Comment by rzwitserloot
3 days ago
OpenJDK redesigns massive swaths of the compiler every other month.
The true explanation, at least the way OpenJDK says it, is that designing language features is more complex than a casual glancer can fathom, and there's 30 years of "Java is in the top 5 most used languages on the planet, probably #1 especially if focussing on stuff that was meant to be supported for a long time" to think about.
From personal experience, essentially every single last "Just do X to add (some lang feature) to java; languages A and B do it and it works great!" would have been bad for java. Usually because it would cause a 'cultural split' - where you can tell some highly used library in the ecosystem was clearly designed before the feature's introduction.
Even if you introduce a new feature in a way that doesn't break existing code, it's still going to cause maintainability headaches if you've cornered the pillars of the ecosystem into total rewrites if they want to remain up to date with the language. Because they will (or somebody will write an alternative that will) and you've _still_ 'python 2 v python 3'd the language and split the baby in the half.
For what its worth, I think the OpenJDK team doesn't take this seriously enough, and a number of recently introduced features have been deployed too hastily without thinking this through. For example, `LocalDate`, which has 'this should be a record' written all over it, is not a record. Or how the securitymanager is being ditched without replacements for what it is most commonly used for here in the 2020s. (To be clear: Ditching it is a good idea, but having no in-process replacement for "let me stop attempts to access files and shut down the JVM, not for security purposes but simply for 'plan B' style fallback purposes" - that's a bit regrettable).
I'm nitpicking on those points because on the whole OpenJDK is doing a far better job than most languages on trying to keep its ecosystem and sizable existing codebase on board _without_ resorting to the crutch of: "Well, users of this language, get to work refactoring everything or embrace obsoletion".
But ... LocalDate predated records by 6 years?
Eventually, I guess there'll be backwards compatible "pattern extractors" functionality retrofittable to existing "record-like" classes. This has been hinted at on several occasions.
> But ... LocalDate predated records by 6 years?
Yes, exactly - now you're getting it. Or rather I get the feeling I failed to explain it well.
ArrayList predates generics.
However, ArrayList does have generics.
That's because generics were added to the language in a 'culturally backwards compatible' way: Existing libraries (i.e. libraries that predate the introduction of generics, such as ArrayList) could modify themselves to add support for generics in a way that is backwards compatible for the library: Code written before they added it continues to work and compile fine even against the new release that added generics.
The same principle applied to records would mean that LocalDate could have been updated to turn into a record in a way that is backwards compatible.
And it really works that way.. almost. You really can take an existing class (defined with `class`) and change it into a record (defined with `record`) and all existing code continues to work just fine. However, this means all your properties now neccessarily get an accessor function that is named after the property. And that is a problem for LocalDate specifically: It already has accessors and they are named e.g. `getYear()`. Not `year()`. That means if LocalDate were to be rewritten as a record, one of two very nasty options must be chosen:
* Break backwards compatibility: As part of upgrading code you must change all calls to `.getYear()` into calls to `.year()`. It's a total ecosystem split: Every dependency you use comes in 2 flavours, one with calls to getYear and one with year, and you must use the right ones. This is truly catastrophic.
* Have both methods. There's year() and also getYear() and they do the same thing. Which is the lesser evil by far, but it makes very clear that LocalDate predates `record`. Contrast to ArrayList: It is not all that obvious that ArrayList predates generics. It does, but if you were to design ArrayList from scratch after generics are introduced you'd probably make the same code. Maybe the signature of `remove` would have been `remove(T)` instead of the current `remove(Object)`.
Instead, obviously then, the best choice is to not make it a record. And that's my point: The best possible (possibly here perfection is the enemy of good, but, I'd have done it differently) way to deploy records would have included some way for localdate to turn into a record without the above dilemma.
Perhaps simply a way to explicitly write your accessors with some marker to indicate 'dont generate the default `year()` - THIS is the accessor for it'.
Had that feature been part of record, then LocalDate could have turned into one in a way that you can't really tell.
Records are great, but objects work.
Date flat out doesn’t. We needed something in the standard library to fix that. It should’ve happened long before it did.
Totally agree, I was just pointing out to GP why LocalDate wasn’t a record. (Date and Calendar should never be used in new code. Note how the new date/time API doesn’t have any conversions to/from old types (Date+Calendar), they only in the opposite direction)
Can’t wait for destructuring support for classes, though.
1 reply →