Coinbase Logo

Continuous Integration at Coinbase: How we optimized CircleCI for speed and cut our build times by 75%

By Author


, July 10, 2019

Tuning a continuous integration server presents an interesting challenge — infrastructure engineers need to balance build speed, cost, and queue times on a system that many developers do not have extensive experience managing at scale. The results, when done right, can be a major benefit to your company as illustrated by the recent journey we took to improve our CI setup.

Continuous Integration at Coinbase

As Coinbase has grown, keeping our developers happy with our internal tools has been a high priority. For most of Coinbase’s history we have used CircleCI server, which has been a performant and low-maintenance tool. As the company and our codebase have grown, however, the demands on our CI server have increased as well. Prior to the optimizations described here, builds for the monorail application that runs had increased significantly in length (doubling or tripling the previous average build times) and developers commonly complained about lengthy or non-finishing builds.

Our CI builds were no longer meeting our expectations, and it was with the previous issues in mind that we decided to embark on a campaign to get our setup back into shape.

It’s worth sharing here that Coinbase specifically uses the on-premise server version of CircleCI rather than their cloud offering — hosting our own infrastructure is important to us for security reasons, and these concepts specifically apply to self-managed CI clusters.

The Four Golden Signals

We found the first key to optimizing any CI system to be observability, as without a way to measure the effects of your tweaks and changes it’s impossible to truly know whether or not you actually made an improvement. In our case, server-hosted CircleCI uses a nomad cluster for builds, and at the time did not provide any method of monitoring your cluster or the nodes within. We had to build systems of our own, and we decided a good approach would be using the framework of the four golden signals, Latency, Traffic, Errors, and Saturation.


Latency is the total amount of time it takes to service a request. In a CI system, this can be considered to be the total amount of time a build takes to run from start to finish. Latency is better measured on a per-repo or even per-build basis as build length can vary hugely based on the project.

To measure this, we built a small application that queried CircleCI’s API regularly for build lengths, and then shipped over that information to Datadog to allow us to build graphs and visualizations of average build times. This allowed us to chart the results of our improvement experiments empirically and automatically rather than relying on anecdotal or manually curated results as we had done previously.


CircleCI API results


Traffic is the amount of demand being placed on your system at any one time. In a CI system, this can be represented by the total number of concurrently running builds.

We were able to measure this by using the same system we built to measure latency metrics. This came in handy when determining the upper and lower bounds for the use of our build resources as it allowed us to see exactly how many jobs were running at any one time.


Errors are the total amount of requests or calls that fail. In a CI system this can be represented by the total number of builds that fail due to infrastructural reasons. It’s important here to make a distinction between builds that fail correctly, due to tests, linting, code errors, etc. rather than builds that fail due to platform issues.

One issue we encountered was that occasionally AWS would give us “bad” instances when spinning up new builders that would run much slower than a normal “good” instance. Adding error detection into our builder startup scripts allowed us to terminate these and spin up new nodes before they could slow down our running builds.


Saturation is how “full” your service is, or how much of your system resources are being used. In a CI system, this is fairly straightforward — how much I/O, CPU, and memory are the builders under load using.

To measure saturation for our setup we were able to tap into cluster metrics by installing a Datadog Agent on each of our builders, which allowed us to get a view into system stats across the cluster.


Datadog job statistics

Identifying the Root Cause

Once your monitoring setup is in place it becomes easier to dig into the root cause of build slowdowns. One of the difficulties in diagnosing CI problems without cluster-wide monitoring is that it can be hard to identify which builders are experiencing load at any one time or how that load affects your builds. Latency monitoring can allow you to figure out which builds are taking the longest, and saturation monitoring can allow you to identify the nodes running those builds for closer investigation.

For us, the new latency measuring we added allowed us to quickly confirm what we had previously guessed: not every build was equal. Some builds ran at the quick speeds we had previously been experiencing but other builds would drag on for far longer than we expected.

In our case this discovery was the big breakthrough — once we could quickly identify builds with increased latency and find the saturated nodes the problem quickly revealed itself: resource contention between starting builds! Due to the large number of tests for our larger builds we use CircleCI’s parallelization feature to split up our tests and run them across the fleet in separate docker containers. Each test container also requires another set of support containers (Redis, MongoDB, etc.) in order to replicate the production environment. Starting all of the necessary containers for each build is a resource-intensive operation, requiring significant amounts of I/O and CPU. Since Nomad uses bin-packing for job distributions our builders would sometimes launch up to 5 different sets of these containers at once, causing massive slow-downs before tests could even start running.

Build Experimentation

Setting up a development environment is key to debugging CI problems once found as it allows you to push your system to its limits while ensuring that none of your testing affects productivity in production. Coinbase maintains a development cluster for CircleCI that we use to test out new versions before pushing them out to production, but in order to investigate our options we turned the cluster into a smaller replica of our production instance, allowing us to effectively load test CircleCI builders. Keeping your development cluster as close as possible to production can help ensure any solutions you find are reflective of what can actually help in a real environment.

Once we had identified why our builds were encountering issues, and we’d set up an environment to run experiments in, we could start developing a solution. We repeatedly ran the same large builds that were causing the problems on our production cluster on different sizes and types of EC2 instances in order to figure out which was the most time and cost-effective options to use.

1* LIQcO G7e9U6N1aIMVt6Q

EC2 instance type comparison

While we previously had been using smaller numbers of large instances to run our builds it turns out the optimal setup for our cluster was actually a very large number of smaller instances (m5.larges in our case) — small enough that CircleCI would only ship one parallelized build container to each instance, preventing the build trampling issues that were the cause of the slow downs. A nice side effect of identifying the correct instance types was that it actually allowed us to reduce our server cost footprint significantly as we were able to size our cluster more closely to its usage.

Problem? Solved!

Applying your changes to a production environment is the final step. Determining whether the effects of the tuning worked can be done the same way the problems were identified — with the four golden signals.

After we had identified what worked best on our development cluster we quickly implemented the new builder sizing in production. The results? A 75% decrease in build time for our largest builds, significant cost savings due to the right-sizing of our cluster, and most important of all: happy developers!


Builds, before and after

This website may contain links to third-party websites or other content for information purposes only (“Third-Party Sites”). The Third-Party Sites are not under the control of Coinbase, Inc., and its affiliates (“Coinbase”), and Coinbase is not responsible for the content of any Third-Party Site, including without limitation any link contained in a Third-Party Site, or any changes or updates to a Third-Party Site. Coinbase is not responsible for webcasting or any other form of transmission received from any Third-Party Site. Coinbase is providing these links to you only as a convenience, and the inclusion of any link does not imply endorsement, approval or recommendation by Coinbase of the site or any association with its operators.

Unless otherwise noted, all images provided herein are by Coinbase.

Coinbase logo