OrchardCore Clusters

[jtkech]

Fixes #13636

Distributes requests across tenant clusters by using the Microsoft Yarp.ReverseProxy.

Work in progress but some first info.

• We first use the Yarp Configuration allowing to define Routes and Clusters with many options. Each Route is tied to a Cluster composed of Destination(s) on which load balancing can be applied ...

• We only need one catch-all RouteTemplate and multiple Clusters on which we can configure a custom SlotRange[min, max] property (up to 16384 slots).

• Each Tenant has an unique slot hash, so an unique Slot and then belongs to the Cluster having the slot in its SlotRange, the Cluster having multiple Destination(s). Note: We could have used a Cluster having Nodes but we follow the Yarp Config having a Clusters list of Cluster type.

• The same application can run as a proxy or behind it (we check the headers), the advantage with our distributed services, is when as a proxy we are still aware of all tenants data. So on a request we can use the same RunningShellTable to know the Tenant, then select the right Cluster based on the Tenant slot hash (in a custom middleware), and let Yarp select one of its Destination(s).

• To compute a Tenant slot hash we use the CRC-16/XMODEM algorithm (as Redis use for clustering keys) applied on the new TenantId property, it allows to automatically spread out new tenants on the slots and then on the configured Clusters. This knowing that the CRC-16 is fast to compute and always return the same number for the same TenantId, so a tenant stays on the same Cluster.

• The distribution is not perfect with few tenants but gets better and better as the number increases.

TODO: Also coupled to a simple feature allowing to release a Tenant if not requested since a given time.

[jtkech]

@sebastienros Just for info

• I removed the proxy Hosts options as discussed in the meeting.

• So, for now we can add Tenants dynamically and we would also want to add Clusters dynamically.

  Because a Yarp Cluster as many properties, as we already have a RouteTemplate, we could also have a ClusterTemplate config on which we could tweak the Destinations. Hmm, but the Cluster SlotRange would also need to be dynamic, for example to evenly distribute Tenants.

• On Azure a VM or instance level Public IP addresses (PIP) can be configured and could be used in our current configuration. But when instances are created dynamically, how an OC running instance would know that its IP is reachable by our reverse Proxy, maybe a config defining a range of IPs.

• I also thought about the case that a Proxy targets an instance that should also act as a Proxy, maybe too much otherwise the header check to prevent loop would not be good. We would need a config to know which instances should act as a Proxy, maybe by reusing the proxy Hosts config option.

Already too much questions, I will think about it ;)

When I will have time, as a first step I will try the Yarp in memory configuration that allow to update dynamically the Yarp config by passing Routes and Clusters, maybe make it easier by exposing a kind of Destinations provider interface, so that when the list of Destinations change we update the config.

[github-actions]

This pull request has merge conflicts. Please resolve those before requesting a review.

[hishamco]

@Piedone is this related to Lombiq? If yes can anyone else continue on this, otherwise what's the progress?

[Piedone]

Nope, and we haven't worked on this with JT. It would be useful though for multi-node hosting.

[github-actions]

It seems that this pull request didn't really move for quite a while. Is this something you'd like to revisit any time soon or should we close? Please comment if you'd like to pick it up and remove the "stale" label.

[github-actions]

Closing this pull request because it has been stale for very long. If you think this is still relevant, feel free to reopen it.