* They now issue the delete() write before COMMIT of
the relevant DB (or immediately if no trx is active)
* This can avoid some stale write race conditions
* Updated the WAN cache delete() docs
Change-Id: Id54887976051120b76528070d5f2ceb357d57897
* If lockTSE is set, make sure it applies during during
the entire duration of a key being tombstoned. This lets
regenerations happen during that whole time, which lowers
the chance of seeing stale data if the DBs are lagged.
* If lockTSE is not set, do not apply tempTTL (for tomstoned keys).
If traffic is high, a stale value would usually be "stashed" and
used for 5 seconds. If the lag was only 1 second, then this
is suboptimal.
* Determine tempTTL from lockTSE as they no longer make sense
being separate. This makes things easier to understand and
also makes the lockTSE value account for the last regeneration
time (via stash key TTL). Since LOCK_TSE << HOLDOFF_TTL, this
helps avoid stale reads without adding stampede risk.
Change-Id: I3b01f0ec67935a238b30e02e42004fd3b2dcfb9d
* This uses a non-blocking $wgMemc lock to reserve the user
name in question. This should prevent two threads from
reaching LOCK IN SHARE MODE and getting stuck on INSERT.
The lock is global to better cover auth plugins.
* This adds a BagOStuff::getScopedLock() convenience method.
It uses less queries than LockManager by being EX only.
* Avoid extra lock attempt in lock() in non-blocking mode.
* Removed (un)lock() HashBagOStuff overrides that made it
behave differently than other caches (wrt to re-entrance).
Bug: T106850
Change-Id: Iecf95206d712367f5d202f76ab0eaa9d7bdabf2b
The last time we had encountered APC errors related to serialization /
unserialization was 2011. PHP's implementation has had many bugfixes since
then, and HHVM's implementation is a complete rewrite. So let's stop working
around alleged bugs.
To prevent errors resulting from HHVM code receiving serialize()d values when
it isn't expecting them, add a key suffix.
Change-Id: I4b2cf1715538aa3d9163787f43eb31984a380d35
* This still allows If-Modified-Since logic but does not
need to broadcast initialization values just because a
key fell out of cache. The value can differ between DCs
anyway via skew, this just lets them drift more. Actual
purge events are still broadcasted, which is what matters.
* The User class has now been simplified given this change.
* Added more general comments to getCheckKeyTime().
Change-Id: Ic1f4bbb1947e0d1dd47499c9e9dc86991c30580c
Fixed PHP runtime warnings:
> Declaration of ReplicatedBagOStuff::getMulti() should be compatible with BagOStuff.
> Declaration of ReplicatedBagOStuff::decr() should be compatible with BagOStuff.
Change-Id: Icf1a0bf2c30408c4a5bef2de0b69ae2162b234d5
* Do not use lock() for tombstoned keys unless lockTSE was set
* Moved the is_callable() check down a bit further
* Also removed one inaccurate comment
Change-Id: I904e0681faa48b1dc2bc2a3c005a29d2f8347065
