet our SLO. If we first have an event that causes 9 errors and then an event that causes 2 errors, we want to devote more attention to the first event, because it used a greater percentage of our error budget. Yet, if we only have the SLO as a measurement device, we’re more likely to investigate the second event, because it caused our service to move from passing to failing the SLO.

Error budget policies translate the theoretical concepts of SLOs and error budgets into concrete allocations of developer resources. They codify the idea that all changes to services have a concrete risk, and provide consistent guidance around how to approach said risk.

For some, the theoretical strictness of the error budget policy is concerning. Maybe a service just isn’t that important, so they can’t imagine prioritizing stability investments once an error budget is exhausted. Or maybe clients demand new features at a regular cadence, so stopping the deployment of new features is a non-starter. While valid concerns, I’d argue that in those instances, the issue is not with the concept of the error budget policy, but rather with the definition of the SLO. If the team responsible for the service doesn’t foresee themselves wanting, or being able, to take actions to address an SLO failure, then the SLO is false advertising. In that situation, skipping the error budget policy is not the best option; rather, teams should define a more relaxed, and achievable SLO. If the SLO cannot be relaxed, but also there is not support for an error budget policy, then stressful times may be on the horizon… at least SLOs and error budget policies surface these impossible expectations in advance, and hopefully support a discussion of what is realistic.

I’ll be an applying an error budget policy to this blog for much of the same reason I applied an SLO. For one, I’m interested to see how this process works in practice, and this blog seems like a safe way to test it. In addition, I know I have a number of changes I want to make to this blog, and an error budget policy seems like the perfect tool for balancing investing in new features and ensuring this blog remains useful.

Deriving an Error Budget from an SLO

via GIPHY

As I mentioned before, this blog will track its SLO over a four week rolling window. Unfortunately, I don’t have four weeks of historic usage data, so I will need to guess the number of requests this blog received over the last four weeks. To make the math easy, let’s assume 10,000 requests. I’ll update these values on a rolling cadence as we have actual usage data. Remember, our SLO is the following:

  • Availability: The proportion of successful requests, as measured from the web server and where any status code other than 5XX is considered successful, is > 99%.
  • Latency: The proportion of sufficiently fast requests, as measured from the web server and where sufficiently fast is defined as < 1s, is > 99%.

Now, we just need to do some easy math. If we take 10,000 * (1 - .99), we get the number of requests which can fail/not be sufficiently fast, and still be within SLO.

We can define our error budget as follows:

  • Availability: Our error budget for unsuccessful requests, where an unsuccessful request is a request which returns a status code of 5XX, is 100 requests.
  • Latency: Our error budget for not sufficiently fast requests, where not sufficiently fast is defined as > 1s, is 100 requests.

Proposing an Error Budget Policy

via GIPHY

All stakeholders should be involved in the error budget policy’s creation. The involvement of those who approve how developers spend their time is particularly important, as the error budget policy proscribes concrete investments of development effort.

Since this blog is an independent project, I’m the only stakeholder, so I can choose whatever error budget policy I want :)

My error budget policy enumerates the following actions when the error budget is exhausted:

  • Writing new blog posts will be halted until blog is back within SLO.
  • Changes to blog infrastructure (i.e. Kubernetes configuration) will be halted until blog is back within SLO.
  • I’ll try and devote ~1 hr a week to addressing stability issues until the blog is back within SLO.

Publicizing the Error Budget Policy

via GIPHY

Its important to publicize an error budget policy, especially in a collaborative context, so that all stakeholders understand why certain actions are being taken or not taken.

I’ve added this blog’s error budget to the SLO page.

Coming Soon: Monitoring

Thanks for reading this post and hopefully also part 1. In the next post, we’ll wrap up this SLO journey by talking about how we can effectively monitor our SLO.

comments powered by Disqus