[RFC] [Discussion] Add bundled user_cache extension

Hi internals,

This is a follow-up to the previous "OPcache Static Cache" discussion.

I have written a new RFC and implementation. It addresses the issues
raised in the previous thread, drops the parts that led to feature
creep, and moves from OPcache integration to a new bundled
user_cache extension.

(The OPcache User Cache entry in the commit history is a remnant of
a completely new implementation built from scratch.
After realizing partway through that it wasn't necessary to depend on
OPcache, I separated it.)

RFC: https://wiki.php.net/rfc/impl_user_cache
Implementation: https://github.com/php/php-src/pull/22730

The scope and API have changed a lot from the previous proposal, so I
am starting this as a new RFC with its own discussion period. Because
of that, I am not sure it can make the PHP 8.6 feature freeze. If the
release managers think 8.6 is not realistic, I will target 8.7.

As mentioned in the previous thread, I think library developers would
benefit from having this available sooner, so feedback is welcome.

Best regards
Go Kudo

Hi Go. First off, I think it's unrealistic for this to make into 8.6 at this point. Especially as it is technically a new RFC now. 8.7 is fine, though. (Or 9, if that's what we call it.)

Which means it's worth considering if the TTL and Lease should use the Duration class rather than ints, as it seems likely to pass.

As to the proposal itself, I overall like it. The remember() method is especially useful, and what I suspect most people (or at least me) will use in practice, especially now that we have FCC.

class ExpensiveThing {
public array $lookup { get => $this->lookup ??= Cache::getPool('expensive')->remember($this->compute(...), 10); }

private function compute() {
// Lots of expensive stuff here.
}
}

That said, remember() doesn't seem to have any lock interaction, which I could see being a problem for exactly that sort of scenario. I'm not entirely certain what the interaction should be, but it should be thought through.

I'm not sure why store() and add() are both necessary. What are the practical uses cases for using add() instead of store()?

It's worth noting explicitly that the cache layering is transparent and largely invisible to the user.

Question: Is it possible to cache a closure? What layer would that live in?

The Locking Semantics section looks to have numerous typos and grammatical errors, making it hard to follow. Please clean it up. (The rest of the doc seems fine.)

--Larry Garfield