This year has not been short of successful events for the Gerrit Community with the oversubscribed GerritMeets in Munich and the one in Sunnyvale earlier in the year, but the User Summit, co-located with the jj mini-con, really did beat every expectation. With over 70 industry experts and practitioners from companies like Google, Nvidia, Qualcomm, HRT, LibreOffice, and of course GerritForge, it felt like a great gathering of minds—particularly focused on how to scale and evolve code review for the AI era. It was of course a pleasure to be hosted again by Google in the Bay, 10 years after the last Google-hosted event.
What’s new in Gerrit 3.15
The event opened with the usual Keynote by our CEO, Gerrit Maintainer, Release Manager, and member of the ESC, Luca. Gerrit 3.15 promises to be an amazing release with “hard to believe” performance improvements, especially in the rendering speed of various user journeys on the UI. It also massively improves resource utilisation with the jump to Java 25 and its UseCompactObjectHeaders Java option support.
Then there are the various native AI features coming to Gerrit—in particular, the new icons to better identify which votes or comments were posted by an agent versus an actual person. After the talk, there was a lot of conversation around how to better give permissions to one’s agents that substantially differ from the actual person themselves! Expect to see more on this topic soon as the community decides which approach to take.
And how can we forget the live demo, where Luca switched LIVE the JGit repo to the reftable ref storage system, giving more than a few people in the room a minor heart attack, I’m sure.
Finally, a note on Gerrit 3.15’s contribution chart: we saw an increase in both core Gerrit commits as well as plugin contributions, and for the first time GerritForge contributed more than anyone else, including Google! Congrats to the team (and myself, I guess!).
On the topic of AI, another interesting talk was the one on task quotas by Nvidia’s Adithya! Adithya, a recent addition to the Maintainers group, covered some very interesting extensions they’ve added to core Gerrit (that complement the Quota plugin well, IMHO) that give incredible power to the Gerrit admin to limit how much work can happen in parallel for a given set of users. This alleviates a concern shared by many at the conference (and elsewhere) about how to limit agents from essentially DDoSing your Gerrit system. Adithya explored some clever solutions to this problem that I’m sure many will find useful.
Agents, Gerrit and Zuul — What a time to be alive
Still on the topic of agent interactions with Gerrit, Monty, one of the co-founders of OpenStack and Zuul and a huge personality in the DevOps world, gave a great presentation that hit the nail on the head in quite a few different ways. In particular, two points gave everyone some food for thought: treating agents like you’d treat an unknown committer in an open-source project (with the difference that you have more agency over the agent), and treating agents a bit like your secretary—trust them to do work for you, but don’t give them your credentials (“that’d be crazy!!”, as Monty put it).
His historical context about how Zuul came to be over 10 years ago to solve much of the same huge-scale challenges we’re trying to solve nowadays was extremely interesting and made me realise how AI isn’t really reinventing how we deliver software, but rather requires further scaling of already known and well-proven methodologies.
Tracking metadata in Git
It was also great to have Patrick Hiesel, Staff Software Engineer at Google and longtime Gerrit maintainer, back at a Gerrit event and back in the version control space—in particular, Google’s new VCS, jj. He gave some great hindsight into how Gerrit’s NoteDB came to be, how tracking metadata as part of the Git history makes a huge amount of sense, and how it allows that metadata to move around right along with the code.
Cloud-JJ
Daniel, Engineering Manager for the jj team at Google, presented cloud-jj, which was also extremely interesting and something that I’ll be keeping an eye on! There is a huge amount of buzz around jj as a tool and in the jj community about its workflow improvements over Git—it’s really exciting to see. Daniel’s team is working on providing an open-source commit cloud that allows companies to operate at Google scale. It’s great to see Google still committed to the open-source ecosystem with great tools and thought leadership, and I’m very excited to see how the project evolves over the coming months and years!
JGit Cassandra
Luca, who felt a bit lazy this time round and only gave three presentations, also talked about the resurgence of JGit-cassandra. Shawn Pearce provided a first trivial implementation 16 years ago which I believe we can all agree was definitely ahead of its time. Luca instead decided to review a prototype he and Haithem worked on at the Gerrit Hackathon in Mountain View over 10 years ago and that can be found at [1]. JGit has had the ability to inject different object storages since its very inception, and that’s exactly the way Google has been using it at the huge scale they operate at all these years; however, outside of Google, concrete implementations of other backends have never really made it to production, mainly because of a lack of need for it.
Of course, the advent of AI changed all of this, and with companies trying to grapple with the huge loads being put on their SCM systems, it was only natural that such solutions would come back alive! Luca showcased how the JGit Cassandra backend, despite still being a prototype version, could resolve many of the issues faced by modern installations—providing huge scale and delegating replication of the Git data to an external system, potentially simplifying Gerrit installations around the world. If you’re interested in seeing it in action on your system, you should hit him up!
Large object promisor support in Git
It was an honour having Christian Couder from GitLab present at the User Summit this year. This was particularly important for me as we love to have people from other forges participate at the Gerrit User Summit—last year we had Patrick Steinhardt, lead of GitLab’s Git team, so hopefully this is the start of a good tradition! Christian presented the latest iteration of supporting large objects in Git with large object promisors, which helps resolve a longstanding pain point in the Git world: Git LFS! Thanks to these promisors, we can more efficiently manage how large objects are handled on the Git side, removing some of the friction points inherent to Git LFS. I’m looking forward to seeing how these evolve in the years to come!
Sustaining Git Performance at High Speed
Lastly—and I’ll cut it here as I’ve gone on long enough already—Shane’s presentation was, as usual, one of the most engaging of the event! Shane is an Associate Professor at the University of Waterloo and one of the most important researchers in the field of software development worldwide. He’s been collaborating with the GerritForge team for quite a few years now, and this time he presented the work we’ve been doing on ensuring Git operations stay performant even under the heaviest of AI loads!
I found particularly interesting the graph showing how, when using GHS [optional: spell out acronym here], all the other parts that comprise a Git Fetch/Clone operation stay incredibly constant and effectively the entirety of the clone time is spent writing the objects over the network! In a test that we did in a real production environment with one of our customers, we were able to obtain over 30x performance improvements, vastly exceeding even the numbers that Anthropic claim broke Git for them! Using AI agents to optimise Git usage for other agents—I love the symbolism here!
There are honestly so many really interesting talks I could talk about, but this is already far too long, like Tony’s talk on the sticky owners and what ownership means in the age of AI, Ponch’s talk on how to improve Gerrit performances by over 100% with a set of relatively minor tweaks and Kaushik’s improvement to the replication plugin!
All in all, it was a great user summit, one that I’ll remember for a while and that certified how code storage and review is definitely evolving, but it’s also here to stay. The classic SCMs are still the defaults of course, but everyone is re-evaluating their options for the AI era and Gerrit is definitely getting its fair share of attention.