Introduction to multi-cluster mode

Multi-cluster mode is a set of features allowing management and interoperations of multiple StorPool clusters located in the same data-center, which allows for much higher scalability and better reliability in large deployments. Instead of creating one big storage cluster with many hosts, the multi-clustering approach adds the flexibility to have multiple purpose-built clusters, each using fewer hosts. The set of features is built as an extension to multi-site mode, and each location could now have multiple clusters. For a different analogy, a multi-cluster would relate to multi-site similarly as a single datacenter to a city.

This document provides and overview of the multi-cluster mode. For more information about using this mode, see 17. Multi-site and multi-cluster.

Basic concepts

The first concept is that of an exported volume. Just like snapshots can be currently exported to another cluster through the bridge, now volumes can be exported too for all nearby clusters. An exported volume can be attached to any StorPool client in a target cluster, and all read and write operations for that volume will be forwarded via the storpool_bridge service to the cluster where the volume actually resides. For more information, see 12.10.2. Listing exported volumes.

The second concept is the ability to move volumes and their snapshots between sub-clusters. The volume is moved by snapshotting it in the sub-cluster where it currently resides - that is, the source sub-cluster, exporting that snapshot and instructing the destination sub-cluster to re-create the volume with that same snapshot as parent. This all happens transparently with a single command. While the volume’s data is being transferred, all read requests targeting data not yet moved to the destination sub-cluster are forwarded through the bridge service to the source sub-cluster. When moving a volume there is an option to also export it back to the source sub-cluster where it came from and updating the attachments there, so that if the volume is attached to a running VM it will not notice the move. For more information, see 17.13. Volume and snapshot move.

Both features have minimal impact on the user input and output (IO) operations. An exported volume adds 2-4 times the storpool_bridge’s link’s RTT to the IO operations (which is usually tens of microseconds). So for everything - except for some extremely demanding workloads - a volume move would go practically unnoticed.

Use cases

A target use case for volume move is the live migration of a virtual machine between hosts that are in different sub-clusters in the multi-cluster. The move to the destination sub-cluster involves also exporting the volume back and leaving the existing attachment on the host in the source cluster intact, almost like a normal live migration between hosts in the same cluster. If the migration fails for some reason and will not be retried, then the volume can be returned back to the original cluster. Its data both in the volume and its snapshots will transparently move back to the source cluster.

Implementation

The functionality is implemented only in extensions to the API - there is no change in what all current API calls do.

There are two extensions to allow using these features explained below.

The first one is the ability to execute API commands in another sub-cluster in a multi-cluster setup. For an example, see 17.13. Volume and snapshot move.

The second extension is the multi-cluster mode of operation. It is enabled by specifying -M as a parameter to the storpool CLI command or executing multiCluster on in interactive mode (example). This mimics the single cluster operations in a multi-cluster environment without placing additional burden on the integration.

For example, a multi-cluster attach volume command will first check if a volume is currently present in the local sub-cluster. If it is, the command proceeds as a normal non-multi-cluster attach operation. If the volume is not present in this sub-cluster, the API will search for the volume in all connected sub-clusters that are part of the multi-cluster, and if it is found will issue a volume move to the local cluster (exporting it back to the source cluster if the volume is attached there) and will then attach the volume locally.

The same is in effect for all other volume and snapshot related commands. The idea is that the multi-cluster mode can be used as a drop-in replacement of the non multi-cluster mode, having no effect for local volumes. The only difference is that an attach operation must be targeted to the cluster with the target client in it. This is the main reason for the first extensions, so that the integration doesn’t need to connect to the different API endpoints in each cluster to do its job.

Volume naming

When using the multi-cluster mode you should be careful with volume naming.

The whole multi-cluster acts as a single namespace for volume names, however there is no StorPool enforced uniqueness constraint. We thought long and hard on this decision, but decided at the end against enforcing it, as there is no way to guarantee the uniqueness in cases of connectivity or other disruption in communication between clusters without blocking local cluster operations. We believe that having the ability to work with a single cluster in the multi-cluster in all cases is far more important.

This decision however places the burden of keeping the volume names unique on the integration. There are few ways this can be handled.

The first option is to never reuse a volume name, even when retrying a failed operation, or operation that has an unknown status (for example, timeout). So if the integration can guarantee this, the uniqueness constraint is satisfied.

Another option is to use another feature - instead of providing a name for the volume create operation, the StorPool-assigned unique ID that StorPool already maintains for each volume in existence - the globalId in the JSON output. So the feature is to be able to use the globalId as identifier for a volume instead of a name and the ability to create a volume without specifying a name and having StorPool return the unique identifier. This mode however is a significant change to the way an integration works. For our integrations we have chosen to move to this mode as they are plugged in a complex management systems that we can’t guarantee will never reuse a volume name, for example during a retry of a volume create operation. This also comes with a new directory - /dev/storpool-byid holding symlinks to the /dev/sp-X devices but instead of using the volume names as in /dev/storpool, using the globalId.

Violating the uniqueness constraint will eventually result in data corruption for a given volume, so handling it is a must before considering the multi-cluster features. A simple example of how data corruption will occur if there are two volumes with the same name. Let’s assume the volume name in question is test and we have one volume test in the source sub-cluster A with a VM working with it, and another volume test in the target sub-cluster B. If trying to migrate the VM from cluster A to cluster B, the multi-cluster volume attach command will find the wrong volume test already existing in the target sub-cluster B and will simply attach it to the client. So when the VM is migrated the data in the volume test will be totally different, resulting in data corruption and VM crash.

There are two more considerations related to naming. They have lower impact:

Volume move will recreate the volume in the target cluster with a template named as the one in the source cluster. So every cluster in a multi-cluster must have all the templates created with the same names.
Different clusters in the multi-cluster should have the same names in each of the clusters, so that the remote cluster commands can be executed from any API endpoint with the same effect. Just for completeness in each cluster the local cluster can also be (and should be) named, so a storpool cluster cmd NAME will work for it too.