API Enable persistent FileHashCaching

[maxime-rainville]

Updates the FileHashCaching to be persistent and always be on.

Parent issue

#279

[chillu]

I think we need to customise the cache layer here: The default cache will choose APC if available, and fall back to filesystem. APC is a relatively small in memory cache, here are the defaults in my VM:

vagrant@44:/var/www/mysite/www$ php -i | grep -i apc
...
apc.entries_hint => 4096 => 4096
apc.shm_segments => 1 => 1
apc.shm_size => 32M => 32M

So by default, you can only have 4k cache entries across all caches built by SilverStripe. Most systems will have more files than that. If combined with an infinite cache lifetime (which I'm still recommending), the SHA cache entries will crowd out other ("hotter") caches over time, meaning those caches will use the slower filesystem layer.

Note that APC caches are cleared on webserver restart (or when a new VM is provisioned as part of a full deployment like in SCP). Filesystem caches are often cleared as part of a deployment as well (either by launching a new VM, or keeping the cache folder connected to a specific deployment).

I think we should only use filesystem caches for this, and leave APC to other more important things.

[chillu]

This ignores any other settings on a filesystem adapter. I'm really keen to have non-expiring caches on those hashes, but that's assuming the visibility alone is enough of a unique cache key. In other words, changes in config or metadata couldn't influence the file contents. For example, an AwsS3Adapater could have its bucket location changed (https://github.com/thephpleague/flysystem-aws-s3-v3/blob/master/src/AwsS3Adapter.php), which somehow changes the file contents. There's no toString() method on the Filesystem interface, and no common way to get "all the state which influences this adapter" other than spl_object_hash.

I think that's just a limitation we need to live with, unless we want to force everyone using other flysystem adapters to implement a wrapper that adds our own interface (toString(), getKey(), etc)

[maxime-rainville]

If you're going to change your adapter settings, I don't think it's an unreasonable expectation that you should flush your settings as well. In most cases you'll have to do that anyway to have your new settings take affect.

In any case, if your changes are somehow causing the hash of your files to change, you'll be completely screwed any way, because your database won't know anything about those new hashes.

[chillu]

Actually, it's a bit more subtle: ChainCache is initialised with APCu and Filesystem by default, and will write to every cache (see https://github.com/silverstripe/silverstripe-framework/blob/4/src/Core/Cache/DefaultCacheFactory.php). The APCu cache is defined with $defaultLifetime / 5, which is something that's been there since day one, I've copied the implementation from Symfony.

So this means APC keeps caches for less time than Filesystem caches, making it a more appropriate "hot cache". But it also means if you're setting cache lifetime to 0, the APC cache becomes '0 / 5 = 0`, and you'll start crowding out other caches. I think for this specific cacse, it's better to have longer lived caches which are a bit slower (filesystem), than shorter live caches which are faster (memory). The assumption here is that filesystem access is generally optimised via OPCode caching in PHP.

[chillu]

See silverstripe/silverstripe-framework#9045

[maxime-rainville]

FYI I've squashed those warnings in FileIDHelperResolutionStrategyTest.php

[chillu]

b21bfd3 - isn't this the opposite of the fixes you've done yesterday in #287? There's other tests which started using DI, what makes this one special that it needed to be reverted, when the other ones don't?

There's still a test failure: https://travis-ci.org/silverstripe/silverstripe-assets/jobs/542541684#L883

[maxime-rainville]

You can use the Injector in dataProvider methods because those get call before the manifest has been initialised. That's why I've undone those changes.

I just set the defaultLifetime to 0.

Looking at that last failure now.

[maxime-rainville]

That last commit should make all the test go green.