APCu support
Closed, ResolvedPublic
Actions

Assigned To

Authored By

	Reedy
	Jul 20 2016, 11:14 PM

Description

PHP 5.5 has apc_ and apcu_. PHP 7 only has apcu_ (without installing apc-bc in pecl. Ugh). HHVM only has apc_

For HHVM supporting apcu_ see https://github.com/facebook/hhvm/issues/6784

For T140587 and https://gerrit.wikimedia.org/r/#/c/299498 I originally just replaced apc_ with apcu_, but for HHVM support, this isn't gonna work.

So, do we create an APCuBagOStuff, and add some if/elseif function_exists in MemoizedCallable as appropriate?

Or, do we "move" everything over to apcu_ calls, and add in GlobalFunctions some wrappers around apc_ per the Compatibility functions?

/**
 * Compatibility functions
 *
 * We support PHP 5.5.9 and up.
 * Re-implementations of newer functions or functions in non-standard
 * PHP extensions may be included here.
 */

As is, upgrading to PHP 7 can break login and other stuff if you're trying to use APC, as @Krenair found out... So, it could be suggested our PHP 7 support in REL1_27 is buggy at least

Details

	Subject	Repo	Branch	Lines +/-
	Detect/use APCu properly	mediawiki/core	REL1_27	+133 -21

Customize query in gerrit

Related Objects
Search...

		Status	Subtype	Assigned	Task
		Resolved		Reedy	T140961 APCu support
		Resolved		Reedy	T140587 web configuration does not properly detect apcu

Event Timeline

Reedy created this task.Jul 20 2016, 11:14 PM

Restricted Application added a subscriber: Aklapper. · View Herald TranscriptJul 20 2016, 11:14 PM

Reedy added a subtask: T140587: web configuration does not properly detect apcu.Jul 20 2016, 11:14 PM

Reedy updated the task description. (Show Details)

Reedy added a subscriber: Krenair.

From @PleaseStand on gerrit:

In APCu 5.1.0 and newer, there is a behavior change between apc_inc (as provided by older versions of APCu, as well as apcu-bc 1.0.2 and newer) and apcu_inc: a value is now initialized to zero with no TTL (prior to performing the operation) if it does not already exist. https://github.com/krakjoe/apcu/issues/166
The doc comment for RedisBagOStuff::incr() states:

Non-atomic implementation of incr(). *

Probably all callers actually want incr() to atomically initialise

values to zero if they don't exist, as provided by the Redis INCR

command. But we are constrained by the memcached-like interface to

return null in that case. Once the key exists, further increments are

atomic.

Actually, at least one caller, incrWithInit(), assumes that incr() will fail if the key does not exist (and expects false in that case, not null), in order to set the specified initial value and/or TTL.