Superhuman reduces their build time from 22 minutes to 8 minutes with Semaphore

  • ⛔️️ Different tools responsible for the CI process
  • ⛔️ A slow and costly feedback loop
  • βœ… One tool to rule them all
  • βœ… A faster and cheaper feedback loop

Try out Semaphore

The Challenge

Superhuman prides itself on being the fastest email client in the world, and wants to live the value of speed for all of their internal tooling. Before Semaphore, running the tests and deploying was too slow. The team behind Superhuman used Wercker for the backend (which was acquired by Oracle a few years ago). Their frontend CI service used the β€œold-school model” of charging for peak parallelism, and they also had yet another CI service to run the builds for the desktop client.

Superhuman needed a service that:

  1. Was built from the ground up to be fast.
  2. Could support macOS code-signing.
  3. Charged for actual time used, not peak parallelism.
  4. Has a high level of peak parallelism.
  5. Who’d reply to them when they emailed.

The Solution

The Superhuman backend is written in Go. When code is pushed to any branch, Semaphore builds and runs the tests. When the tests pass on 'master' or on 'staging', Semaphore builds Docker images, uploads those to Google Container Registry, and then gradually rolls them out to our Google Container Engine cluster.

The frontend is written in JavaScript. When code is pushed to any branch, Semaphore runs Zen (a test runner built by one of their engineers) to run all the tests. Zen compiles the code, uploads it to S3 and boots up 600 AWS Lambda instances to run the tests in parallel (the Semaphore worker co-ordinates this). When the tests pass on the 'staging', or 'production' branches, Semaphore automatically builds a bundle and uploads it to Google Cloud Storage.

our solution

The Results

Semaphore has significantly better performance, the machines are not overloaded and you can run many builds in parallel. It is also way cheaper than the old model reserving capacity and resources for 10 parallel jobs (which is a waste 80% of the time, and not enough 10% of the time).

  • βœ”οΈ Better performance
  • βœ”οΈ The machines are not overloaded
  • βœ”οΈ Run many builds in parallel
  • βœ”οΈ Cheaper than the old model
Industry

Job search & Recruitment

Stack

Go

Javascript

Docker

GKE

Aws Lambda

Star us on GitHub