Skip to content
Notifications
Clear all

Has anyone run a cost-benefit on using more powerful, expensive runners?

1 Posts
1 Users
0 Reactions
59 Views
(@datadog)
Reputable Member
Joined: 4 months ago
Posts: 365
Topic starter   [#11461]

We’ve been running a mix of t3.large and c5.2xlarge EC2 spot instances as self-hosted GitHub runners. The c5s cost about 2.5x more per hour.

The raw data from our cost allocation tags and GitHub Actions metrics over the last quarter:

* **t3.large (2 vCPU, 8 GiB):** Average build time: 22 minutes. Cost per build: ~$0.08
* **c5.2xlarge (8 vCPU, 16 GiB):** Average build time: 6 minutes. Cost per build: ~$0.09

The faster runner finishes the build in roughly 1/4 of the time for only a ~12% higher cost per build.

But the real metric is developer wait time for merges. Our `main` branch queue backs up during peak hours. Faster runners clear the backlog significantly quicker, which reduces the chance of merge conflicts and devs context-switching.

The break-even analysis is simple: if your team's hourly cost (fully loaded) is greater than the marginal cost of the faster hardware, you're losing money by saving on compute. For us, that threshold is very low.

Here's the Prometheus query we used to correlate runner type with job duration from the actions metrics:
```promql
avg by (runner_label) (github_actions_workflow_job_duration_seconds{repo="myapp"})
```

Has anyone else run the numbers and found a different inflection point? I'm curious about the cost vs. time trade-off for memory-bound jobs (Java, Go) versus pure CPU-bound jobs.


Metrics don't lie.


   
Quote