We recently put together some automated testing which evaluates the performance of various software forges without bias1: the Software Forge Performance Index. The goal is to push our community of software forges to strive for higher performance, to tighten your workflow loop and to better service communities without access to high-speed internet or newer computer hardware. I also included results for accessibility testing, hoping that we can push each other to improve on that front as well.
These tests show that SourceHut has room for improvement. For one, some important features are conspicuously absent, like a git blame view. GitHub beats out our performance at rendering commits and tickets, and on several pages GitHub and Bitbucket both have better accessibility scores than the equivalent sr.ht page. But the overall results are clear: SourceHut is the best performing code forge, and by a wide margin. The data also shows that our approach delivers the best reliability: our friends at GitHub and GitLab have struggled with outages 15 of the last 90 days and 18 of the last 90 days, respectively. SourceHut has had two outages in 2020, both of which were planned and announced a week in advance, and only one of which affected more than one subsystem.
This culture of simplicity extends to our backend servers, too, which is how we deliver on reliability. We have no Kubernetes here, no Docker, no CloudFormation, no monorepos. SourceHut has 10 dedicated servers and about 30 virtual machines, all of which were provisioned by hand. We prefer Alpine Linux because it’s lightweight and simple — our average production server has only 20 to 30 packages installed, which means no moving parts that we don’t understand. Our application design helps as well: many pages are designed to make only one or two round-trips to the database, the queries for which can be carefully optimized with joins and indices built to-purpose.
Our distributed “miniservice” application design also allows any of our services to fail independently without causing a total outage. And, because we embrace email for collaboration, our entire service could be offline without affecting your workflow — the email system has 50 years of battle-tested fault tolerance and redelivery mechanisms, you can keep working normally and our servers automatically get caught up when they’re back online. We achieve reliability on several levels: simple servers which don’t break unexpectedly, monitoring which alerts us to problems before they’re serious, and a design which works even if we do experience an outage.
Microsoft bought GitHub for 7 billion dollars. GitLab is valued at 3 billion dollars. Together, the two companies have over 2,000 employees. SourceHut made 4 thousand dollars in Q1 with two employees and an intern. How do we deliver on this level of reliability and performance compared to these giants? The answer is a fundamental difference in our approach and engineering ethos. I hope that forgeperf.org is a good start for pushing our industry to mount a serious attack on the eternal problems of high performance and reliability once more.
If you want to learn more about our approach to reliability, check out our public operations manual, join our #sr.ht.ops channel on irc.freenode.net to watch our operations in real-time, or query the raw monitoring data yourself from our public Prometheus instance.
I am confident that these results are accurate and without bias towards SourceHut. We run only standard benchmarks, maintained by a neutral third-party (Google Lighthouse), and the results are corrected for latency and bandwidth differences. This was verified by having the same test suite run in several locations across America and Europe, showing consistent results even with more generous routing to our competitors. You can run it, too, and verify the results for yourself — it’s pretty easy to set up the software on Linux and it only takes about an hour to produce the full report. I’m also sure that the results are not cherry-picking pages for which SourceHut has an advantage — it’s easy to contribute new pages if you think we missed a spot. ↩︎