Fun facts

A grab-bag of things you'll only learn by spending time in this codebase.

• The first commit is from 2011-10-09. The repo has accumulated ~528,000 commits since.
• About 935 distinct authors committed in the last 90 days, including release bots that account for ~5% of commit volume.
• db/structure.sql weighs in around 24 MB. It's autogenerated from migrations.
• There are ~591 pre-deploy migrations and ~946 post-deploy migrations. The split exists so deploys can be zero-downtime.
• ~463 active feature flags are defined under config/feature_flags/. Each must declare a default, a feature category, and a milestone.
• There are ~155 feature categories and ~75 bounded contexts — categories are operational metadata, contexts are namespaces in code.
• A "ghost user" exists at the database level to be the author/assignee of records whose creator was deleted. There's also a "support bot" and a "security bot" — about 15 user types in total.
• Some integrations ship as full STI subclasses of Integration — there are ~70 of them (Slack, Mattermost, Jira, Telegram, etc.).
• GitLab Duo has its own catalog of LLM features: prompt families live in ee/lib/gitlab/llm/templates/ and agent flows in ee/lib/gitlab/duo/workflows/. The set is large enough that the abuse harness is itself a sub-feature.
• The monorepo includes 26 in-tree gems under gems/, loaded via path: in the Gemfile so they ship with the monolith.
• Hashed storage means a project's on-disk path is a sha256 of its ID, so renaming a group is zero-cost on disk. The legacy "named storage" path is gone.
• The Workhorse → Gitaly → Postgres path is the hottest in production. A Git push goes Workhorse → Rails (for the pre-receive check) → Gitaly → object pool, all instrumented end-to-end via correlation_id.
• Sidekiq's "dead set" is monitored as carefully as the queue itself. Repeated failures end up here, and a feature category with growing dead-set entries is a paged alert.
• The config/vue3migration/ directory tracks every component's Vue 3 readiness. The Vue 2 → Vue 3 migration is one component at a time.
• Rails console in production is gated by audit and recorded — it's not a free-for-all.
• The CI catalog (Ci::Catalog) is a feature-store-of-features: project owners publish reusable CI templates, others consume them by version.
• EE prepend means there's a parallel set of files under ee/ that prepends behavior into FOSS classes. About half the controllers, models, and services have an EE counterpart.
• The repo has its own markdown-renderer-as-pipeline (Banzai), and changes to it run a separate CI verification.
• Quick actions like /assign @alice work in any comment box. There are ~30 of them, defined declaratively.
• The CI configuration for this repo is itself huge — .gitlab-ci.yml is mostly a manifest of includes; the actual rules are spread across .gitlab/ci/ and rule generators in tooling/.
• The Danger rules in danger/ enforce conventions like "PRs touching db/migrate/ must include a database review" and "MRs adding feature flags must declare a rollout issue".
• Lefthook runs as the local git hook, gated by lefthook.yml. New contributors are gently nudged onto it.
• The codebase contains 3 different "feature flag" concepts: instance-level Flipper toggles, per-licensed-feature flags, and frontend client-side flags. They've been unified mostly but not entirely.
• Reading ee/app/models/ee/project.rb is a small adventure. It prepends to Project and adds a few thousand lines of EE-only behavior. The discipline of using extend ::Gitlab::Utils::Override on every method is what keeps refactors safe.
• The Postgres extension layer uses pg_trgm, btree_gin, plpgsql, and (optionally) pgvector for vector search.
• Object pools for forks are stored as their own Gitaly repos. A fork "alternates" off the pool, deduplicating gigabytes for popular projects.
• The CommonMark renderer is custom-built atop commonmarker to integrate Banzai filters; the AsciiDoc renderer goes through Asciidoctor with a sandboxed extension set.