Dualstack Backends

Overview

In cases where an endpoint (i.e., a gRPC server) has both an IPv4 and IPv6 address, it is useful for clients to know both of those addresses, so that they can reach the endpoint on either of them. Because only one of the two protocols may actually be routable between a given client and endpoint, clients need to try both addresses to see which one will work. To minimize connection attempt latency, it is also desirable for clients to attempt to connect to multiple addresses in parallel, which is done using the “Happy Eyeballs” algorithm described in RFC-8305.

In gRPC, it is particularly important for the Load Balancing Policy to understand which addresses are associated with the same endpoint as opposed to being different endpoints, so that the client does not send twice as much load to a endpoint with two addresses. To that end, the Name Resolver can return multiple addresses per endpoint (where an endpoint is a single endpoint), and that information is communicated to the Load Balancing Policy.

Note that the only name resolver implementation that gRPC provides is a DNS resolver, which does not support dualstack backends. This is because DNS inherently returns a flat list of addresses, with no way to indicate whether or not multiple addresses actually point to the same endpoint.

There are two ways to use dualstack endpoints in gRPC:

1. Write your own custom name resolver that returns multiple addresses per endpoint. For more information, see Custom Name Resolution.
2. Use an xDS control plane that can communicate multiple addresses per endpoint to the client. For more information, consult the documentation for your xDS control plane.

For more information on gRPC dualstack backend support, see gRFC A61.

Custom Resolver Dualstack Examples