this for ~10 years. • Everything from Search to GCE VMs run in containers. • We launch 2B containers every week • Deep understanding of what it takes to launch, run and manage containers at scale. • We have lots of experience/scars • Core contributor to linux cgroups and Docker libcontainer. • In some ways, we invented the core technology for containers on Linux • Containers are a keystone technology for cluster management systems.
Packaging and portability • Easy to capture and move systems between environments • Density • Use every part of the "animal." Drive utilization up. Both in VM clouds and on bare metal. • Security • !! Note: We don't think that container technology is there yet as a hard security boundary. Surface area to secure is too big. But, when combined with other mitigations, it is doable.
utilization and efficiency • Larger pool of workloads and compute allows for more complementary workloads to be scheduled. • Easy platform for horizontal scaling • Running, tracking, managing arrays processes is easy. • Self healing • Workloads can be dynamically rescheduled in the face of individual machine failure. • Allows for specialization of ops roles. Cluster ops vs. individual app ops. • Internal Example: GMail SRE rarely has to talk to Borg SRE • Insights into performance and metrics at the app server level vs. machine level. • Pairs well with microservices • Reducing the management burden allows you to manage more things • Keep deployed/managed items small enough for efficient teams.
• Focused on portability -- run anywhere • Launched in June • Active developer community • 100+ contributors, 5100+ GitHub stars, active IRC channel • Open roadmap: roadmap.md on GitHub • Built on Docker
moving parts are in place • Road to v1: • Improve usability • Introspectability • Production reliability • Cluster upgrade • Finalize API with deprecation policy • Pluggable auth model
machine failure • Groups of containers that must be co-scheduled: Pods • Native support for sets of containers • Labels on on containers with query/selection syntax • Horizontal scaling based on template: ReplicationController • Critical for visualization, logging and monitoring • Services: connections between containers and beyond • Find containers based on name/labels • Easily communicate with a set of containers • Import/export services from cluster
is your "cluster ops" Now: • One click/API call to spin up cluster • Builds on Kubernetes API/tools for using the cluster. Coming: • Automatic scaling of cluster pool. • Integration with GCP around storage, logging, monitoring, LB, auth. • Regional/Global solutions.
is a building block for compute workload management Missing parts: • Application description and deployments, upgrade • HTTP router • "Just bring the code" There will be many solutions to these problems both on GCP and beyond. App Engine is will be built on GKE -- Navneet to detail later.