🏏 450 million fans watched the last IPL. What is 'Cricket Scale' for SREs? Know More

Mar 16th, ‘22/3 min read

Why Service Level Objectives?

Understanding how to measure the health of your servcie, benefits of using SLOs, how to set compliances and much more...

Why Service Level Objectives?

Assuming that you have read the Primer on Service Level Objectives (SLOs), let’s talk about the benefits of adopting SLOs.

How do Service Level Objectives help over regular (manual or per-interval threshold-based) alerting?

Short answer

Engineers and Engineering is finite resource. So would you spend them on new Features or fixing the existing problems?

Long answer (keep reading)

📊 Let’s illustrate the need using a sample service’s metrics over 15 minutes.

Timestamp Throughput 5xx
18:00 500 10
18:01 600 0
18:02 500 0
18:03 2000 20
18:04 150 5
18:05 150 10
18:06 150 15
18:07 600 12
18:08 600 5
18:09 700 5
18:10 2500 20
18:11 250 5
18:12 700 7
18:13 800 12
18:14 1250 5
18:15 1500 10
18:16 1500 5
18:17 2500 10
18:18 1000 10

3️⃣ Three questions that come to mind

  1. Is this Service reliable?
  2. Should it alert and when?
  3. Third, how severe is the situation?

📈 Traditionally, your favorite dashboard tool would show you something like this.

🕡 There is a spike at 6:03 PM and 6:06 PM but look at the quantum of those errors.

🪂 SREs usually fall into this trap of misleading Spikes.

100% uptime is not possible. A downtime, no matter how small, is inevitable. Because you need some time off to provide for the following, some requests are bound to fail.
  • 🔺Upgrades
  • 🔧 Maintenance
  • 🏖️ Comfort to Engineers
Fair to conclude that the only way to make a service reliable is by allowing it to fail a little

But this downtime, albeit small, must NOT:

  • Bring down other services in the system.
  • Impact a large majority of the users.
  • Become a trend.
The key is to stop chasing every single error. Instead, look at the Service health.
But, you cannot improve health, or anything else for that matter, without an Objective! So let's say our objective for this service is 99% health

Look at the graphs again. Is the Service Healthy?

Maybe it is?

Total Duration 15 minutes
Total Requests (Sum of throughput over the duration) 17,950
Total Failed (Sum of 5xx over the duration) 156
Number of minutes where error_rate (5xx/throughput) > 1% 53%
Non-error Requests / Total Requests 99.13%

Here’s a cumulative representation of the Service’s health, or what we call a Service Level Objective.

🟢 The green line indicates the proportion of total non-error requests till that corresponding time point.

🔴 The red line is a static line at 99% health.

Regular Alerting:
There would be 17 alarms for every time the Service’s error_rate was > 1%. And this would result in:
- Alert fatigue is real bro!
- Constant distraction towards Maintenance and no time for features.

Service Level Objective:
The only alert that mattered was when the Service Level Objective dropped below the desired 99% because they are not sensitive towards point-in-time fluctuations.

How to set Service Level Objectives?

At first, the thought of not paying attention to every Error seems counterintuitive, but as the business matures and the scale increases, one agrees that failures are inevitable.

How do you define the “acceptable” behavior of this Service? And the variables are:

  • 🎯 How many requests should be allowed to fail, or Compliance Target.
  • ⏱️ Say you are ok if 1% of the requests fail. 1% over what Compliance Period? A year, a week, a month, an hour ......?

How to choose the correct compliance?

⏱️ Compliance Period

Usually 1x or 2x of your engineering sprint cycle; this ensures that errors from one sprint do not break the budget of the next sprint. However, shorter periods also tend to be more sensitive to failure and may create more noise.

🎯 Compliance Target

Make sure to choose a limit that serves you well without multiple breakages. For example, if you desire to set 99.9%, it’s wise to start with 99% and see if it served you well.

Further read: head over to our blog SLA vs SLO to better understand the differences between the two.



Stay updated on the latest from Last9.


Piyush Verma

Co-Founder, CTO at Last9

Handcrafted Related Posts