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