170k Jobs a Day on GKE: Scaling Mercari's CI Platform - and What's Next for AI-Native Development

Mercari's self-hosted CI platform on GKE now completes roughly 170,000 GitHub Actions jobs on an average day — about twice as many as a year ago. This talk covers the engineering that absorbed that growth — machine-family-agnostic capacity with GKE ComputeClass, pod-density optimization with ScaleOps, and upstream contributions to Actions Runner Controller — and where the platform goes next as AI agents become first-class users of CI

Transcript

1. 170k Jobs a Day on GKE: Scaling Mercari's CI Platform - and What's Next for AI-Native Development KubeAuto Day Japan 2026 1
2. Junya Okabe ・Software Engineer at Mercari, Inc. ・CI/CD, Developer eXperience ・Kubernetes contributor / maintainer (SIG-Docs) GitHub: @Okabe-Junya LinkedIn: @junya-okabe X: @junya__okabe ・CNCF Ambassador 2
3. Mercari: One of the Biggest Marketplaces in Japan Accumulated listings 4 billion+ MAU Approx. GMV 23 million 1.1 trillion JPY+ 1/5 of people in Japan are using Mercari 3
4. Mercari's CI Platform and Its Scale 4
5. CI jobs completed per day 5
6. Mercari's CI Platform at a Glance 170k+ CI Jobs/day on GKE nearly 2x YoY Weekday avg: 230k+ 6
7. Architecture GitHub Actions self-hosted runner platform ・GKE ・actions-runner-controller ・in-house k8s CRDs ・Middleware ・Go module proxy ・In-cluster cache ・Security tools ・etc … 7
8. Our Scale At peak, on a single cluster 300+ nodes ~2k concurrent CI jobs 11k+ vCPUs And in a single day 200k+ ephemeral runner pods 120k+ vCPU-hours 8
9. Scaling Challenges 01 Capacity 02 Cost vs. performance vs. resource efficiency 03 Limits of actions-runner-controller 9
10. Tackling Capacity Relying on one machine type is risky at this scale ・Multi-region: not there yet ・Multi-machine-type: GKE ComputeClass fallback 10
11. Balancing Cost, Performance, and Efficiency ・Rightsizing with ScaleOps ・Optimal requests from per-job usage history ・Applied via admission webhook at pod creation 11
12. Contributing Back to actions-runner-controller Scaling pains, upstream gains ・Rate limiter option to raise controller throughput — #4451 ・pprof support to profile the controller — #4449 …and more OSS contributions!! Cf. PRs by @Okabe-Junya 12
13. CI at the Core of AI-Native Development 13
14. Becoming an AI-Native Company careers.mercari.com (Career Blog) engineering.mercari.com (Engineering Blog) 14
15. CI at the Core of AI Enablement ・CI was becoming a bottleneck in our SDLC ・Demand nearly doubled — AI keeps accelerating it ・We must scale 2x in a year — without degrading performance 15
16. CI as the Guardrail for AI-Native Development Guardrails for shipping changes quickly and safely: Fast, reliable unit & e2e tests Regression tests Security Scanning ・・・ CI as the Platform for Guardrails Scale the platform → improve the core component accelerating AI-native development CI is not just infrastructure: it enables UX and AX 16
17. What is a Great CI — for Humans and for AI A platform you can always trust — fast, actionable feedback, and code quality guaranteed before deploy The operations we're starting: ・SLOs on platform reliability ・Measuring change failure rate (DORA metrics) ・Measuring time to actionable feedback 17
18. Plenty of Challenges Ahead ・CI, build & test remain top DX pain points across the entire company — we need “user enablement” ・Scalability: many open problems ・disk I/O optimization ・flaky components ・networking ・etc … ・Growing and maturing the internal ecosystem 18
19. Wrap up 19
20. Key Takeaways 01 No silver bullet for scalability — invest from every angle 02 Not just supporting AI — a platform that enables AI development 03 Define key metrics from the user and AI experience 20
21. We are Hiring! Check here to view the entire Recruitment information. 21