Adam Litke

Kubernetes clusters increasingly host applications that manipulate data in persistent storage. Preserving that data is important but how does it get there in the first place? Unless it can generate its own data, the application should expect the volume to be populated prior to initialization. This may seem like a trivial problem with many simple solutions but we can build on kubernetes design principles to produce a superior solution. In this session you will discover a declarative PV population pattern, observe some concrete examples how this can be applied to different sources of data, learn about some common problems in this area and how kubernetes can help to solve them. After the session you will be ready to contribute your own populator to the kubernetes community.

As kubernetes clusters take on more workloads, the need for persistent storage has only increased. Users want to connect more storage devices and use them efficiently. Cinder has been solving this use case in Openstack for a long time and can be integrated with any kubernetes cluster regardless of its cloud provider. By deploying a standalone cinder service and a new external provisioner to a cluster, pods can access almost any volume that cinder is able to provision. Complex operations such as cloning, snapshots, and migration become services of the cluster. This presentation will describe and demonstrate a cinder external provisioner, enumerate its advantages, discuss challenges especially around deployment, and propose collaboration opportunities between Openstack and kubernetes.

In oVirt datacenter virtualization environments, a manager directs hosts to initiate operations to shared storage. These operations create or remove volumes, copy data between volumes, create or merge snapshots, and various other actions related to virtual machine storage. For efficiency and balance these operations should be distributed across multiple hosts and run in parallel when possible. Maintaining reliability under real world conditions requires careful management and resilient algorithms. This talk will introduce some of the problems that can arise including: dropped communications, scheduling conflicts, and host or storage array failure. Next, a solution to these problems using shared storage locking, atomic operations, volume generations, and forensic analysis of the storage will be presented. Through step by step examples, the audience will understand how the proposed solution can solve all of the outlined problems.

In modern virtualization platforms virtual machine disks are stored in various kinds of shared repositories which are accessed by thousands of hypervisors simultaneously. Changes are constant. How can you provide high performance access to these resources while ensuring a consistent view of the world across all systems? When bad things happen, how do you restore the integrity of the data?

This talk will share an approach being taken by the oVirt virtualization platform to address these challenging storage problems. Complex storage tasks are broken down into simpler metadata and datapath operations. A set of rules for handling each type will be established. Through several examples, the audience will learn how a combination of SAN-based locking, atomic commitment, and garbage collection can help to achieve a high performance, fault-tolerant storage architecture.