Merge "registration: Allow loading composer's autoloader if it exists"
[lhc/web/wiklou.git] / includes / libs / objectcache / WANObjectCache.php
1 <?php
2 /**
3 * This program is free software; you can redistribute it and/or modify
4 * it under the terms of the GNU General Public License as published by
5 * the Free Software Foundation; either version 2 of the License, or
6 * (at your option) any later version.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11 * GNU General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public License along
14 * with this program; if not, write to the Free Software Foundation, Inc.,
15 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
16 * http://www.gnu.org/copyleft/gpl.html
17 *
18 * @file
19 * @ingroup Cache
20 * @author Aaron Schulz
21 */
22
23 use Psr\Log\LoggerAwareInterface;
24 use Psr\Log\LoggerInterface;
25 use Psr\Log\NullLogger;
26
27 /**
28 * Multi-datacenter aware caching interface
29 *
30 * All operations go to the local datacenter cache, except for delete(),
31 * touchCheckKey(), and resetCheckKey(), which broadcast to all datacenters.
32 *
33 * This class is intended for caching data from primary stores.
34 * If the get() method does not return a value, then the caller
35 * should query the new value and backfill the cache using set().
36 * When querying the store on cache miss, the closest DB replica
37 * should be used. Try to avoid heavyweight DB master or quorum reads.
38 * When the source data changes, a purge method should be called.
39 * Since purges are expensive, they should be avoided. One can do so if:
40 * - a) The object cached is immutable; or
41 * - b) Validity is checked against the source after get(); or
42 * - c) Using a modest TTL is reasonably correct and performant
43 *
44 * The simplest purge method is delete().
45 *
46 * Instances of this class must be configured to point to a valid
47 * PubSub endpoint, and there must be listeners on the cache servers
48 * that subscribe to the endpoint and update the caches.
49 *
50 * Broadcasted operations like delete() and touchCheckKey() are done
51 * synchronously in the local datacenter, but are relayed asynchronously.
52 * This means that callers in other datacenters will see older values
53 * for however many milliseconds the datacenters are apart. As with
54 * any cache, this should not be relied on for cases where reads are
55 * used to determine writes to source (e.g. non-cache) data stores.
56 *
57 * All values are wrapped in metadata arrays. Keys use a "WANCache:" prefix
58 * to avoid collisions with keys that are not wrapped as metadata arrays. The
59 * prefixes are as follows:
60 * - a) "WANCache:v" : used for regular value keys
61 * - b) "WANCache:s" : used for temporarily storing values of tombstoned keys
62 * - c) "WANCache:t" : used for storing timestamp "check" keys
63 *
64 * @ingroup Cache
65 * @since 1.26
66 */
67 class WANObjectCache implements IExpiringStore, LoggerAwareInterface {
68 /** @var BagOStuff The local datacenter cache */
69 protected $cache;
70 /** @var HashBagOStuff Script instance PHP cache */
71 protected $procCache;
72 /** @var string Cache pool name */
73 protected $pool;
74 /** @var EventRelayer Bus that handles purge broadcasts */
75 protected $relayer;
76 /** @var LoggerInterface */
77 protected $logger;
78
79 /** @var int ERR_* constant for the "last error" registry */
80 protected $lastRelayError = self::ERR_NONE;
81
82 /** Max time expected to pass between delete() and DB commit finishing */
83 const MAX_COMMIT_DELAY = 3;
84 /** Max replication+snapshot lag before applying TTL_LAGGED or disallowing set() */
85 const MAX_READ_LAG = 7;
86 /** Seconds to tombstone keys on delete() */
87 const HOLDOFF_TTL = 11; // MAX_COMMIT_DELAY + MAX_READ_LAG + 1
88
89 /** Seconds to keep dependency purge keys around */
90 const CHECK_KEY_TTL = self::TTL_YEAR;
91 /** Seconds to keep lock keys around */
92 const LOCK_TTL = 10;
93 /** Default remaining TTL at which to consider pre-emptive regeneration */
94 const LOW_TTL = 30;
95 /** Default time-since-expiry on a miss that makes a key "hot" */
96 const LOCK_TSE = 1;
97
98 /** Idiom for getWithSetCallback() callbacks to avoid calling set() */
99 const TTL_UNCACHEABLE = -1;
100 /** Idiom for getWithSetCallback() callbacks to 'lockTSE' logic */
101 const TSE_NONE = -1;
102 /** Max TTL to store keys when a data sourced is lagged */
103 const TTL_LAGGED = 30;
104 /** Idiom for delete() for "no hold-off" */
105 const HOLDOFF_NONE = 0;
106
107 /** Tiny negative float to use when CTL comes up >= 0 due to clock skew */
108 const TINY_NEGATIVE = -0.000001;
109
110 /** Cache format version number */
111 const VERSION = 1;
112
113 const FLD_VERSION = 0;
114 const FLD_VALUE = 1;
115 const FLD_TTL = 2;
116 const FLD_TIME = 3;
117 const FLD_FLAGS = 4;
118 const FLD_HOLDOFF = 5;
119
120 /** @var integer Treat this value as expired-on-arrival */
121 const FLG_STALE = 1;
122
123 const ERR_NONE = 0; // no error
124 const ERR_NO_RESPONSE = 1; // no response
125 const ERR_UNREACHABLE = 2; // can't connect
126 const ERR_UNEXPECTED = 3; // response gave some error
127 const ERR_RELAY = 4; // relay broadcast failed
128
129 const VALUE_KEY_PREFIX = 'WANCache:v:';
130 const STASH_KEY_PREFIX = 'WANCache:s:';
131 const TIME_KEY_PREFIX = 'WANCache:t:';
132
133 const PURGE_VAL_PREFIX = 'PURGED:';
134
135 const MAX_PC_KEYS = 1000; // max keys to keep in process cache
136
137 /**
138 * @param array $params
139 * - cache : BagOStuff object
140 * - pool : pool name
141 * - relayer : EventRelayer object
142 * - logger : LoggerInterface object
143 */
144 public function __construct( array $params ) {
145 $this->cache = $params['cache'];
146 $this->pool = $params['pool'];
147 $this->relayer = $params['relayer'];
148 $this->procCache = new HashBagOStuff( array( 'maxKeys' => self::MAX_PC_KEYS ) );
149 $this->setLogger( isset( $params['logger'] ) ? $params['logger'] : new NullLogger() );
150 }
151
152 public function setLogger( LoggerInterface $logger ) {
153 $this->logger = $logger;
154 }
155
156 /**
157 * Get an instance that wraps EmptyBagOStuff
158 *
159 * @return WANObjectCache
160 */
161 public static function newEmpty() {
162 return new self( array(
163 'cache' => new EmptyBagOStuff(),
164 'pool' => 'empty',
165 'relayer' => new EventRelayerNull( array() )
166 ) );
167 }
168
169 /**
170 * Fetch the value of a key from cache
171 *
172 * If supplied, $curTTL is set to the remaining TTL (current time left):
173 * - a) INF; if $key exists, has no TTL, and is not expired by $checkKeys
174 * - b) float (>=0); if $key exists, has a TTL, and is not expired by $checkKeys
175 * - c) float (<0); if $key is tombstoned, stale, or existing but expired by $checkKeys
176 * - d) null; if $key does not exist and is not tombstoned
177 *
178 * If a key is tombstoned, $curTTL will reflect the time since delete().
179 *
180 * The timestamp of $key will be checked against the last-purge timestamp
181 * of each of $checkKeys. Those $checkKeys not in cache will have the last-purge
182 * initialized to the current timestamp. If any of $checkKeys have a timestamp
183 * greater than that of $key, then $curTTL will reflect how long ago $key
184 * became invalid. Callers can use $curTTL to know when the value is stale.
185 * The $checkKeys parameter allow mass invalidations by updating a single key:
186 * - a) Each "check" key represents "last purged" of some source data
187 * - b) Callers pass in relevant "check" keys as $checkKeys in get()
188 * - c) When the source data that "check" keys represent changes,
189 * the touchCheckKey() method is called on them
190 *
191 * Source data entities might exists in a DB that uses snapshot isolation
192 * (e.g. the default REPEATABLE-READ in innoDB). Even for mutable data, that
193 * isolation can largely be maintained by doing the following:
194 * - a) Calling delete() on entity change *and* creation, before DB commit
195 * - b) Keeping transaction duration shorter than delete() hold-off TTL
196 *
197 * However, pre-snapshot values might still be seen if an update was made
198 * in a remote datacenter but the purge from delete() didn't relay yet.
199 *
200 * Consider using getWithSetCallback() instead of get() and set() cycles.
201 * That method has cache slam avoiding features for hot/expensive keys.
202 *
203 * @param string $key Cache key
204 * @param mixed $curTTL Approximate TTL left on the key if present [returned]
205 * @param array $checkKeys List of "check" keys
206 * @return mixed Value of cache key or false on failure
207 */
208 final public function get( $key, &$curTTL = null, array $checkKeys = array() ) {
209 $curTTLs = array();
210 $values = $this->getMulti( array( $key ), $curTTLs, $checkKeys );
211 $curTTL = isset( $curTTLs[$key] ) ? $curTTLs[$key] : null;
212
213 return isset( $values[$key] ) ? $values[$key] : false;
214 }
215
216 /**
217 * Fetch the value of several keys from cache
218 *
219 * @see WANObjectCache::get()
220 *
221 * @param array $keys List of cache keys
222 * @param array $curTTLs Map of (key => approximate TTL left) for existing keys [returned]
223 * @param array $checkKeys List of check keys to apply to all $keys. May also apply "check"
224 * keys to specific cache keys only by using cache keys as keys in the $checkKeys array.
225 * @return array Map of (key => value) for keys that exist
226 */
227 final public function getMulti(
228 array $keys, &$curTTLs = array(), array $checkKeys = array()
229 ) {
230 $result = array();
231 $curTTLs = array();
232
233 $vPrefixLen = strlen( self::VALUE_KEY_PREFIX );
234 $valueKeys = self::prefixCacheKeys( $keys, self::VALUE_KEY_PREFIX );
235
236 $checkKeysForAll = array();
237 $checkKeysByKey = array();
238 $checkKeysFlat = array();
239 foreach ( $checkKeys as $i => $keys ) {
240 $prefixed = self::prefixCacheKeys( (array)$keys, self::TIME_KEY_PREFIX );
241 $checkKeysFlat = array_merge( $checkKeysFlat, $prefixed );
242 // Is this check keys for a specific cache key, or for all keys being fetched?
243 if ( is_int( $i ) ) {
244 $checkKeysForAll = array_merge( $checkKeysForAll, $prefixed );
245 } else {
246 $checkKeysByKey[$i] = isset( $checkKeysByKey[$i] )
247 ? array_merge( $checkKeysByKey[$i], $prefixed )
248 : $prefixed;
249 }
250 }
251
252 // Fetch all of the raw values
253 $wrappedValues = $this->cache->getMulti( array_merge( $valueKeys, $checkKeysFlat ) );
254 // Time used to compare/init "check" keys (derived after getMulti() to be pessimistic)
255 $now = microtime( true );
256
257 // Collect timestamps from all "check" keys
258 $purgeValuesForAll = $this->processCheckKeys( $checkKeysForAll, $wrappedValues, $now );
259 $purgeValuesByKey = array();
260 foreach ( $checkKeysByKey as $cacheKey => $checks ) {
261 $purgeValuesByKey[$cacheKey] =
262 $this->processCheckKeys( $checks, $wrappedValues, $now );
263 }
264
265 // Get the main cache value for each key and validate them
266 foreach ( $valueKeys as $vKey ) {
267 if ( !isset( $wrappedValues[$vKey] ) ) {
268 continue; // not found
269 }
270
271 $key = substr( $vKey, $vPrefixLen ); // unprefix
272
273 list( $value, $curTTL ) = $this->unwrap( $wrappedValues[$vKey], $now );
274 if ( $value !== false ) {
275 $result[$key] = $value;
276
277 // Force dependant keys to be invalid for a while after purging
278 // to reduce race conditions involving stale data getting cached
279 $purgeValues = $purgeValuesForAll;
280 if ( isset( $purgeValuesByKey[$key] ) ) {
281 $purgeValues = array_merge( $purgeValues, $purgeValuesByKey[$key] );
282 }
283 foreach ( $purgeValues as $purge ) {
284 $safeTimestamp = $purge[self::FLD_TIME] + $purge[self::FLD_HOLDOFF];
285 if ( $safeTimestamp >= $wrappedValues[$vKey][self::FLD_TIME] ) {
286 // How long ago this value was expired by *this* check key
287 $ago = min( $purge[self::FLD_TIME] - $now, self::TINY_NEGATIVE );
288 // How long ago this value was expired by *any* known check key
289 $curTTL = min( $curTTL, $ago );
290 }
291 }
292 }
293 $curTTLs[$key] = $curTTL;
294 }
295
296 return $result;
297 }
298
299 /**
300 * @since 1.27
301 * @param array $timeKeys List of prefixed time check keys
302 * @param array $wrappedValues
303 * @param float $now
304 * @return array List of purge value arrays
305 */
306 private function processCheckKeys( array $timeKeys, array $wrappedValues, $now ) {
307 $purgeValues = array();
308 foreach ( $timeKeys as $timeKey ) {
309 $purge = isset( $wrappedValues[$timeKey] )
310 ? self::parsePurgeValue( $wrappedValues[$timeKey] )
311 : false;
312 if ( $purge === false ) {
313 // Key is not set or invalid; regenerate
314 $newVal = $this->makePurgeValue( $now, self::HOLDOFF_TTL );
315 $this->cache->add( $timeKey, $newVal, self::CHECK_KEY_TTL );
316 $purge = self::parsePurgeValue( $newVal );
317 }
318 $purgeValues[] = $purge;
319 }
320 return $purgeValues;
321 }
322
323 /**
324 * Set the value of a key in cache
325 *
326 * Simply calling this method when source data changes is not valid because
327 * the changes do not replicate to the other WAN sites. In that case, delete()
328 * should be used instead. This method is intended for use on cache misses.
329 *
330 * If the data was read from a snapshot-isolated transactions (e.g. the default
331 * REPEATABLE-READ in innoDB), use 'since' to avoid the following race condition:
332 * - a) T1 starts
333 * - b) T2 updates a row, calls delete(), and commits
334 * - c) The HOLDOFF_TTL passes, expiring the delete() tombstone
335 * - d) T1 reads the row and calls set() due to a cache miss
336 * - e) Stale value is stuck in cache
337 *
338 * Setting 'lag' and 'since' help avoids keys getting stuck in stale states.
339 *
340 * Example usage:
341 * @code
342 * $dbr = wfGetDB( DB_SLAVE );
343 * $setOpts = Database::getCacheSetOptions( $dbr );
344 * // Fetch the row from the DB
345 * $row = $dbr->selectRow( ... );
346 * $key = $cache->makeKey( 'building', $buildingId );
347 * $cache->set( $key, $row, $cache::TTL_DAY, $setOpts );
348 * @endcode
349 *
350 * @param string $key Cache key
351 * @param mixed $value
352 * @param integer $ttl Seconds to live. Special values are:
353 * - WANObjectCache::TTL_INDEFINITE: Cache forever
354 * @param array $opts Options map:
355 * - lag : Seconds of slave lag. Typically, this is either the slave lag
356 * before the data was read or, if applicable, the slave lag before
357 * the snapshot-isolated transaction the data was read from started.
358 * Default: 0 seconds
359 * - since : UNIX timestamp of the data in $value. Typically, this is either
360 * the current time the data was read or (if applicable) the time when
361 * the snapshot-isolated transaction the data was read from started.
362 * Default: 0 seconds
363 * - pending : Whether this data is possibly from an uncommitted write transaction.
364 * Generally, other threads should not see values from the future and
365 * they certainly should not see ones that ended up getting rolled back.
366 * Default: false
367 * - lockTSE : if excessive replication/snapshot lag is detected, then store the value
368 * with this TTL and flag it as stale. This is only useful if the reads for
369 * this key use getWithSetCallback() with "lockTSE" set.
370 * Default: WANObjectCache::TSE_NONE
371 * @return bool Success
372 */
373 final public function set( $key, $value, $ttl = 0, array $opts = array() ) {
374 $lockTSE = isset( $opts['lockTSE'] ) ? $opts['lockTSE'] : self::TSE_NONE;
375 $age = isset( $opts['since'] ) ? max( 0, microtime( true ) - $opts['since'] ) : 0;
376 $lag = isset( $opts['lag'] ) ? $opts['lag'] : 0;
377
378 // Do not cache potentially uncommitted data as it might get rolled back
379 if ( !empty( $opts['pending'] ) ) {
380 $this->logger->info( "Rejected set() for $key due to pending writes." );
381
382 return true; // no-op the write for being unsafe
383 }
384
385 $wrapExtra = array(); // additional wrapped value fields
386 // Check if there's a risk of writing stale data after the purge tombstone expired
387 if ( $lag === false || ( $lag + $age ) > self::MAX_READ_LAG ) {
388 // Case A: read lag with "lockTSE"; save but record value as stale
389 if ( $lockTSE >= 0 ) {
390 $ttl = max( 1, (int)$lockTSE ); // set() expects seconds
391 $wrapExtra[self::FLD_FLAGS] = self::FLG_STALE; // mark as stale
392 // Case B: any long-running transaction; ignore this set()
393 } elseif ( $age > self::MAX_READ_LAG ) {
394 $this->logger->warning( "Rejected set() for $key due to snapshot lag." );
395
396 return true; // no-op the write for being unsafe
397 // Case C: high replication lag; lower TTL instead of ignoring all set()s
398 } elseif ( $lag === false || $lag > self::MAX_READ_LAG ) {
399 $ttl = $ttl ? min( $ttl, self::TTL_LAGGED ) : self::TTL_LAGGED;
400 $this->logger->warning( "Lowered set() TTL for $key due to replication lag." );
401 // Case D: medium length request with medium replication lag; ignore this set()
402 } else {
403 $this->logger->warning( "Rejected set() for $key due to high read lag." );
404
405 return true; // no-op the write for being unsafe
406 }
407 }
408
409 // Wrap that value with time/TTL/version metadata
410 $wrapped = $this->wrap( $value, $ttl ) + $wrapExtra;
411
412 $func = function ( $cache, $key, $cWrapped ) use ( $wrapped ) {
413 return ( is_string( $cWrapped ) )
414 ? false // key is tombstoned; do nothing
415 : $wrapped;
416 };
417
418 return $this->cache->merge( self::VALUE_KEY_PREFIX . $key, $func, $ttl, 1 );
419 }
420
421 /**
422 * Purge a key from all datacenters
423 *
424 * This should only be called when the underlying data (being cached)
425 * changes in a significant way. This deletes the key and starts a hold-off
426 * period where the key cannot be written to for a few seconds (HOLDOFF_TTL).
427 * This is done to avoid the following race condition:
428 * - a) Some DB data changes and delete() is called on a corresponding key
429 * - b) A request refills the key with a stale value from a lagged DB
430 * - c) The stale value is stuck there until the key is expired/evicted
431 *
432 * This is implemented by storing a special "tombstone" value at the cache
433 * key that this class recognizes; get() calls will return false for the key
434 * and any set() calls will refuse to replace tombstone values at the key.
435 * For this to always avoid stale value writes, the following must hold:
436 * - a) Replication lag is bounded to being less than HOLDOFF_TTL; or
437 * - b) If lag is higher, the DB will have gone into read-only mode already
438 *
439 * Note that set() can also be lag-aware and lower the TTL if it's high.
440 *
441 * When using potentially long-running ACID transactions, a good pattern is
442 * to use a pre-commit hook to issue the delete. This means that immediately
443 * after commit, callers will see the tombstone in cache in the local datacenter
444 * and in the others upon relay. It also avoids the following race condition:
445 * - a) T1 begins, changes a row, and calls delete()
446 * - b) The HOLDOFF_TTL passes, expiring the delete() tombstone
447 * - c) T2 starts, reads the row and calls set() due to a cache miss
448 * - d) T1 finally commits
449 * - e) Stale value is stuck in cache
450 *
451 * Example usage:
452 * @code
453 * $dbw->begin( __METHOD__ ); // start of request
454 * ... <execute some stuff> ...
455 * // Update the row in the DB
456 * $dbw->update( ... );
457 * $key = $cache->makeKey( 'homes', $homeId );
458 * // Purge the corresponding cache entry just before committing
459 * $dbw->onTransactionPreCommitOrIdle( function() use ( $cache, $key ) {
460 * $cache->delete( $key );
461 * } );
462 * ... <execute some stuff> ...
463 * $dbw->commit( __METHOD__ ); // end of request
464 * @endcode
465 *
466 * The $ttl parameter can be used when purging values that have not actually changed
467 * recently. For example, a cleanup script to purge cache entries does not really need
468 * a hold-off period, so it can use HOLDOFF_NONE. Likewise for user-requested purge.
469 * Note that $ttl limits the effective range of 'lockTSE' for getWithSetCallback().
470 *
471 * If called twice on the same key, then the last hold-off TTL takes precedence. For
472 * idempotence, the $ttl should not vary for different delete() calls on the same key.
473 *
474 * @param string $key Cache key
475 * @param integer $ttl Tombstone TTL; Default: WANObjectCache::HOLDOFF_TTL
476 * @return bool True if the item was purged or not found, false on failure
477 */
478 final public function delete( $key, $ttl = self::HOLDOFF_TTL ) {
479 $key = self::VALUE_KEY_PREFIX . $key;
480
481 if ( $ttl <= 0 ) {
482 // Update the local datacenter immediately
483 $ok = $this->cache->delete( $key );
484 // Publish the purge to all datacenters
485 $ok = $this->relayDelete( $key ) && $ok;
486 } else {
487 // Update the local datacenter immediately
488 $ok = $this->cache->set( $key,
489 $this->makePurgeValue( microtime( true ), self::HOLDOFF_NONE ),
490 $ttl
491 );
492 // Publish the purge to all datacenters
493 $ok = $this->relayPurge( $key, $ttl, self::HOLDOFF_NONE ) && $ok;
494 }
495
496 return $ok;
497 }
498
499 /**
500 * Fetch the value of a timestamp "check" key
501 *
502 * The key will be *initialized* to the current time if not set,
503 * so only call this method if this behavior is actually desired
504 *
505 * The timestamp can be used to check whether a cached value is valid.
506 * Callers should not assume that this returns the same timestamp in
507 * all datacenters due to relay delays.
508 *
509 * The level of staleness can roughly be estimated from this key, but
510 * if the key was evicted from cache, such calculations may show the
511 * time since expiry as ~0 seconds.
512 *
513 * Note that "check" keys won't collide with other regular keys.
514 *
515 * @param string $key
516 * @return float UNIX timestamp of the check key
517 */
518 final public function getCheckKeyTime( $key ) {
519 $key = self::TIME_KEY_PREFIX . $key;
520
521 $purge = self::parsePurgeValue( $this->cache->get( $key ) );
522 if ( $purge !== false ) {
523 $time = $purge[self::FLD_TIME];
524 } else {
525 // Casting assures identical floats for the next getCheckKeyTime() calls
526 $now = (string)microtime( true );
527 $this->cache->add( $key,
528 $this->makePurgeValue( $now, self::HOLDOFF_TTL ),
529 self::CHECK_KEY_TTL
530 );
531 $time = (float)$now;
532 }
533
534 return $time;
535 }
536
537 /**
538 * Purge a "check" key from all datacenters, invalidating keys that use it
539 *
540 * This should only be called when the underlying data (being cached)
541 * changes in a significant way, and it is impractical to call delete()
542 * on all keys that should be changed. When get() is called on those
543 * keys, the relevant "check" keys must be supplied for this to work.
544 *
545 * The "check" key essentially represents a last-modified field.
546 * When touched, keys using it via get(), getMulti(), or getWithSetCallback()
547 * will be invalidated. It is treated as being HOLDOFF_TTL seconds in the future
548 * by those methods to avoid race conditions where dependent keys get updated
549 * with stale values (e.g. from a DB slave).
550 *
551 * This is typically useful for keys with hardcoded names or in some cases
552 * dynamically generated names where a low number of combinations exist.
553 * When a few important keys get a large number of hits, a high cache
554 * time is usually desired as well as "lockTSE" logic. The resetCheckKey()
555 * method is less appropriate in such cases since the "time since expiry"
556 * cannot be inferred.
557 *
558 * Note that "check" keys won't collide with other regular keys.
559 *
560 * @see WANObjectCache::get()
561 * @see WANObjectCache::getWithSetCallback()
562 * @see WANObjectCache::resetCheckKey()
563 *
564 * @param string $key Cache key
565 * @param int $holdoff HOLDOFF_TTL or HOLDOFF_NONE constant
566 * @return bool True if the item was purged or not found, false on failure
567 */
568 final public function touchCheckKey( $key, $holdoff = self::HOLDOFF_TTL ) {
569 $key = self::TIME_KEY_PREFIX . $key;
570 // Update the local datacenter immediately
571 $ok = $this->cache->set( $key,
572 $this->makePurgeValue( microtime( true ), $holdoff ),
573 self::CHECK_KEY_TTL
574 );
575 // Publish the purge to all datacenters
576 return $this->relayPurge( $key, self::CHECK_KEY_TTL, $holdoff ) && $ok;
577 }
578
579 /**
580 * Delete a "check" key from all datacenters, invalidating keys that use it
581 *
582 * This is similar to touchCheckKey() in that keys using it via get(), getMulti(),
583 * or getWithSetCallback() will be invalidated. The differences are:
584 * - a) The timestamp will be deleted from all caches and lazily
585 * re-initialized when accessed (rather than set everywhere)
586 * - b) Thus, dependent keys will be known to be invalid, but not
587 * for how long (they are treated as "just" purged), which
588 * effects any lockTSE logic in getWithSetCallback()
589 *
590 * The advantage is that this does not place high TTL keys on every cache
591 * server, making it better for code that will cache many different keys
592 * and either does not use lockTSE or uses a low enough TTL anyway.
593 *
594 * This is typically useful for keys with dynamically generated names
595 * where a high number of combinations exist.
596 *
597 * Note that "check" keys won't collide with other regular keys.
598 *
599 * @see WANObjectCache::get()
600 * @see WANObjectCache::getWithSetCallback()
601 * @see WANObjectCache::touchCheckKey()
602 *
603 * @param string $key Cache key
604 * @return bool True if the item was purged or not found, false on failure
605 */
606 final public function resetCheckKey( $key ) {
607 $key = self::TIME_KEY_PREFIX . $key;
608 // Update the local datacenter immediately
609 $ok = $this->cache->delete( $key );
610 // Publish the purge to all datacenters
611 return $this->relayDelete( $key ) && $ok;
612 }
613
614 /**
615 * Method to fetch/regenerate cache keys
616 *
617 * On cache miss, the key will be set to the callback result via set()
618 * (unless the callback returns false) and that result will be returned.
619 * The arguments supplied to the callback are:
620 * - $oldValue : current cache value or false if not present
621 * - &$ttl : a reference to the TTL which can be altered
622 * - &$setOpts : a reference to options for set() which can be altered
623 *
624 * It is strongly recommended to set the 'lag' and 'since' fields to avoid race conditions
625 * that can cause stale values to get stuck at keys. Usually, callbacks ignore the current
626 * value, but it can be used to maintain "most recent X" values that come from time or
627 * sequence based source data, provided that the "as of" id/time is tracked. Note that
628 * preemptive regeneration and $checkKeys can result in a non-false current value.
629 *
630 * Usage of $checkKeys is similar to get() and getMulti(). However, rather than the caller
631 * having to inspect a "current time left" variable (e.g. $curTTL, $curTTLs), a cache
632 * regeneration will automatically be triggered using the callback.
633 *
634 * The simplest way to avoid stampedes for hot keys is to use
635 * the 'lockTSE' option in $opts. If cache purges are needed, also:
636 * - a) Pass $key into $checkKeys
637 * - b) Use touchCheckKey( $key ) instead of delete( $key )
638 *
639 * Example usage (typical key):
640 * @code
641 * $catInfo = $cache->getWithSetCallback(
642 * // Key to store the cached value under
643 * $cache->makeKey( 'cat-attributes', $catId ),
644 * // Time-to-live (in seconds)
645 * $cache::TTL_MINUTE,
646 * // Function that derives the new key value
647 * function ( $oldValue, &$ttl, array &$setOpts ) {
648 * $dbr = wfGetDB( DB_SLAVE );
649 * // Account for any snapshot/slave lag
650 * $setOpts += Database::getCacheSetOptions( $dbr );
651 *
652 * return $dbr->selectRow( ... );
653 * }
654 * );
655 * @endcode
656 *
657 * Example usage (key that is expensive and hot):
658 * @code
659 * $catConfig = $cache->getWithSetCallback(
660 * // Key to store the cached value under
661 * $cache->makeKey( 'site-cat-config' ),
662 * // Time-to-live (in seconds)
663 * $cache::TTL_DAY,
664 * // Function that derives the new key value
665 * function ( $oldValue, &$ttl, array &$setOpts ) {
666 * $dbr = wfGetDB( DB_SLAVE );
667 * // Account for any snapshot/slave lag
668 * $setOpts += Database::getCacheSetOptions( $dbr );
669 *
670 * return CatConfig::newFromRow( $dbr->selectRow( ... ) );
671 * },
672 * array(
673 * // Calling touchCheckKey() on this key invalidates the cache
674 * 'checkKeys' => array( $cache->makeKey( 'site-cat-config' ) ),
675 * // Try to only let one datacenter thread manage cache updates at a time
676 * 'lockTSE' => 30
677 * )
678 * );
679 * @endcode
680 *
681 * Example usage (key with dynamic dependencies):
682 * @code
683 * $catState = $cache->getWithSetCallback(
684 * // Key to store the cached value under
685 * $cache->makeKey( 'cat-state', $cat->getId() ),
686 * // Time-to-live (seconds)
687 * $cache::TTL_HOUR,
688 * // Function that derives the new key value
689 * function ( $oldValue, &$ttl, array &$setOpts ) {
690 * // Determine new value from the DB
691 * $dbr = wfGetDB( DB_SLAVE );
692 * // Account for any snapshot/slave lag
693 * $setOpts += Database::getCacheSetOptions( $dbr );
694 *
695 * return CatState::newFromResults( $dbr->select( ... ) );
696 * },
697 * array(
698 * // The "check" keys that represent things the value depends on;
699 * // Calling touchCheckKey() on any of them invalidates the cache
700 * 'checkKeys' => array(
701 * $cache->makeKey( 'sustenance-bowls', $cat->getRoomId() ),
702 * $cache->makeKey( 'people-present', $cat->getHouseId() ),
703 * $cache->makeKey( 'cat-laws', $cat->getCityId() ),
704 * )
705 * )
706 * );
707 * @endcode
708 *
709 * Example usage (hot key holding most recent 100 events):
710 * @code
711 * $lastCatActions = $cache->getWithSetCallback(
712 * // Key to store the cached value under
713 * $cache->makeKey( 'cat-last-actions', 100 ),
714 * // Time-to-live (in seconds)
715 * 10,
716 * // Function that derives the new key value
717 * function ( $oldValue, &$ttl, array &$setOpts ) {
718 * $dbr = wfGetDB( DB_SLAVE );
719 * // Account for any snapshot/slave lag
720 * $setOpts += Database::getCacheSetOptions( $dbr );
721 *
722 * // Start off with the last cached list
723 * $list = $oldValue ?: array();
724 * // Fetch the last 100 relevant rows in descending order;
725 * // only fetch rows newer than $list[0] to reduce scanning
726 * $rows = iterator_to_array( $dbr->select( ... ) );
727 * // Merge them and get the new "last 100" rows
728 * return array_slice( array_merge( $new, $list ), 0, 100 );
729 * },
730 * // Try to only let one datacenter thread manage cache updates at a time
731 * array( 'lockTSE' => 30 )
732 * );
733 * @endcode
734 *
735 * @see WANObjectCache::get()
736 * @see WANObjectCache::set()
737 *
738 * @param string $key Cache key
739 * @param integer $ttl Seconds to live for key updates. Special values are:
740 * - WANObjectCache::TTL_INDEFINITE: Cache forever
741 * - WANObjectCache::TTL_UNCACHEABLE: Do not cache at all
742 * @param callable $callback Value generation function
743 * @param array $opts Options map:
744 * - checkKeys: List of "check" keys. The key at $key will be seen as invalid when either
745 * touchCheckKey() or resetCheckKey() is called on any of these keys.
746 * - lowTTL: Consider pre-emptive updates when the current TTL (sec) of the key is less than
747 * this. It becomes more likely over time, becoming a certainty once the key is expired.
748 * Default: WANObjectCache::LOW_TTL seconds.
749 * - lockTSE: If the key is tombstoned or expired (by checkKeys) less than this many seconds
750 * ago, then try to have a single thread handle cache regeneration at any given time.
751 * Other threads will try to use stale values if possible. If, on miss, the time since
752 * expiration is low, the assumption is that the key is hot and that a stampede is worth
753 * avoiding. Setting this above WANObjectCache::HOLDOFF_TTL makes no difference. The
754 * higher this is set, the higher the worst-case staleness can be.
755 * Use WANObjectCache::TSE_NONE to disable this logic.
756 * Default: WANObjectCache::TSE_NONE.
757 * - pcTTL : process cache the value in this PHP instance with this TTL. This avoids
758 * network I/O when a key is read several times. This will not cache if the callback
759 * returns false however. Note that any purges will not be seen while process cached;
760 * since the callback should use slave DBs and they may be lagged or have snapshot
761 * isolation anyway, this should not typically matter.
762 * Default: WANObjectCache::TTL_UNCACHEABLE.
763 * @return mixed Value to use for the key
764 */
765 final public function getWithSetCallback( $key, $ttl, $callback, array $opts = array() ) {
766 $pcTTL = isset( $opts['pcTTL'] ) ? $opts['pcTTL'] : self::TTL_UNCACHEABLE;
767
768 // Try the process cache if enabled
769 $value = ( $pcTTL >= 0 ) ? $this->procCache->get( $key ) : false;
770
771 if ( $value === false ) {
772 // Fetch the value over the network
773 $value = $this->doGetWithSetCallback( $key, $ttl, $callback, $opts );
774 // Update the process cache if enabled
775 if ( $pcTTL >= 0 && $value !== false ) {
776 $this->procCache->set( $key, $value, $pcTTL );
777 }
778 }
779
780 return $value;
781 }
782
783 /**
784 * Do the actual I/O for getWithSetCallback() when needed
785 *
786 * @see WANObjectCache::getWithSetCallback()
787 *
788 * @param string $key
789 * @param integer $ttl
790 * @param callback $callback
791 * @param array $opts
792 * @return mixed
793 */
794 protected function doGetWithSetCallback( $key, $ttl, $callback, array $opts ) {
795 $lowTTL = isset( $opts['lowTTL'] ) ? $opts['lowTTL'] : min( self::LOW_TTL, $ttl );
796 $lockTSE = isset( $opts['lockTSE'] ) ? $opts['lockTSE'] : self::TSE_NONE;
797 $checkKeys = isset( $opts['checkKeys'] ) ? $opts['checkKeys'] : array();
798
799 // Get the current key value
800 $curTTL = null;
801 $cValue = $this->get( $key, $curTTL, $checkKeys ); // current value
802 $value = $cValue; // return value
803
804 // Determine if a regeneration is desired
805 if ( $value !== false && $curTTL > 0 && !$this->worthRefresh( $curTTL, $lowTTL ) ) {
806 return $value;
807 }
808
809 // A deleted key with a negative TTL left must be tombstoned
810 $isTombstone = ( $curTTL !== null && $value === false );
811 // Assume a key is hot if requested soon after invalidation
812 $isHot = ( $curTTL !== null && $curTTL <= 0 && abs( $curTTL ) <= $lockTSE );
813 // Decide whether a single thread should handle regenerations.
814 // This avoids stampedes when $checkKeys are bumped and when preemptive
815 // renegerations take too long. It also reduces regenerations while $key
816 // is tombstoned. This balances cache freshness with avoiding DB load.
817 $useMutex = ( $isHot || ( $isTombstone && $lockTSE > 0 ) );
818
819 $lockAcquired = false;
820 if ( $useMutex ) {
821 // Acquire a datacenter-local non-blocking lock
822 if ( $this->cache->lock( $key, 0, self::LOCK_TTL ) ) {
823 // Lock acquired; this thread should update the key
824 $lockAcquired = true;
825 } elseif ( $value !== false ) {
826 // If it cannot be acquired; then the stale value can be used
827 return $value;
828 } else {
829 // Use the stash value for tombstoned keys to reduce regeneration load.
830 // For hot keys, either another thread has the lock or the lock failed;
831 // use the stash value from the last thread that regenerated it.
832 $value = $this->cache->get( self::STASH_KEY_PREFIX . $key );
833 if ( $value !== false ) {
834 return $value;
835 }
836 }
837 }
838
839 if ( !is_callable( $callback ) ) {
840 throw new InvalidArgumentException( "Invalid cache miss callback provided." );
841 }
842
843 // Generate the new value from the callback...
844 $setOpts = array();
845 $value = call_user_func_array( $callback, array( $cValue, &$ttl, &$setOpts ) );
846 // When delete() is called, writes are write-holed by the tombstone,
847 // so use a special stash key to pass the new value around threads.
848 if ( $useMutex && $value !== false && $ttl >= 0 ) {
849 $tempTTL = max( 1, (int)$lockTSE ); // set() expects seconds
850 $this->cache->set( self::STASH_KEY_PREFIX . $key, $value, $tempTTL );
851 }
852
853 if ( $lockAcquired ) {
854 $this->cache->unlock( $key );
855 }
856
857 if ( $value !== false && $ttl >= 0 ) {
858 // Update the cache; this will fail if the key is tombstoned
859 $setOpts['lockTSE'] = $lockTSE;
860 $this->set( $key, $value, $ttl, $setOpts );
861 }
862
863 return $value;
864 }
865
866 /**
867 * @see BagOStuff::makeKey()
868 * @param string ... Key component
869 * @return string
870 * @since 1.27
871 */
872 public function makeKey() {
873 return call_user_func_array( array( $this->cache, __FUNCTION__ ), func_get_args() );
874 }
875
876 /**
877 * @see BagOStuff::makeGlobalKey()
878 * @param string ... Key component
879 * @return string
880 * @since 1.27
881 */
882 public function makeGlobalKey() {
883 return call_user_func_array( array( $this->cache, __FUNCTION__ ), func_get_args() );
884 }
885
886 /**
887 * Get the "last error" registered; clearLastError() should be called manually
888 * @return int ERR_* constant for the "last error" registry
889 */
890 final public function getLastError() {
891 if ( $this->lastRelayError ) {
892 // If the cache and the relayer failed, focus on the later.
893 // An update not making it to the relayer means it won't show up
894 // in other DCs (nor will consistent re-hashing see up-to-date values).
895 // On the other hand, if just the cache update failed, then it should
896 // eventually be applied by the relayer.
897 return $this->lastRelayError;
898 }
899
900 $code = $this->cache->getLastError();
901 switch ( $code ) {
902 case BagOStuff::ERR_NONE:
903 return self::ERR_NONE;
904 case BagOStuff::ERR_NO_RESPONSE:
905 return self::ERR_NO_RESPONSE;
906 case BagOStuff::ERR_UNREACHABLE:
907 return self::ERR_UNREACHABLE;
908 default:
909 return self::ERR_UNEXPECTED;
910 }
911 }
912
913 /**
914 * Clear the "last error" registry
915 */
916 final public function clearLastError() {
917 $this->cache->clearLastError();
918 $this->lastRelayError = self::ERR_NONE;
919 }
920
921 /**
922 * Do the actual async bus purge of a key
923 *
924 * This must set the key to "PURGED:<UNIX timestamp>:<holdoff>"
925 *
926 * @param string $key Cache key
927 * @param integer $ttl How long to keep the tombstone [seconds]
928 * @param integer $holdoff HOLDOFF_* constant controlling how long to ignore sets for this key
929 * @return bool Success
930 */
931 protected function relayPurge( $key, $ttl, $holdoff ) {
932 $event = $this->cache->modifySimpleRelayEvent( array(
933 'cmd' => 'set',
934 'key' => $key,
935 'val' => 'PURGED:$UNIXTIME$:' . (int)$holdoff,
936 'ttl' => max( $ttl, 1 ),
937 'sbt' => true, // substitute $UNIXTIME$ with actual microtime
938 ) );
939
940 $ok = $this->relayer->notify( "{$this->pool}:purge", $event );
941 if ( !$ok ) {
942 $this->lastRelayError = self::ERR_RELAY;
943 }
944
945 return $ok;
946 }
947
948 /**
949 * Do the actual async bus delete of a key
950 *
951 * @param string $key Cache key
952 * @return bool Success
953 */
954 protected function relayDelete( $key ) {
955 $event = $this->cache->modifySimpleRelayEvent( array(
956 'cmd' => 'delete',
957 'key' => $key,
958 ) );
959
960 $ok = $this->relayer->notify( "{$this->pool}:purge", $event );
961 if ( !$ok ) {
962 $this->lastRelayError = self::ERR_RELAY;
963 }
964
965 return $ok;
966 }
967
968 /**
969 * Check if a key should be regenerated (using random probability)
970 *
971 * This returns false if $curTTL >= $lowTTL. Otherwise, the chance
972 * of returning true increases steadily from 0% to 100% as the $curTTL
973 * moves from $lowTTL to 0 seconds. This handles widely varying
974 * levels of cache access traffic.
975 *
976 * @param float $curTTL Approximate TTL left on the key if present
977 * @param float $lowTTL Consider a refresh when $curTTL is less than this
978 * @return bool
979 */
980 protected function worthRefresh( $curTTL, $lowTTL ) {
981 if ( $curTTL >= $lowTTL ) {
982 return false;
983 } elseif ( $curTTL <= 0 ) {
984 return true;
985 }
986
987 $chance = ( 1 - $curTTL / $lowTTL );
988
989 return mt_rand( 1, 1e9 ) <= 1e9 * $chance;
990 }
991
992 /**
993 * Do not use this method outside WANObjectCache
994 *
995 * @param mixed $value
996 * @param integer $ttl [0=forever]
997 * @return array
998 */
999 protected function wrap( $value, $ttl ) {
1000 return array(
1001 self::FLD_VERSION => self::VERSION,
1002 self::FLD_VALUE => $value,
1003 self::FLD_TTL => $ttl,
1004 self::FLD_TIME => microtime( true )
1005 );
1006 }
1007
1008 /**
1009 * Do not use this method outside WANObjectCache
1010 *
1011 * @param array|string|bool $wrapped
1012 * @param float $now Unix Current timestamp (preferrable pre-query)
1013 * @return array (mixed; false if absent/invalid, current time left)
1014 */
1015 protected function unwrap( $wrapped, $now ) {
1016 // Check if the value is a tombstone
1017 $purge = self::parsePurgeValue( $wrapped );
1018 if ( $purge !== false ) {
1019 // Purged values should always have a negative current $ttl
1020 $curTTL = min( $purge[self::FLD_TIME] - $now, self::TINY_NEGATIVE );
1021 return array( false, $curTTL );
1022 }
1023
1024 if ( !is_array( $wrapped ) // not found
1025 || !isset( $wrapped[self::FLD_VERSION] ) // wrong format
1026 || $wrapped[self::FLD_VERSION] !== self::VERSION // wrong version
1027 ) {
1028 return array( false, null );
1029 }
1030
1031 $flags = isset( $wrapped[self::FLD_FLAGS] ) ? $wrapped[self::FLD_FLAGS] : 0;
1032 if ( ( $flags & self::FLG_STALE ) == self::FLG_STALE ) {
1033 // Treat as expired, with the cache time as the expiration
1034 $age = $now - $wrapped[self::FLD_TIME];
1035 $curTTL = min( -$age, self::TINY_NEGATIVE );
1036 } elseif ( $wrapped[self::FLD_TTL] > 0 ) {
1037 // Get the approximate time left on the key
1038 $age = $now - $wrapped[self::FLD_TIME];
1039 $curTTL = max( $wrapped[self::FLD_TTL] - $age, 0.0 );
1040 } else {
1041 // Key had no TTL, so the time left is unbounded
1042 $curTTL = INF;
1043 }
1044
1045 return array( $wrapped[self::FLD_VALUE], $curTTL );
1046 }
1047
1048 /**
1049 * @param array $keys
1050 * @param string $prefix
1051 * @return string[]
1052 */
1053 protected static function prefixCacheKeys( array $keys, $prefix ) {
1054 $res = array();
1055 foreach ( $keys as $key ) {
1056 $res[] = $prefix . $key;
1057 }
1058
1059 return $res;
1060 }
1061
1062 /**
1063 * @param string $value Wrapped value like "PURGED:<timestamp>:<holdoff>"
1064 * @return array|bool Array containing a UNIX timestamp (float) and holdoff period (integer),
1065 * or false if value isn't a valid purge value
1066 */
1067 protected static function parsePurgeValue( $value ) {
1068 if ( !is_string( $value ) ) {
1069 return false;
1070 }
1071 $segments = explode( ':', $value, 3 );
1072 if ( !isset( $segments[0] ) || !isset( $segments[1] )
1073 || "{$segments[0]}:" !== self::PURGE_VAL_PREFIX
1074 ) {
1075 return false;
1076 }
1077 if ( !isset( $segments[2] ) ) {
1078 // Back-compat with old purge values without holdoff
1079 $segments[2] = self::HOLDOFF_TTL;
1080 }
1081 return array(
1082 self::FLD_TIME => (float)$segments[1],
1083 self::FLD_HOLDOFF => (int)$segments[2],
1084 );
1085 }
1086
1087 /**
1088 * @param float $timestamp
1089 * @param int $holdoff In seconds
1090 * @return string Wrapped purge value
1091 */
1092 protected function makePurgeValue( $timestamp, $holdoff ) {
1093 return self::PURGE_VAL_PREFIX . (float)$timestamp . ':' . (int)$holdoff;
1094 }
1095 }