Ie161e0f was done in a hurry, and so didn't do things in the best ways.
This introduces a new "CachedBagOStuff" that transparently handles all
the logic that had been copy-pasted all over in Ie161e0f.
The differences between CachedBagOStuff and MultiWriteBagOStuff are:
* CachedBagOStuff supports only one "backend".
* There's a flag for writes to only go to the in-memory cache.
* The in-memory cache is always updated.
* Locks go to the backend cache (with MultiWriteBagOStuff, it would wind
up going to the HashBagOStuff used for the in-memory cache).
Change-Id: Iea494729bd2e8c6c5ab8facf4c241232e31e8215
* This can happen due to incr/add races. Use incrWithInit()
instead to handle such cases.
* Also made BagOStuff:incrWithInit() return the new value like incr().
Change-Id: I0e3b02a4cff7c20544a9db2eaabd3f61e5a470b1
APCu's object serialization causes memory corruption when apc.serializer is set
to 'default' (see https://github.com/krakjoe/apcu/issues/38). We can work
around this bug by falling back to the pre-I4b2cf17155 behavior of
APCBagOStuff, which is not to trust apc_store() with anything other than
strings and integers, and instead serialize everything in user-space PHP code.
Bug: T120267
Change-Id: If34a1d959e2700792514b372af52919940222d83
MessageBlobStore class:
* Make logger aware.
* Log an error if json encoding fails.
* Stop using the DB table. WANObjectCache supports everything we need:
- Batch retrieval.
- Invalidate keys with wildcard selects or cascading check keys.
* Update tests slightly since the actual update now happens on-demand as
part of get() instead of within updateMessage().
ResourceLoader class:
* Remove all interaction with the msg_resource table. Remove db table later.
* Refactor code to use a hash of the blob instead of a timestamp.
Timestamps are unreliable and roll over too frequently for message blob store
because there is no authoritative source. The timestamps were inferred based on
when a change is observed. Message overrides from the local wiki have an
explicit update event when the page is edited. All other messages, such as
from MediaWiki core and extensions using LocalisationCache, have a single
timestamp for all messages which rolls over every time the cache is rebuilt.
A hash is deterministic, and won't cause needless invalidation (T102578).
* Remove redundant pre-fetching in makeModuleResponse().
This is already done by preloadModuleInfo() in respond().
* Don't bother storing and retreiving empty "{}" objects.
Instead, detect whether a module's message list is empty at runtime.
ResourceLoaderModule class:
* Make logger aware.
* Log if a module's message blob was not preloaded.
cleanupRemovedModules:
* Now that blobs have a TTL, there's no need to prune old entries.
Bug: T113092
Bug: T92357
Change-Id: Id8c26f41a82597e34013f95294cdc3971a4f52ae
It was already used everywhere as non-static via $this.
This is needed in order to allow MessageBlobStore unit tests
to disable the holdoff via a mock (mocks can't override static methods).
Change-Id: I3aad5b6e780addf1b6ce9de56c81b91f5ab358b2
If a set() happened around the exact same time as check key was
initialized via get(), the curTTL in future get() calls could
sometimes be 0 instead of a negative value, even though hold-off
should apply.
Change-Id: Ide1fd65112aff425a4798e2ec406d71f2a8e84a7
Move the holdoff period into the purge value instead of deciding
it at runtime. This opens the way for touchCheckKey() to support
a custom $holdoff parameter, which will allow callers to invalidate
keys without a holdoff period. Similar to what we already support
in delete().
Right now the holdoff period is decided at run time.
Change-Id: Id10c036272e92ae4429effc823b75e08fb11a48b
To allow batch queries for multiple keys that themselves have different check
keys. Previously check keys always applied to all keys being retrieved.
Change-Id: I9e5ba198d79020ce05a802a510762e29fcfb2f1b
Follows-up f36b73e96c, which moved these classes to libs/objectcache.
* Fix wrong @throws in MemcachedPeclBagOStuff.
* Fix wrong @returns in MemcachedBagOStuff::getClient().
* Rename MWMemcached to MemcachedClient.
* Remove mention of 'memcached.php', which doesn't exist anymore.
Change-Id: I34dbc859be4778cea489fd2344f233f30452605c
* Match behaviour of MapCacheLRU and ProcessCacheLRU.
* Add missing unit tests for TTL and maxCacheKeys eviction behaviour.
Change-Id: I559eae1cd336274b21728e86775cfbad7e2f2c6d
Especially useful when interacting with getWithSetCallback of
HashBagOStuff where a $exptime is required before the callback.
In set() it is typically left out and thus not as counter-intuitive.
Change-Id: Ic2f7adda3c00cefe701d77bea91a7e8e77ef1439
* Keys that take a long time to generate would run into
the MAX_SNAPSHOT_LAG check and have set() no-op. This
would be fine except that leaves no key there to figure
out the time since expiry and therefore whether to use
the mutex, so it didn't. This now saves the keys but with
a FLG_STALE bit set, making the next caller that sees it
perform a regeneration (unless it can't get the mutex).
* Bumped LOCK_TTL so that keys can stay locked much longer.
* This is easy to test via sleep(5) in tagUsageStatistics()
and two Special:Tags browser tabs.
Bug: T91535
Change-Id: I549e70ace3d2e7da40d3c4346ebacc36024cd522
Also consistently use self:: instead of BagOStuff:: for constants
referenced within the BagOStuff class.
Change-Id: I20fde9fa5cddcc9e92fa6a02b05dc7effa846742
* In the common ~0 second lag case, transactions up to ~7 seconds
long will have normal set() behavior (instead of just ~5 seconds).
* Like wise with ~0 second transactions tolerating ~7 seconds of
lag (instead of just ~5).
* The lower hold-off time lets caching resume 3 seconds sooner.
Change-Id: I21e2a0a6915492cec422b6a6af5541937745c15b
This generally only effects wikis with no slave DBs,
but also matters if the master has non-zero LB load.
If the master ends up being used for DB_SLAVE, care
should be shown for cache-aside writes
Interesting WAN cache events are now logged.
Change-Id: I2cd8e84138263c13ea23beb9ab3d7562340e1fd3
Also removed a few MW dependencies from MemcachedPeclBagOStuff.
It still uses an IP class method, so it has to stay for now.
Change-Id: I8c5c83046c58fb58091d6ce11b2385208262460f
For the sake of safety and correctness, the following BagOStuff::makeKey()
invocations should return distinct keys:
$cache->makeKey( 'ab:', 'cd' );
$cache->makeKey( 'ab', ':cd' );
That is not currently the case, because while we use ':' as a key path
separator, we don't escape ':' in the input supplied to makeKey(). So, make
BagOStuff::makeKeyInternal() URL-encode colons.
To prevent this from messing up the max. key length calculations, reproduce
this logic in MemcachedBagOStuff::makeKeyInternal(), in lieu of having the
method call its parent.
Change-Id: I83ea7e7336a1c9e64aa42284c2517089a736efe5
* This blocks on writing to all replicas
and returns false if any failed.
* This is useful if ChronologyProtector is to work across
domains by having the writes go everywhere so that later
reads will see them (and be local at the same time).
* Redundant doc comments were also removed.
Change-Id: I9ed098d563c64dba605e7809bc96731da3b3e79d
