Open Source Summit Europe 2024

Git monorepos that are under heavy load quickly become inefficient or even inaccessible. This impacts client operations of all types (e.g., git-upload-pack, git-receive-pack), especially during search-for-reuse phase.
Currently, the only countermeasures are a full GC or a geometric repacking on a regular basis or based on metrics.

As the repo history grows, running a full GC takes longer, is more expensive, and risks introducing additional workload at inopportune times (e.g., running GC during a burst of repository activity may bring nodes to a standstill).
In this talk, we will introduce an AI-driven approach to maintaining the performance of large Git monorepos that undergo heavy workloads.
The AI model will explore and learn different strategies, including partial repacking, bitmap regeneration, empty directory removals, and more, by evaluating its success using reinforcement learning.