Sure monads are cool and I’d be tempted to use them. They make it impossible for forget to check for errors and if you don’t care you can panic.

But JS is not Rust. And the default is obviously to use exceptions.

You’ll have to rewrap every API under the moon. So for Monads in JS to make sense you need a lot of weird code that’s awkward to write with exceptions to justify the costs.

I’m not sure the example of doing a retry in the API is “enough” to justify the cost. Also in the example, I’m not sure you should retry. Retries can be dangerous especially if you pile them on top of other retries: https://devblogs.microsoft.com/oldnewthing/20051107-20/?p=33...

> In JS world this could be true, but for Rust (and statically typed compiled languages in general) this is actually not the case… GO pointers are the only exceptions to this. There are no nil check protection at compile level. But Rust, kotlin etc are solid.

Yes it actually is the case. You cannot check/validate for every error, not even in rust. I recommend getting over it.

For a stupid-simple example: You can't even check if disk is going to be full!

The disk being full is a real error you have to deal with, and it could happen at any line in your code through no fault of your own, and no it doesn't always happen at write() but can also when you allocate pages for writing (e.g. SIGSEGV). You cannot really do anything about this with code- aborting or unwinding will only ever annoy users, but you can do something.

We live in a multitasking world, so our users can deal with out-of-disk and out-of-memory errors by deleting files, adding more storage, closing other (lower priority) processes, paging/swapping, and so on. So you can wait: maybe alert the user/operator that there is trouble but then wait for the trouble to clear.

Also: Dynamic-wind is a useful general-purpose programming technique awkward to emulate, and I personally dislike subclassing BackTrack from Error because of what can only be a lack of imagination.

Call it a thought experiment. We start with a clean implementation that satisfies requirements. It makes a bold assumption that every star in the universe will align to help us achieve to goal.

Now we add logging and error handling.

Despite my best intentions and years of experience, starting with clean code, the outcome was a complete mess.

It brings back memories when in 2006 I was implementing deep linking for Wikia. I started with a "true to the documention" implemention which was roughly 10 lines of code. After handling all edge cases and browser incompatibilites I ended up with a whooping 400 lines.

Doing exactly the same as the original lines did, but cross compatible.

This approach is great when:

* program requirements are clear

* correctness is more important than prototyping speed, because every error has to be handled

* no need for concise stack trace, which would require additional layer above simple tuples

* language itself has a great support for binding and mapping values, e.g. first class monads or a bind operator

Good job by the author on acknowledging that this error handling approach is not a solver bullet and has tradeoffs.

In my experience I've used exceptions for things that really should never fail, and optional for things that are more likely to.

1. returning a tuple with an ok or fail value (so errors as values) plus

2. pattern matching on return values (which makes error values bearable) possibly using the with do end macro plus

3. failing on unmatched errors and trying again to execute the failed operation (fail fast) thanks to supervision trees.

Maybe that's because the latter feature is not available nearly for free in most runtimes and because Erlang style pattern matching is also uncommon.

The approach requires a language that's built on those concepts and not one in which they are added unnaturally as an afterthought (the approach becomes burdensome.)

Pattern matching: https://hexdocs.pm/elixir/pattern-matching.html

With: https://hexdocs.pm/elixir/1.18.1/Kernel.SpecialForms.html#wi...

Supervisors: https://hexdocs.pm/elixir/1.18.1/supervisor-and-application....

1. Stacktraces with fields/context besides a string 2. Wrapping errors 3. Combining multiple errors

- Program is broken. Probably need to abort program. Example: subscript out of range.

- Data from an external source is corrupted. Probably need to unwind transaction but program can continue. Example: bad UTF-8 string from input.

- Connection to external device or network reports a problem.

-- Retryable. Wait and try again a few times. Example: HTTP 5xx errors.

-- Non-retryable. Give up now. Example: HTTP 4xx errors.

Python 2 came close to that, but the hierarchy for Python 3 was worse. They tried; all errors are subclasses of a standard error hierarchy, but it doesn't break down well into what's retryable and what isn't.

Rust never got this right, even with Anyhow.

will immediately throw if b == 0, because

    a / b

    () => a / b

Similarly, withRetry()'s argument needs to have type "() => ResultAsync<T, ApiError>" -- at present, it is passed a result, and if that result is a RateLimit error, it will just return the same error again 1s later.

That means good UX, intuitive interfaces, good affordances, user guidance (often, without requiring them to read text), and simplicity.

When an error is encountered, then it needs to be reported to the user in as empathetic and useful manner as possible. It also needs to be as “bare bones” simple as can reasonably be managed.

Designing for low error rates, starts from requirements. Good error reporting requires a lot of [early] input from non-technical stakeholders.

It go, use go style, in js use try/catch/finally.

Some junior engineer is going to stumble upon this and create a mess /bad week for the next poor soul who has to work with this.

However, many make the mistake to handle any errors at the wrong level. This leads to really buggy and hard to reason about code and in some cases really bad data inconsistency issues.

A rule of thumb is to never catch a specific error which you are not in a good position to handle correctly at that precise level of code. Just let them pass through.

Compared to try / catch with await, falling back to promises at least makes the error handling explicit for each request — more along the lines of what Go does without having to introduce a new pattern.

Well, IIUC, Java had (and still has) something called “checked exceptions”, but people have, by and large, elected to not use those kind of exceptions, since it makes the rest of the code balloon out with enormous lists of exceptions, each of which must be changed when some library at the bottom of the stack changes slightly.

That's a weird take. I've been working for multiple decades now with systems that have no UI to speak of; their end-users are barely aware that there's a whole system behind what they can see, and that's a good thing because they become aware of it when it causes them trouble.

I take from my mentor in programming this stance for many things, including error handling: the best solution to a problem is to avoid it. That's something everybody knows actually, but we can forget that when designing/programming because one has so many things to deal with and worry about. Making the thing barely work can be a challenge in itself.

For errors, this usually means: don't let them happen. E.g. avoid OOM by avoiding dynamic allocation as much as possible; statically pre-allocate everything, even if it means megabytes of unused reserved space. Don't design your serialization format with quotes around your keys just to allow "weird" key names, a feature that nobody will ever use and that creates opportunities for errors.

Of course it is not always possible, but don't miss the opportunity when it is.

Isn’t this addressed by preallocating data files in advance of writing application data? It’s pretty common practice for databases for both ensuring space and sometimes performance (by ensuring a contiguous extent allocation).

I hate checked exceptions too, but in fairness to them this specific problem can be handled by intermediate code throwing its own exceptions rather than allowing the lower-level ones to bubble up.

In Go (which uses error values instead) the pattern (if one doesn’t go all the way to defining a new error type) is typically to do:

    if err := doSomething(…); err != nil {
      return fmt.Errorf("couldn’t do something: %w", err)
    }

A similar pattern could be used in languages with checked exceptions.

Checked exceptions should indicate conditions that are expected to be handled by the caller. If a method is throwing a laundry list of checked exceptions then something went wrong in the design of that method’s interface.

That's mostly developer laziness: They write a layer that calls the exception-throwing code, but they don't want to to think about how to model the problem in their own level of abstraction. "Leaking" them upwards by slapping on a "throws" clause is one of the lowest-effort reactions.

What ought to happen is that each layer has its own exception classes, capturing its own model for what kinds of things can go wrong and what kinds of distinctions are necessary. These would abstract-away the lower-level ones, but carrying them along as linked "causes" so that diagnostic detail isn't lost when it comes time for bug-reports.

Ex: If I'm writing a tool to try to analyze and recommend music that has to handle multiple different file types, I might catch an MP3 library's Mp3TagCorruptException and wrap it into my own FileFormatException.

    type FooError = YoureHoldingItWrong | FileError
    type BarError = YoureHoldingItWrong | NetworkError
    fn foo() -> Result<int, FooError> { ... }
    fn bar() -> Result<int, BarError> { ... }
    fn baz() -> Result<String, BarError> { ... }

Rust needs a bit more boilerplate to declare FooError, but the ? syntax automatically calling into(), and into() being free to rearrange errors it bubbles up really help a lot too.

The big problem with Java's checked exceptions was that you need to list all the exceptions on every function, every time.

In surprising twist: Java has ConcurrentModificationException. And, to counter its own culture of exception misuse, the docs have a stern reminder that this exception is supposed to be thrown when there are bugs. You are not supposed to use it to, I dunno, iterate over the collection and bail out (control flow) based on getting this exception.