Operating System: All
Platform: All

We have three options here:

1) The approach we have taken in Binding V2, where we have an explicit callback interface of two methods

2) Augment 1) with using BiConsumer<Throwable, T result>, allowing the use of lambdas, but having callers having to check failures first

3) Switch to using java.util.CompletionStage

Options 1&2 have the benefit of not being tied to a particular Future-like contract, hence are open to bridging:

RpcService s;
CompletableFuture<T> f = new CompletableFuture();

s.invokeRpc(..., (t, r) -> {
if (t != null)

else

}

With the same approach being possible for Guava, Scala and others. It also does not give semblance of having a synchronous get() interface.

CompletionStage is similarly flexible, with converters for Scala being readily available – but has one wrinkle, that of CompletionStage#toCompletableFuture(). That method exposes complete(), which is not appropriate and can actually be not implemented. This would leave us vulnerable to users assuming CompletionStages can be converted and that being broken when a different implementation is substituted.

I think I prefer #2 as the core API with a library of utilities bridging callbacks to popular Future-likes:

RpcService s:
ListenableFuture<T> f = GuavaRpcServiceAdaptor.invokeRpc(s, ...);

where GuavaRpcServiceAdaptor will provide briding to the future, which can then be composited as usual.

Since the bug is unassigned I'm currently assigning it to you.

Please assign to the relevant person.