Kubernetes, 2015–2026: a decade of the world's orchestrator, measured statement by statement

The measurement

Take the Kubernetes repository at the start and end of each year, 2015 to mid-2026. Classify every changed logical statement as added, deleted, or edited in place, whole tree, one command per window. REWORK is the share of all churn spent editing statements that already existed — the number that separates building from maintaining. (Definitions: code churn, explained properly.)

Three eras in one table

The explosion (2015–2017). Post-1.0 Kubernetes churned over a million statements a year, adding 6.5 statements for every one deleted in 2015 — and editing almost nothing in place: REWORK 0.5%, one statement in two hundred. Rapid human growth looks like this; it is also why a low rework number alone never proves AI involvement.

The great shedding (2017). The only year deletions beat additions (0.8 : 1) — nearly 600,000 statements removed at peak churn. That is the numeric shape of a project cutting weight while running: vendored-dependency and codebase restructuring at full speed.

Maturity (2018–2026). Churn settles near a quarter-million statements a year and the editing share climbs steadily: 7% in 2018, hovering 10–11% since 2020, and 2026 H1 is the highest in the project's history at 11.6% — one churned statement in nine is now a repair of existing code. For scale: mega-projects (Chromium, the Linux kernel, WebKit) run 5–10%, and established mid-size projects like curl and PostgreSQL sit near 17%. Kubernetes has spent a decade travelling from the hypergrowth floor toward the maintenance band — the numeric signature of a project growing up.

The question everyone will ask

AI-assisted development shows up in churn shape as the opposite movement — the editing share collapsing toward zero, because generated code is replaced rather than repaired (measured in the REP_CHURN paper: an intensively agent-built codebase reworks 0.19%). Kubernetes shows no such bend: through the AI-tooling era its rework share has kept rising. Whatever role assistants play in individual contributions, the project's change process has, so far, only become more repair-dominant. The 2025 dip to 7.6% and rebound is worth watching, not over-reading — one year is weather, the decade is climate.

Reproduce any row

Each window is two public commits and one command. The snapshot pairs (old → new, default branch):

2015  82a72c805d83 → 708fda2683ee     2021  d378fca35a5e → c5d8354e0e65
2016  708fda2683ee → f5d9c430e916     2022  c5d8354e0e65 → 886c4b0cf262
2017  f5d9c430e916 → da9a4d5da9ec     2023  886c4b0cf262 → 3be9a8cc7326
2018  da9a4d5da9ec → de02d00989cb     2024  3be9a8cc7326 → d8d369128012
2019  de02d00989cb → 6b5cf90ecd2f     2025  d8d369128012 → 73307ba402df
2020  6b5cf90ecd2f → d378fca35a5e     2026  73307ba402df → b8a20977dbfc

These figures are drawn from the same 136-codebase, 16-year measurement campaign reported in paper D (1,437 repository-year measurements, 393 million statements of churn). Method, edge cases and instrument validation: the papers.