Comment by nothrabannosir | Hacker News Reader

Comment by nothrabannosir | Hacker News Reader

> Error is only returned by this function when the secret isn't 32 characters so as long as your secret is definitely 32 characters you can safely ignore the err returned by this function.

Personal preference, but in this case I'd advise using panic() instead of error. Ignoring errors is rarely a good idea: APIs change, other programmers read their code and don't know about your API, wonder why they ignored the error, &c. You can now also not change your API to ever actually return meaningful errors from that function anymore.

Meanwhile, if the failure mode is so well-defined, just put that in the documentation: "panics if the secret key is not exactly 32 characters long." Let the user of your library worry about it---they will have to, anyway: if they don't, they get an error they have to deal with, and suddenly they're dealing with it after all.

EDIT: From a quick glance, it looks like master gets results back from workers using .Get(<workid>)? Is it easily possible for master to "select" a bunch of workers and get the first completed one? If you use (or at least somehow allow the use of) channels for this, you get all the built-in Go goodness for this kind of control. Even for dynamic number of channels, the reflect package has a dynamic .Select that handles it for you. I.e.: provide a channel that I can <- work results from, and let Go (and the user of your lib) handle the rest. Including timeouts :) select { result, ok := <- workchan: ... ; <-time.NewTimer(3 * time.Seconds).C : <timeout> }.

I strongly disagree.

The package author means the error will be nil when the secret is 32 characters long; he doesn't mean you can completely ignore checking it.

Regarding when to use panic: "The convention in the Go libraries is that even when a package uses panic internally, its external API still presents explicit error return values." http://blog.golang.org/defer-panic-and-recover

Panics should only be used in cases of programmer error or when the process is in an unrecoverable state.

> The package author means the error will be nil when the secret is 32 characters long; he doesn't mean you can completely ignore checking it.

But that's exactly what will happen. This comment will lead to ws, _ := gowork.NewServer("32 character secret"). Which is a worse situation than a panic(). I understand the normal rule for this, but "a foolish consistency..". In this case, the init value being 32 long can be compared to "the init value must be non-nil." A lot of non-error APIs will panic() if you pass nil, even in the stdlib.

Doing panic() here will lead to less room for error and confusion, in my opinion. Especially because the error is explicit and loud (panic()), it's worth considering breaking the rules for.

If a developer is frequently ignoring returned errors, then they have larger problems and it's their fault. Panic was not added to assist lazy programmers. Even though some API has solidified, an author could later add different return errors. Heck, if that function was calling other functions with returned errors, then the list of possible unique errors increases greatly. You should never ignore an err because you think you know the possible reasons.

The package author should modify the phrasing of that comment.

Thanks for the feedback (both of you!) - when writing this, I envisioned a use case where the secret was static within the application, so it is easy enough for the developer to check their secret is 32 characters and not have to bother with checking that particular error. I didn't use panic because although a non 32 character secret puts my library in an unrecoverable state, it doesn't mean the application implementing the library is also in an unrecoverable state, hence why it returns an error instead of panicing.

Also, in terms of the top comment in this chain, the work server (master) never gets work from the workers, the workers push completed work back to the server, maybe this isn't clear within the current docs. All communication between the worker and server has to be coordinated by the application implementing the library.

An idea is to include a function in the library called MustNewServer() that panics instead of returning an error. A usability perk of having single return is that the function can be chained, i. e.,

    ws := gowork.MustNewServer(key).AddParams(...)

Great suggestion---I recant my original comment and second this.

Thanks for the suggestion, i'll add something similar to this when i've got a bit of time. Cheers!

As much as I hate chaining functions, I do like the idiom of declaring that a function may panic with the 'Must' prefix.

As long as there is an alternative function which returns errors :)

Reply on news.ycombinator.com