Merge "Defer the redirect table update in WikiPage::insertRedirect()"
[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 $now = microtime( true );
255
256 // Collect timestamps from all "check" keys
257 $purgeValuesForAll = $this->processCheckKeys( $checkKeysForAll, $wrappedValues, $now );
258 $purgeValuesByKey = array();
259 foreach ( $checkKeysByKey as $cacheKey => $checks ) {
260 $purgeValuesByKey[$cacheKey] =
261 $this->processCheckKeys( $checks, $wrappedValues, $now );
262 }
263
264 // Get the main cache value for each key and validate them
265 foreach ( $valueKeys as $vKey ) {
266 if ( !isset( $wrappedValues[$vKey] ) ) {
267 continue; // not found
268 }
269
270 $key = substr( $vKey, $vPrefixLen ); // unprefix
271
272 list( $value, $curTTL ) = $this->unwrap( $wrappedValues[$vKey], $now );
273 if ( $value !== false ) {
274 $result[$key] = $value;
275
276 // Force dependant keys to be invalid for a while after purging
277 // to reduce race conditions involving stale data getting cached
278 $purgeValues = $purgeValuesForAll;
279 if ( isset( $purgeValuesByKey[$key] ) ) {
280 $purgeValues = array_merge( $purgeValues, $purgeValuesByKey[$key] );
281 }
282 foreach ( $purgeValues as $purge ) {
283 $safeTimestamp = $purge[self::FLD_TIME] + $purge[self::FLD_HOLDOFF];
284 if ( $safeTimestamp >= $wrappedValues[$vKey][self::FLD_TIME] ) {
285 $curTTL = min( $curTTL, $purge[self::FLD_TIME] - $now );
286 }
287 }
288 }
289 $curTTLs[$key] = $curTTL;
290 }
291
292 return $result;
293 }
294
295 /**
296 * @since 1.27
297 * @param array $timeKeys List of prefixed time check keys
298 * @param array $wrappedValues
299 * @param float $now
300 * @return array List of purge value arrays
301 */
302 private function processCheckKeys( array $timeKeys, array $wrappedValues, $now ) {
303 $purgeValues = array();
304 foreach ( $timeKeys as $timeKey ) {
305 $purge = isset( $wrappedValues[$timeKey] )
306 ? self::parsePurgeValue( $wrappedValues[$timeKey] )
307 : false;
308 if ( $purge === false ) {
309 // Key is not set or invalid; regenerate
310 $this->cache->add( $timeKey,
311 $this->makePurgeValue( $now, self::HOLDOFF_TTL ),
312 self::CHECK_KEY_TTL
313 );
314 $purge = array( self::FLD_TIME => $now, self::FLD_HOLDOFF => self::HOLDOFF_TTL );
315 }
316 $purgeValues[] = $purge;
317 }
318 return $purgeValues;
319 }
320
321 /**
322 * Set the value of a key in cache
323 *
324 * Simply calling this method when source data changes is not valid because
325 * the changes do not replicate to the other WAN sites. In that case, delete()
326 * should be used instead. This method is intended for use on cache misses.
327 *
328 * If the data was read from a snapshot-isolated transactions (e.g. the default
329 * REPEATABLE-READ in innoDB), use 'since' to avoid the following race condition:
330 * - a) T1 starts
331 * - b) T2 updates a row, calls delete(), and commits
332 * - c) The HOLDOFF_TTL passes, expiring the delete() tombstone
333 * - d) T1 reads the row and calls set() due to a cache miss
334 * - e) Stale value is stuck in cache
335 *
336 * Setting 'lag' and 'since' help avoids keys getting stuck in stale states.
337 *
338 * Example usage:
339 * @code
340 * $dbr = wfGetDB( DB_SLAVE );
341 * $setOpts = Database::getCacheSetOptions( $dbr );
342 * // Fetch the row from the DB
343 * $row = $dbr->selectRow( ... );
344 * $key = $cache->makeKey( 'building', $buildingId );
345 * $cache->set( $key, $row, $cache::TTL_DAY, $setOpts );
346 * @endcode
347 *
348 * @param string $key Cache key
349 * @param mixed $value
350 * @param integer $ttl Seconds to live. Special values are:
351 * - WANObjectCache::TTL_INDEFINITE: Cache forever
352 * @param array $opts Options map:
353 * - lag : Seconds of slave lag. Typically, this is either the slave lag
354 * before the data was read or, if applicable, the slave lag before
355 * the snapshot-isolated transaction the data was read from started.
356 * Default: 0 seconds
357 * - since : UNIX timestamp of the data in $value. Typically, this is either
358 * the current time the data was read or (if applicable) the time when
359 * the snapshot-isolated transaction the data was read from started.
360 * Default: 0 seconds
361 * - pending : Whether this data is possibly from an uncommitted write transaction.
362 * Generally, other threads should not see values from the future and
363 * they certainly should not see ones that ended up getting rolled back.
364 * Default: false
365 * - lockTSE : if excessive replication/snapshot lag is detected, then store the value
366 * with this TTL and flag it as stale. This is only useful if the reads for
367 * this key use getWithSetCallback() with "lockTSE" set.
368 * Default: WANObjectCache::TSE_NONE
369 * @return bool Success
370 */
371 final public function set( $key, $value, $ttl = 0, array $opts = array() ) {
372 $lockTSE = isset( $opts['lockTSE'] ) ? $opts['lockTSE'] : self::TSE_NONE;
373 $age = isset( $opts['since'] ) ? max( 0, microtime( true ) - $opts['since'] ) : 0;
374 $lag = isset( $opts['lag'] ) ? $opts['lag'] : 0;
375
376 // Do not cache potentially uncommitted data as it might get rolled back
377 if ( !empty( $opts['pending'] ) ) {
378 $this->logger->info( "Rejected set() for $key due to pending writes." );
379
380 return true; // no-op the write for being unsafe
381 }
382
383 $wrapExtra = array(); // additional wrapped value fields
384 // Check if there's a risk of writing stale data after the purge tombstone expired
385 if ( $lag === false || ( $lag + $age ) > self::MAX_READ_LAG ) {
386 // Case A: read lag with "lockTSE"; save but record value as stale
387 if ( $lockTSE >= 0 ) {
388 $ttl = max( 1, (int)$lockTSE ); // set() expects seconds
389 $wrapExtra[self::FLD_FLAGS] = self::FLG_STALE; // mark as stale
390 // Case B: any long-running transaction; ignore this set()
391 } elseif ( $age > self::MAX_READ_LAG ) {
392 $this->logger->warning( "Rejected set() for $key due to snapshot lag." );
393
394 return true; // no-op the write for being unsafe
395 // Case C: high replication lag; lower TTL instead of ignoring all set()s
396 } elseif ( $lag === false || $lag > self::MAX_READ_LAG ) {
397 $ttl = $ttl ? min( $ttl, self::TTL_LAGGED ) : self::TTL_LAGGED;
398 $this->logger->warning( "Lowered set() TTL for $key due to replication lag." );
399 // Case D: medium length request with medium replication lag; ignore this set()
400 } else {
401 $this->logger->warning( "Rejected set() for $key due to high read lag." );
402
403 return true; // no-op the write for being unsafe
404 }
405 }
406
407 // Wrap that value with time/TTL/version metadata
408 $wrapped = $this->wrap( $value, $ttl ) + $wrapExtra;
409
410 $func = function ( $cache, $key, $cWrapped ) use ( $wrapped ) {
411 return ( is_string( $cWrapped ) )
412 ? false // key is tombstoned; do nothing
413 : $wrapped;
414 };
415
416 return $this->cache->merge( self::VALUE_KEY_PREFIX . $key, $func, $ttl, 1 );
417 }
418
419 /**
420 * Purge a key from all datacenters
421 *
422 * This should only be called when the underlying data (being cached)
423 * changes in a significant way. This deletes the key and starts a hold-off
424 * period where the key cannot be written to for a few seconds (HOLDOFF_TTL).
425 * This is done to avoid the following race condition:
426 * - a) Some DB data changes and delete() is called on a corresponding key
427 * - b) A request refills the key with a stale value from a lagged DB
428 * - c) The stale value is stuck there until the key is expired/evicted
429 *
430 * This is implemented by storing a special "tombstone" value at the cache
431 * key that this class recognizes; get() calls will return false for the key
432 * and any set() calls will refuse to replace tombstone values at the key.
433 * For this to always avoid stale value writes, the following must hold:
434 * - a) Replication lag is bounded to being less than HOLDOFF_TTL; or
435 * - b) If lag is higher, the DB will have gone into read-only mode already
436 *
437 * Note that set() can also be lag-aware and lower the TTL if it's high.
438 *
439 * When using potentially long-running ACID transactions, a good pattern is
440 * to use a pre-commit hook to issue the delete. This means that immediately
441 * after commit, callers will see the tombstone in cache in the local datacenter
442 * and in the others upon relay. It also avoids the following race condition:
443 * - a) T1 begins, changes a row, and calls delete()
444 * - b) The HOLDOFF_TTL passes, expiring the delete() tombstone
445 * - c) T2 starts, reads the row and calls set() due to a cache miss
446 * - d) T1 finally commits
447 * - e) Stale value is stuck in cache
448 *
449 * Example usage:
450 * @code
451 * $dbw->begin(); // start of request
452 * ... <execute some stuff> ...
453 * // Update the row in the DB
454 * $dbw->update( ... );
455 * $key = $cache->makeKey( 'homes', $homeId );
456 * // Purge the corresponding cache entry just before committing
457 * $dbw->onTransactionPreCommitOrIdle( function() use ( $cache, $key ) {
458 * $cache->delete( $key );
459 * } );
460 * ... <execute some stuff> ...
461 * $dbw->commit(); // end of request
462 * @endcode
463 *
464 * The $ttl parameter can be used when purging values that have not actually changed
465 * recently. For example, a cleanup script to purge cache entries does not really need
466 * a hold-off period, so it can use HOLDOFF_NONE. Likewise for user-requested purge.
467 * Note that $ttl limits the effective range of 'lockTSE' for getWithSetCallback().
468 *
469 * If called twice on the same key, then the last hold-off TTL takes precedence. For
470 * idempotence, the $ttl should not vary for different delete() calls on the same key.
471 *
472 * @param string $key Cache key
473 * @param integer $ttl Tombstone TTL; Default: WANObjectCache::HOLDOFF_TTL
474 * @return bool True if the item was purged or not found, false on failure
475 */
476 final public function delete( $key, $ttl = self::HOLDOFF_TTL ) {
477 $key = self::VALUE_KEY_PREFIX . $key;
478
479 if ( $ttl <= 0 ) {
480 // Update the local datacenter immediately
481 $ok = $this->cache->delete( $key );
482 // Publish the purge to all datacenters
483 $ok = $this->relayDelete( $key ) && $ok;
484 } else {
485 // Update the local datacenter immediately
486 $ok = $this->cache->set( $key,
487 $this->makePurgeValue( microtime( true ), self::HOLDOFF_NONE ),
488 $ttl
489 );
490 // Publish the purge to all datacenters
491 $ok = $this->relayPurge( $key, $ttl, self::HOLDOFF_NONE ) && $ok;
492 }
493
494 return $ok;
495 }
496
497 /**
498 * Fetch the value of a timestamp "check" key
499 *
500 * The key will be *initialized* to the current time if not set,
501 * so only call this method if this behavior is actually desired
502 *
503 * The timestamp can be used to check whether a cached value is valid.
504 * Callers should not assume that this returns the same timestamp in
505 * all datacenters due to relay delays.
506 *
507 * The level of staleness can roughly be estimated from this key, but
508 * if the key was evicted from cache, such calculations may show the
509 * time since expiry as ~0 seconds.
510 *
511 * Note that "check" keys won't collide with other regular keys.
512 *
513 * @param string $key
514 * @return float UNIX timestamp of the check key
515 */
516 final public function getCheckKeyTime( $key ) {
517 $key = self::TIME_KEY_PREFIX . $key;
518
519 $purge = self::parsePurgeValue( $this->cache->get( $key ) );
520 if ( $purge !== false ) {
521 $time = $purge[self::FLD_TIME];
522 } else {
523 // Casting assures identical floats for the next getCheckKeyTime() calls
524 $now = (string)microtime( true );
525 $this->cache->add( $key,
526 $this->makePurgeValue( $now, self::HOLDOFF_TTL ),
527 self::CHECK_KEY_TTL
528 );
529 $time = (float)$now;
530 }
531
532 return $time;
533 }
534
535 /**
536 * Purge a "check" key from all datacenters, invalidating keys that use it
537 *
538 * This should only be called when the underlying data (being cached)
539 * changes in a significant way, and it is impractical to call delete()
540 * on all keys that should be changed. When get() is called on those
541 * keys, the relevant "check" keys must be supplied for this to work.
542 *
543 * The "check" key essentially represents a last-modified field.
544 * When touched, keys using it via get(), getMulti(), or getWithSetCallback()
545 * will be invalidated. It is treated as being HOLDOFF_TTL seconds in the future
546 * by those methods to avoid race conditions where dependent keys get updated
547 * with stale values (e.g. from a DB slave).
548 *
549 * This is typically useful for keys with hardcoded names or in some cases
550 * dynamically generated names where a low number of combinations exist.
551 * When a few important keys get a large number of hits, a high cache
552 * time is usually desired as well as "lockTSE" logic. The resetCheckKey()
553 * method is less appropriate in such cases since the "time since expiry"
554 * cannot be inferred.
555 *
556 * Note that "check" keys won't collide with other regular keys.
557 *
558 * @see WANObjectCache::get()
559 * @see WANObjectCache::getWithSetCallback()
560 * @see WANObjectCache::resetCheckKey()
561 *
562 * @param string $key Cache key
563 * @param int $holdoff HOLDOFF_TTL or HOLDOFF_NONE constant
564 * @return bool True if the item was purged or not found, false on failure
565 */
566 final public function touchCheckKey( $key, $holdoff = self::HOLDOFF_TTL ) {
567 $key = self::TIME_KEY_PREFIX . $key;
568 // Update the local datacenter immediately
569 $ok = $this->cache->set( $key,
570 $this->makePurgeValue( microtime( true ), $holdoff ),
571 self::CHECK_KEY_TTL
572 );
573 // Publish the purge to all datacenters
574 return $this->relayPurge( $key, self::CHECK_KEY_TTL, $holdoff ) && $ok;
575 }
576
577 /**
578 * Delete a "check" key from all datacenters, invalidating keys that use it
579 *
580 * This is similar to touchCheckKey() in that keys using it via get(), getMulti(),
581 * or getWithSetCallback() will be invalidated. The differences are:
582 * - a) The timestamp will be deleted from all caches and lazily
583 * re-initialized when accessed (rather than set everywhere)
584 * - b) Thus, dependent keys will be known to be invalid, but not
585 * for how long (they are treated as "just" purged), which
586 * effects any lockTSE logic in getWithSetCallback()
587 *
588 * The advantage is that this does not place high TTL keys on every cache
589 * server, making it better for code that will cache many different keys
590 * and either does not use lockTSE or uses a low enough TTL anyway.
591 *
592 * This is typically useful for keys with dynamically generated names
593 * where a high number of combinations exist.
594 *
595 * Note that "check" keys won't collide with other regular keys.
596 *
597 * @see WANObjectCache::get()
598 * @see WANObjectCache::getWithSetCallback()
599 * @see WANObjectCache::touchCheckKey()
600 *
601 * @param string $key Cache key
602 * @return bool True if the item was purged or not found, false on failure
603 */
604 final public function resetCheckKey( $key ) {
605 $key = self::TIME_KEY_PREFIX . $key;
606 // Update the local datacenter immediately
607 $ok = $this->cache->delete( $key );
608 // Publish the purge to all datacenters
609 return $this->relayDelete( $key ) && $ok;
610 }
611
612 /**
613 * Method to fetch/regenerate cache keys
614 *
615 * On cache miss, the key will be set to the callback result via set()
616 * (unless the callback returns false) and that result will be returned.
617 * The arguments supplied to the callback are:
618 * - $oldValue : current cache value or false if not present
619 * - &$ttl : a reference to the TTL which can be altered
620 * - &$setOpts : a reference to options for set() which can be altered
621 *
622 * It is strongly recommended to set the 'lag' and 'since' fields to avoid race conditions
623 * that can cause stale values to get stuck at keys. Usually, callbacks ignore the current
624 * value, but it can be used to maintain "most recent X" values that come from time or
625 * sequence based source data, provided that the "as of" id/time is tracked. Note that
626 * preemptive regeneration and $checkKeys can result in a non-false current value.
627 *
628 * Usage of $checkKeys is similar to get() and getMulti(). However, rather than the caller
629 * having to inspect a "current time left" variable (e.g. $curTTL, $curTTLs), a cache
630 * regeneration will automatically be triggered using the callback.
631 *
632 * The simplest way to avoid stampedes for hot keys is to use
633 * the 'lockTSE' option in $opts. If cache purges are needed, also:
634 * - a) Pass $key into $checkKeys
635 * - b) Use touchCheckKey( $key ) instead of delete( $key )
636 *
637 * Example usage (typical key):
638 * @code
639 * $catInfo = $cache->getWithSetCallback(
640 * // Key to store the cached value under
641 * $cache->makeKey( 'cat-attributes', $catId ),
642 * // Time-to-live (in seconds)
643 * $cache::TTL_MINUTE,
644 * // Function that derives the new key value
645 * function ( $oldValue, &$ttl, array &$setOpts ) {
646 * $dbr = wfGetDB( DB_SLAVE );
647 * // Account for any snapshot/slave lag
648 * $setOpts += Database::getCacheSetOptions( $dbr );
649 *
650 * return $dbr->selectRow( ... );
651 * }
652 * );
653 * @endcode
654 *
655 * Example usage (key that is expensive and hot):
656 * @code
657 * $catConfig = $cache->getWithSetCallback(
658 * // Key to store the cached value under
659 * $cache->makeKey( 'site-cat-config' ),
660 * // Time-to-live (in seconds)
661 * $cache::TTL_DAY,
662 * // Function that derives the new key value
663 * function ( $oldValue, &$ttl, array &$setOpts ) {
664 * $dbr = wfGetDB( DB_SLAVE );
665 * // Account for any snapshot/slave lag
666 * $setOpts += Database::getCacheSetOptions( $dbr );
667 *
668 * return CatConfig::newFromRow( $dbr->selectRow( ... ) );
669 * },
670 * array(
671 * // Calling touchCheckKey() on this key invalidates the cache
672 * 'checkKeys' => array( $cache->makeKey( 'site-cat-config' ) ),
673 * // Try to only let one datacenter thread manage cache updates at a time
674 * 'lockTSE' => 30
675 * )
676 * );
677 * @endcode
678 *
679 * Example usage (key with dynamic dependencies):
680 * @code
681 * $catState = $cache->getWithSetCallback(
682 * // Key to store the cached value under
683 * $cache->makeKey( 'cat-state', $cat->getId() ),
684 * // Time-to-live (seconds)
685 * $cache::TTL_HOUR,
686 * // Function that derives the new key value
687 * function ( $oldValue, &$ttl, array &$setOpts ) {
688 * // Determine new value from the DB
689 * $dbr = wfGetDB( DB_SLAVE );
690 * // Account for any snapshot/slave lag
691 * $setOpts += Database::getCacheSetOptions( $dbr );
692 *
693 * return CatState::newFromResults( $dbr->select( ... ) );
694 * },
695 * array(
696 * // The "check" keys that represent things the value depends on;
697 * // Calling touchCheckKey() on any of them invalidates the cache
698 * 'checkKeys' => array(
699 * $cache->makeKey( 'sustenance-bowls', $cat->getRoomId() ),
700 * $cache->makeKey( 'people-present', $cat->getHouseId() ),
701 * $cache->makeKey( 'cat-laws', $cat->getCityId() ),
702 * )
703 * )
704 * );
705 * @endcode
706 *
707 * Example usage (hot key holding most recent 100 events):
708 * @code
709 * $lastCatActions = $cache->getWithSetCallback(
710 * // Key to store the cached value under
711 * $cache->makeKey( 'cat-last-actions', 100 ),
712 * // Time-to-live (in seconds)
713 * 10,
714 * // Function that derives the new key value
715 * function ( $oldValue, &$ttl, array &$setOpts ) {
716 * $dbr = wfGetDB( DB_SLAVE );
717 * // Account for any snapshot/slave lag
718 * $setOpts += Database::getCacheSetOptions( $dbr );
719 *
720 * // Start off with the last cached list
721 * $list = $oldValue ?: array();
722 * // Fetch the last 100 relevant rows in descending order;
723 * // only fetch rows newer than $list[0] to reduce scanning
724 * $rows = iterator_to_array( $dbr->select( ... ) );
725 * // Merge them and get the new "last 100" rows
726 * return array_slice( array_merge( $new, $list ), 0, 100 );
727 * },
728 * // Try to only let one datacenter thread manage cache updates at a time
729 * array( 'lockTSE' => 30 )
730 * );
731 * @endcode
732 *
733 * @see WANObjectCache::get()
734 * @see WANObjectCache::set()
735 *
736 * @param string $key Cache key
737 * @param integer $ttl Seconds to live for key updates. Special values are:
738 * - WANObjectCache::TTL_INDEFINITE: Cache forever
739 * - WANObjectCache::TTL_UNCACHEABLE: Do not cache at all
740 * @param callable $callback Value generation function
741 * @param array $opts Options map:
742 * - checkKeys: List of "check" keys. The key at $key will be seen as invalid when either
743 * touchCheckKey() or resetCheckKey() is called on any of these keys.
744 * - lowTTL: Consider pre-emptive updates when the current TTL (sec) of the key is less than
745 * this. It becomes more likely over time, becoming a certainty once the key is expired.
746 * Default: WANObjectCache::LOW_TTL seconds.
747 * - lockTSE: If the key is tombstoned or expired (by checkKeys) less than this many seconds
748 * ago, then try to have a single thread handle cache regeneration at any given time.
749 * Other threads will try to use stale values if possible. If, on miss, the time since
750 * expiration is low, the assumption is that the key is hot and that a stampede is worth
751 * avoiding. Setting this above WANObjectCache::HOLDOFF_TTL makes no difference. The
752 * higher this is set, the higher the worst-case staleness can be.
753 * Use WANObjectCache::TSE_NONE to disable this logic.
754 * Default: WANObjectCache::TSE_NONE.
755 * - pcTTL : process cache the value in this PHP instance with this TTL. This avoids
756 * network I/O when a key is read several times. This will not cache if the callback
757 * returns false however. Note that any purges will not be seen while process cached;
758 * since the callback should use slave DBs and they may be lagged or have snapshot
759 * isolation anyway, this should not typically matter.
760 * Default: WANObjectCache::TTL_UNCACHEABLE.
761 * @return mixed Value to use for the key
762 */
763 final public function getWithSetCallback( $key, $ttl, $callback, array $opts = array() ) {
764 $pcTTL = isset( $opts['pcTTL'] ) ? $opts['pcTTL'] : self::TTL_UNCACHEABLE;
765
766 // Try the process cache if enabled
767 $value = ( $pcTTL >= 0 ) ? $this->procCache->get( $key ) : false;
768
769 if ( $value === false ) {
770 // Fetch the value over the network
771 $value = $this->doGetWithSetCallback( $key, $ttl, $callback, $opts );
772 // Update the process cache if enabled
773 if ( $pcTTL >= 0 && $value !== false ) {
774 $this->procCache->set( $key, $value, $pcTTL );
775 }
776 }
777
778 return $value;
779 }
780
781 /**
782 * Do the actual I/O for getWithSetCallback() when needed
783 *
784 * @see WANObjectCache::getWithSetCallback()
785 *
786 * @param string $key
787 * @param integer $ttl
788 * @param callback $callback
789 * @param array $opts
790 * @return mixed
791 */
792 protected function doGetWithSetCallback( $key, $ttl, $callback, array $opts ) {
793 $lowTTL = isset( $opts['lowTTL'] ) ? $opts['lowTTL'] : min( self::LOW_TTL, $ttl );
794 $lockTSE = isset( $opts['lockTSE'] ) ? $opts['lockTSE'] : self::TSE_NONE;
795 $checkKeys = isset( $opts['checkKeys'] ) ? $opts['checkKeys'] : array();
796
797 // Get the current key value
798 $curTTL = null;
799 $cValue = $this->get( $key, $curTTL, $checkKeys ); // current value
800 $value = $cValue; // return value
801
802 // Determine if a regeneration is desired
803 if ( $value !== false && $curTTL > 0 && !$this->worthRefresh( $curTTL, $lowTTL ) ) {
804 return $value;
805 }
806
807 // A deleted key with a negative TTL left must be tombstoned
808 $isTombstone = ( $curTTL !== null && $value === false );
809 // Assume a key is hot if requested soon after invalidation
810 $isHot = ( $curTTL !== null && $curTTL <= 0 && abs( $curTTL ) <= $lockTSE );
811 // Decide whether a single thread should handle regenerations.
812 // This avoids stampedes when $checkKeys are bumped and when preemptive
813 // renegerations take too long. It also reduces regenerations while $key
814 // is tombstoned. This balances cache freshness with avoiding DB load.
815 $useMutex = ( $isHot || ( $isTombstone && $lockTSE > 0 ) );
816
817 $lockAcquired = false;
818 if ( $useMutex ) {
819 // Acquire a datacenter-local non-blocking lock
820 if ( $this->cache->lock( $key, 0, self::LOCK_TTL ) ) {
821 // Lock acquired; this thread should update the key
822 $lockAcquired = true;
823 } elseif ( $value !== false ) {
824 // If it cannot be acquired; then the stale value can be used
825 return $value;
826 } else {
827 // Use the stash value for tombstoned keys to reduce regeneration load.
828 // For hot keys, either another thread has the lock or the lock failed;
829 // use the stash value from the last thread that regenerated it.
830 $value = $this->cache->get( self::STASH_KEY_PREFIX . $key );
831 if ( $value !== false ) {
832 return $value;
833 }
834 }
835 }
836
837 if ( !is_callable( $callback ) ) {
838 throw new InvalidArgumentException( "Invalid cache miss callback provided." );
839 }
840
841 // Generate the new value from the callback...
842 $setOpts = array();
843 $value = call_user_func_array( $callback, array( $cValue, &$ttl, &$setOpts ) );
844 // When delete() is called, writes are write-holed by the tombstone,
845 // so use a special stash key to pass the new value around threads.
846 if ( $useMutex && $value !== false && $ttl >= 0 ) {
847 $tempTTL = max( 1, (int)$lockTSE ); // set() expects seconds
848 $this->cache->set( self::STASH_KEY_PREFIX . $key, $value, $tempTTL );
849 }
850
851 if ( $lockAcquired ) {
852 $this->cache->unlock( $key );
853 }
854
855 if ( $value !== false && $ttl >= 0 ) {
856 // Update the cache; this will fail if the key is tombstoned
857 $setOpts['lockTSE'] = $lockTSE;
858 $this->set( $key, $value, $ttl, $setOpts );
859 }
860
861 return $value;
862 }
863
864 /**
865 * @see BagOStuff::makeKey()
866 * @param string ... Key component
867 * @return string
868 * @since 1.27
869 */
870 public function makeKey() {
871 return call_user_func_array( array( $this->cache, __FUNCTION__ ), func_get_args() );
872 }
873
874 /**
875 * @see BagOStuff::makeGlobalKey()
876 * @param string ... Key component
877 * @return string
878 * @since 1.27
879 */
880 public function makeGlobalKey() {
881 return call_user_func_array( array( $this->cache, __FUNCTION__ ), func_get_args() );
882 }
883
884 /**
885 * Get the "last error" registered; clearLastError() should be called manually
886 * @return int ERR_* constant for the "last error" registry
887 */
888 final public function getLastError() {
889 if ( $this->lastRelayError ) {
890 // If the cache and the relayer failed, focus on the later.
891 // An update not making it to the relayer means it won't show up
892 // in other DCs (nor will consistent re-hashing see up-to-date values).
893 // On the other hand, if just the cache update failed, then it should
894 // eventually be applied by the relayer.
895 return $this->lastRelayError;
896 }
897
898 $code = $this->cache->getLastError();
899 switch ( $code ) {
900 case BagOStuff::ERR_NONE:
901 return self::ERR_NONE;
902 case BagOStuff::ERR_NO_RESPONSE:
903 return self::ERR_NO_RESPONSE;
904 case BagOStuff::ERR_UNREACHABLE:
905 return self::ERR_UNREACHABLE;
906 default:
907 return self::ERR_UNEXPECTED;
908 }
909 }
910
911 /**
912 * Clear the "last error" registry
913 */
914 final public function clearLastError() {
915 $this->cache->clearLastError();
916 $this->lastRelayError = self::ERR_NONE;
917 }
918
919 /**
920 * Do the actual async bus purge of a key
921 *
922 * This must set the key to "PURGED:<UNIX timestamp>:<holdoff>"
923 *
924 * @param string $key Cache key
925 * @param integer $ttl How long to keep the tombstone [seconds]
926 * @param integer $holdoff HOLDOFF_* constant controlling how long to ignore sets for this key
927 * @return bool Success
928 */
929 protected function relayPurge( $key, $ttl, $holdoff ) {
930 $event = $this->cache->modifySimpleRelayEvent( array(
931 'cmd' => 'set',
932 'key' => $key,
933 'val' => 'PURGED:$UNIXTIME$:' . (int)$holdoff,
934 'ttl' => max( $ttl, 1 ),
935 'sbt' => true, // substitute $UNIXTIME$ with actual microtime
936 ) );
937
938 $ok = $this->relayer->notify( "{$this->pool}:purge", $event );
939 if ( !$ok ) {
940 $this->lastRelayError = self::ERR_RELAY;
941 }
942
943 return $ok;
944 }
945
946 /**
947 * Do the actual async bus delete of a key
948 *
949 * @param string $key Cache key
950 * @return bool Success
951 */
952 protected function relayDelete( $key ) {
953 $event = $this->cache->modifySimpleRelayEvent( array(
954 'cmd' => 'delete',
955 'key' => $key,
956 ) );
957
958 $ok = $this->relayer->notify( "{$this->pool}:purge", $event );
959 if ( !$ok ) {
960 $this->lastRelayError = self::ERR_RELAY;
961 }
962
963 return $ok;
964 }
965
966 /**
967 * Check if a key should be regenerated (using random probability)
968 *
969 * This returns false if $curTTL >= $lowTTL. Otherwise, the chance
970 * of returning true increases steadily from 0% to 100% as the $curTTL
971 * moves from $lowTTL to 0 seconds. This handles widely varying
972 * levels of cache access traffic.
973 *
974 * @param float $curTTL Approximate TTL left on the key if present
975 * @param float $lowTTL Consider a refresh when $curTTL is less than this
976 * @return bool
977 */
978 protected function worthRefresh( $curTTL, $lowTTL ) {
979 if ( $curTTL >= $lowTTL ) {
980 return false;
981 } elseif ( $curTTL <= 0 ) {
982 return true;
983 }
984
985 $chance = ( 1 - $curTTL / $lowTTL );
986
987 return mt_rand( 1, 1e9 ) <= 1e9 * $chance;
988 }
989
990 /**
991 * Do not use this method outside WANObjectCache
992 *
993 * @param mixed $value
994 * @param integer $ttl [0=forever]
995 * @return array
996 */
997 protected function wrap( $value, $ttl ) {
998 return array(
999 self::FLD_VERSION => self::VERSION,
1000 self::FLD_VALUE => $value,
1001 self::FLD_TTL => $ttl,
1002 self::FLD_TIME => microtime( true )
1003 );
1004 }
1005
1006 /**
1007 * Do not use this method outside WANObjectCache
1008 *
1009 * @param array|string|bool $wrapped
1010 * @param float $now Unix Current timestamp (preferrable pre-query)
1011 * @return array (mixed; false if absent/invalid, current time left)
1012 */
1013 protected function unwrap( $wrapped, $now ) {
1014 // Check if the value is a tombstone
1015 $purge = self::parsePurgeValue( $wrapped );
1016 if ( $purge !== false ) {
1017 // Purged values should always have a negative current $ttl
1018 $curTTL = min( $purge[self::FLD_TIME] - $now, self::TINY_NEGATIVE );
1019 return array( false, $curTTL );
1020 }
1021
1022 if ( !is_array( $wrapped ) // not found
1023 || !isset( $wrapped[self::FLD_VERSION] ) // wrong format
1024 || $wrapped[self::FLD_VERSION] !== self::VERSION // wrong version
1025 ) {
1026 return array( false, null );
1027 }
1028
1029 $flags = isset( $wrapped[self::FLD_FLAGS] ) ? $wrapped[self::FLD_FLAGS] : 0;
1030 if ( ( $flags & self::FLG_STALE ) == self::FLG_STALE ) {
1031 // Treat as expired, with the cache time as the expiration
1032 $age = $now - $wrapped[self::FLD_TIME];
1033 $curTTL = min( -$age, self::TINY_NEGATIVE );
1034 } elseif ( $wrapped[self::FLD_TTL] > 0 ) {
1035 // Get the approximate time left on the key
1036 $age = $now - $wrapped[self::FLD_TIME];
1037 $curTTL = max( $wrapped[self::FLD_TTL] - $age, 0.0 );
1038 } else {
1039 // Key had no TTL, so the time left is unbounded
1040 $curTTL = INF;
1041 }
1042
1043 return array( $wrapped[self::FLD_VALUE], $curTTL );
1044 }
1045
1046 /**
1047 * @param array $keys
1048 * @param string $prefix
1049 * @return string[]
1050 */
1051 protected static function prefixCacheKeys( array $keys, $prefix ) {
1052 $res = array();
1053 foreach ( $keys as $key ) {
1054 $res[] = $prefix . $key;
1055 }
1056
1057 return $res;
1058 }
1059
1060 /**
1061 * @param string $value Wrapped value like "PURGED:<timestamp>:<holdoff>"
1062 * @return array|bool Array containing a UNIX timestamp (float) and holdoff period (integer),
1063 * or false if value isn't a valid purge value
1064 */
1065 protected static function parsePurgeValue( $value ) {
1066 if ( !is_string( $value ) ) {
1067 return false;
1068 }
1069 $segments = explode( ':', $value, 3 );
1070 if ( !isset( $segments[0] ) || !isset( $segments[1] )
1071 || "{$segments[0]}:" !== self::PURGE_VAL_PREFIX
1072 ) {
1073 return false;
1074 }
1075 if ( !isset( $segments[2] ) ) {
1076 // Back-compat with old purge values without holdoff
1077 $segments[2] = self::HOLDOFF_TTL;
1078 }
1079 return array(
1080 self::FLD_TIME => (float)$segments[1],
1081 self::FLD_HOLDOFF => (int)$segments[2],
1082 );
1083 }
1084
1085 /**
1086 * @param float $timestamp
1087 * @param int $holdoff In seconds
1088 * @return string Wrapped purge value
1089 */
1090 protected static function makePurgeValue( $timestamp, $holdoff ) {
1091 return self::PURGE_VAL_PREFIX . (float)$timestamp . ':' . (int)$holdoff;
1092 }
1093 }