[stickfigure]

> It shouldn't look like any other function call.

I'm sorry, This is the stupidest thing I see commonly repeated in public discourse about software.

Every single distributed application I've ever worked on in my 30-year career (including working for multiple companies you've heard of) wrapped remote calls in something that looks like a normal function call. It doesn't matter if your low-level RPC stub throws RemoteException or returns RemoteError, somewhere up the call stack someone has wrapped this into a simple method that looks like this:

    doSomethingUseful();

In real life, you make a call to a function and you live with the consequences. If you're lucky, the docs let you know the performance and failure characteristics. If not, you make some assumptions. When those assumptions are wrong, you spend some time debugging and profiling.

Adding a bunch of syntactic noise to the callsite doesn't help. The first thing any competent programmer will do is abstract your noise away in convenience methods. Because 99% of the time, it doesn't matter that your call is remote.

Take a look at your own codebase that makes client REST calls to some other service - you may hand-wire a bunch of http calls, but somewhere up the stack there's a function that hides the http mess. Everything below that function is accidental complexity.

[naasking]

> Every single distributed application I've ever worked on in my 30-year career (including working for multiple companies you've heard of) wrapped remote calls in something that looks like a normal function call.

This isn't a problem as long as the returned value provides the right failure semantics (like futures). The problem with trying to encapsulate the network is that deep call chains lead to cascading failures for problems that are common in networks (partitions, latency, etc.). These failure modes also lead to more pervasive use of timeouts in deep call chains, which then introduces non-determinism, which itself makes issues impossible to debug.

[stickfigure]

This is also nonsense. 99% of the time these failure modes are irrelevant. A remote call fails, the error propagates up the call stack, and someone gets an error message. Just like any of the thousands of other things that can produce errors in complex systems.

In the rare case you need to harden a particular call, you add caching or retries or whatever other logic fits your use case. It matters not one bit whether you're using futures or synchronous rpc stubs. Actually it does - synchronous code is easier to harden because it's easier to reason about.

Even javascript added await because it's better to pretend that async code looks synchronous. The failure semantics of "throws an exception" are just fine.

[naasking]

> A remote call fails, the error propagates up the call stack, and someone gets an error message.

Uh-huh, but did the message actually get through? Can they safely just retry? These are very uncommon failure modes on local systems but very common on networked systems. Without a proper stateful abstraction beyond just "procedure call", like a promise, you can't address these failure modes properly.

> In the rare case you need to harden a particular call, you add caching or retries or whatever other logic fits your use case

Which now makes your system nondeterministic like I said.

> Even javascript added await because it's better to pretend that async code looks synchronous

Yes, linear code is easier to read. I don't see what this has to do with anything. The use of promises and await indicates a possibility of failure semantics that would otherwise not be apparent in the program's control-flow.

Yes, you can superficially make this look like synchronous code, but it's not synchronous code.

[lelanthran]

> Uh-huh, but did the message actually get through? Can they safely just retry?

...

> The use of promises and await indicates a possibility of failure semantics that would otherwise not be apparent in the program's control-flow.

They don't, though. They don't indicate if the message got through. They don't indicate if you can safely retry. Their failure mode is exactly as opaque or as transparent as synchronous calls.

The reason for their existence and mandatory use in Javascript is due to a deficiency of the platform (single thread, so all synchronous calls block).

If the platform was better they would never have existed.

[naasking]

> They don't, though. They don't indicate if the message got through. They don't indicate if you can safely retry. Their failure mode is exactly as opaque or as transparent as synchronous calls.

The promise is at the remote end. Promise resolution is idempotent, so retries always resolve to the same value. These are correct promise semantics as pioneered in the E language.

[lelanthran]

> The promise is at the remote end.

What? The promise is local, so what do you mean "the promise is at the remote end?"

> Promise resolution is idempotent,

No. It is only run once (whether resolve or success), that does not mean it's idempotent.

Just like exceptions for synchronous code, or error returns.

[stickfigure]

> you can't address these failure modes properly.

Again, total nonsense. There's nothing special about a promise. Whatever logic you can build on promises is easier to build synchronously. Everything that applies to building distributed systems applies whether you use rpc stubs or promises. Promises are just noisier and harder to reason about.

[naasking]

Maybe we're speaking past each other. Idempotent operations are implicit promises. Network partitions require idempotency at some level to ensure robustness. That means any robust distributed protocol requires promises at the core protocol level.

Trying to hide the promises behind a synchronous client interface is unnecessarily constraining and inefficient, like requiring large stack contexts that can't be restarted or persisted, and so can't be simply resumed after partitions.