Re: "TryX" idom for Enumerations

On Fri, Jan 15, 2021, at 9:55 AM, Pierre R. wrote: > Le 10/01/2021 à 22:27, Larry Garfield a écrit : > > This is a little tangent from the Enums RFC, but I want to flag it because it it's > > the sort of in-passing decision that could have far-reaching implications, so shouldn't be done > > implicitly. > > > > At the moment, the Enum RFC for scalar enums includes two methods: > > > > public function has(string $name): bool > > > > public function from(string $name): self throws ValueError > > > > Nikita raised the point that has() seems kinda pointless as it would only ever be used to > > wrap a possibly-unsafe from() call. (In most other cases, calling cases() would be more useful or > > at least equally useful..) One of the proposed alternatives was the following: > > > > public function from(string $name): self throws ValueError > > > > public function tryFrom(string $name): ?self > > > > The "a method that begins with try is nullable, so watch out" idiom is present > > in C# and Rust, but to my knowledge has never existed in PHP. That doesn't make it bad; it > > actually combines quite well with the null coalesce operator to allow for default values, making a > > valueOrDefault() method unnecessary. > > > > $order = SortOrder::tryFrom($input) ?? SortOrder::Asc; > > > > I'm not opposed to following that pattern here; it would allow both a "hard > > fail" and "soft fail" variant of the operation. However, as noted that idiom has > > never appeared in PHP before that I'm aware. If we adopt it here, that means it will either > > start to spread and become a more common PHP idiom over time, OR it won't spread and Enums will > > have this weird one-off naming convention for a nullable method. The former would, of course, be > > considerably preferable to the latter. > > > > So, explicit decision time: Are we OK with introducing that idiom, and then following it > > consistently in the future in similar situations? (viz, tryX() means nullable, and no-try means not > > nullable.) I'm good with it if the consensus is good with it, but I want to see what the > > consensus is first. > > I'm OK with the tryBar(): ?Foo pattern, but my preference goes over just > bar(): ?Foo, as long as API are explicitly typed and correctly documented. Well, the concern is that we also want to have a from(): Foo throws ValueError version, but both cannot be called from(). The possible patterns are: Interface: Suit::from(string): ?self Usage: $s = Suit::from($var) ?? Suit::SomeDefault; Interface: Suit::from(string): self throws ValueError Suit::has(string): bool Usage: if (Suit::has($var)) { $s = Suit::from($var); } Technically, one could create the second from the first more easily than vice versa, but both could implement each other. try { $s = Suit::from($var); } catch (ValueError) { $s = Suit::SomeDefault; } // or $s= Suit::from($var) ?? throw new ValueError(); While the nullable return ?? is very appealing, we know from experience that most people don't think to check the null case, leading to unhelpful "method called on non object" errors 6 months later when someone passes in bad data. An Optional/Maybe type is the ideal solution as that forces people to account for that possibility, but we don't have those yet. (Enums are a step toward being able to build them cleanly.) A thrown exception is next, although nothing forces people to include the try-catch so they could forget that, too. At least static analyzers tend to warn you about it. Nullable is least likely to get handled in practice, even if the resulting code may, in all honestly, look the cleanest if you're OK with ??. Offering both exception and nullable variants is a compromise option, but that requires disambiguating the names somehow. TryX -> nullable is the least-bad suggestion so far, IMO. I'm still not sure of how we want to proceed. :-/ > I will not fight against or for any or another solution, as long as the > convention remains the same for everything. > > Larry, I think your proposal is missing something: if a convention > emerge, it should be documented as such, and be considered as a law for > subsequent language / API addition, the real question, IMHO, is : Are > people OK with writing naming and design conventions by law in PHP core ? > > For that matters, I'm OK, this would avoid many bikesheds in the future: > this is the law, and the law says, name it tryX(), no more useless > votes, no more flavor-oriented flame wars on naming :) While such a formal convention would be nice, that's a separate fight, and the last time I picked that fight (https://wiki.php.net/rfc/php_namespace_policy) we lost. :-) For now I just want to keep enums de facto consistent. --Larry Garfield