Milvus

Milvus is the open source vector database that backs semantic search, recommendation systems, image search, and the retrieval layer for modern AI applications. The project sits inside the LF AI & Data Foundation and runs on contributions from around 30 people, mostly engineers at Zilliz. Five of them have full maintainer access.

The merge policy is non-trivial. Because Milvus is an LF AI & Data project, every PR has to pass a DCO check before it's allowed to merge. On top of that, the team runs two separate test layers: GitHub Actions for the lightweight checks, and an internal Jenkins for the heavier ones. The result is a PR with status updates landing from multiple places at different times, which historically left both contributors and maintainers having to chase down whether a PR was actually ready to merge.

Mergify's job at Milvus is to turn that mess of status updates into a single decision. The team relies heavily on labels: a dco-passed label gets added when DCO clears, and a ci-passed label flips when the PR's relevant tests turn green. Combined with a maintainer's LGTM, those labels tell Mergify whether the PR is actually mergeable. When everything lines up, Mergify auto-merges to master without anyone touching a button.

Because the team has multiple PR types, with different test requirements for code versus docs, the labeling rules differ accordingly. Doc-only PRs skip the heavy Jenkins tests entirely, while code changes go through the full suite. Mergify handles the routing based on file paths and labels, which means contributors don't need to know the rules to file a correct PR.

Across the past 12 months, Mergify has handled 1,742 PRs across the milvus-io organization. The milvus core repo alone accounts for about 1,360 of those, with the rest spread across 11 other repositories like pymilvus and the language SDKs. Five maintainers couldn't review and merge that volume manually if they tried, and the 105 distinct contributors over the same period would lose patience fast if they had to wait for one of those five to be free.

For Xiangyu, the win is the asynchronous flow: contributors push a change, tests run automatically, maintainers leave reviews when they have time, and Mergify merges once everything is green. The "extra team member" framing is his, and after a year of 1,700+ merges that description holds up.