Uploaded November 2026 | Updated November 2026, 3 days ago
SREcon26 Europe/Middle East/Africa - When Partitioning Beats Sharing: Ephemeral MicroVMs for CI at Scale
Emir Beganović and Luka Stout, Booking.com
Your CI hosts show plenty of idle CPU, yet a job you gave 4 cores behaves like it has 2. Where did they go?
At Booking.com we hit exactly this running ~10 million CI jobs a month on a container-based fleet. The bottleneck wasn't hardware but rather a combination of CFS throttling on high thread-count servers. This talk walks through what we did instead: replacing shared-kernel containers with fully isolated ephemeral KVM microVMs on the same hardware. This bought us ~3× more effective CI capacity, better p95 latency, and a hard isolation boundary, while preserving exact developer experience with zero CI syntax changes for users.
Using four laws as a guide: Little's Law, Amdahl's, Universal Scalability, and Hyrum's, you'll leave with a concrete way to tell when a shared scheduler has stopped fitting your workload, and why partitioning can beat sharing for CI workloads.
View the full SREcon26 Europe/Middle East/Africa program at usenix.org/conference/srecon26emea/program
SREcon26 Europe/Middle East/Africa - When Partitioning Beats Sharing: Ephemeral MicroVMs for CI at Scale
Emir Beganović and Luka Stout, Booking.com
Your CI hosts show plenty of idle CPU, yet a job you gave 4 cores behaves like it has 2. Where did they go?
At Booking.com we hit exactly this running ~10 million CI jobs a month on a container-based fleet. The bottleneck wasn't hardware but rather a combination of CFS throttling on high thread-count servers. This talk walks through what we did instead: replacing shared-kernel containers with fully isolated ephemeral KVM microVMs on the same hardware. This bought us ~3× more effective CI capacity, better p95 latency, and a hard isolation boundary, while preserving exact developer experience with zero CI syntax changes for users.
Using four laws as a guide: Little's Law, Amdahl's, Universal Scalability, and Hyrum's, you'll leave with a concrete way to tell when a shared scheduler has stopped fitting your workload, and why partitioning can beat sharing for CI workloads.
View the full SREcon26 Europe/Middle East/Africa program at usenix.org/conference/srecon26emea/program










