Systems Correctness Practices at Amazon Web Services

The problem with coupled tooling is that no one will use it. That's what is cool about antithesis. If they're able to complete their goal, that's basically what will be achieved.

I've worked with AWS and GCS for a while, and I have the opposite opinion. GCS is what you get if you let engineers dictate to customers how they are allowed to do work, and then give them shit interfaces, poor documentation, and more complexity than adds value.

There's "I engineered the ultimate engineery thing", and then there's "I made something people actually like using".

S3 is likely an order of magnitude larger then those others - it's had a lot longer to grow.

We have used Coyote/P# not just for model checking an abstract design (which no doubt is very useful) but testing real implementations of production services at Microsoft.

https://sim.tigerbeetle.com/

Yes, because the Go runtime actually explicitly and deliberately leans in the opposite direction. Additional non-determinism is deliberately added to the system, most notably that a "select" call with multiple valid channels will be psuedorandomly chosen, and iteration on a map is somewhat scrambled on each iteration. It's not entirely random and as such not suitable for pretty much any randomness requirement you may have, but enough to generally prevent accidental dependencies on deterministic iteration.

This has the effect of "spreading out" concurrency issues and hopefully reducing the sort of issues you get when you only discover a major issue when something else in the environment changes three years later. It's a nice default but nothing like a proof, of course.

You can also go the "predictable OS" route, since the entire Go runtime sits on top of it. Nothing can non-deterministically run on top of a fully deterministic OS image. But a fully deterministic OS would probably want some integration with the Go runtime to be able to more deeply control it through mechanisms other than varying inputs to the random function, which is a very indirect way to go down a code path.

But `synctest.Wait` won't work for non-durably blocked goroutines (such as in the middle of a TCP syscall) so requires memory-based implementations of e.g. net.Conn (I've plugged in https://pkg.go.dev/google.golang.org/grpc/test/bufconn with good success)

For example, the first implementations of AWS's internal lock service (Alf) used DST as a key part of the testing strategy, sometime around 2009. Al Vermeulen was influential in introducing it at AWS, and I believe it built on some things he'd worked on before.

Still, Anithesis is super cool, and I really admire how you all are changing the conversation around systems correctness. So this is a minor point.

What is the system state when it does error?

What is the best possible recovery from each error state?

What can the user/caller expect for an error?

E.g. if you're in python and raise a value error when an API is rate limited, someone down stream from you is going to have a bad time.

I have a lot of Go testing shims that look like:

    type UserGetter interface {
        GetUser(userID string) (User, error)
    }

    type TestUsers struct {
        Users map[string]User
        Error error
    }

    func (t TestUsers) GetUser(userID string) (User, error) {
        if t.Error != nil {
            return User{}, t.Error
        }
        user, have := t.Users[userID]
        if !have {
            return User{}, ErrUserNotFound
        }
        return user, nil
    }

It's more complicated when you have something like a byte stream where you want to simulate a failure at arbitrary points in the stream, but similar techniques can get you as close as you like, depending on how close that is.

From there, in terms of "how do you handle non-fatal errors", there really isn't a snap rule to give. Quite often you just propagate because there isn't anything else to do. Sometimes you retry some bounded number of times, maybe with backoff. Sometimes you log things and move on. Sometimes you have a fallback you may try. It just depends on your needs. I write a lot of network code, and I find that once my systems mature it's actually the case that rather a lot of the errors in the system get some sort of handling beyond "just propagate it up", but it's hard for me to ever guess in advance what they will be. It's a lot easy to mentally design all the happy paths than it is to figure out all the ways the perverse external universe will screw them up and how we can try to mitigate the issues.

Typical example would be processing an event that you can't handle from a message queue. You don't want to crashloop, so you'd probably have to throw it away on a dead letter queue and continue processing. But then, is your system still correct? What happens if you later receive another event relating to the same entity, which depends on the first event? Or sometimes you can't even tell which entity the malformed or bug-triggering event relates to, and then it's a real problem.

> In 2014, Yuan et al. found that 92% of catastrophic failures in tested distributed systems were triggered by incorrect handling of nonfatal errors.