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.
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.
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
20 * @author Aaron Schulz
24 * Multi-datacenter aware caching interface
26 * All operations go to the local cache, except the delete()
27 * and touchCheckKey(), which broadcast to all clusters.
28 * This class is intended for caching data from primary stores.
29 * If the get() method does not return a value, then the caller
30 * should query the new value and backfill the cache using set().
31 * When the source data changes, the delete() method should be called.
32 * Since delete() is expensive, it should be avoided. One can do so if:
33 * - a) The object cached is immutable; or
34 * - b) Validity is checked against the source after get(); or
35 * - c) Using a modest TTL is reasonably correct and performant
36 * Consider using getWithSetCallback() instead of the get()/set() cycle.
38 * Instances of this class must be configured to point to a valid
39 * PubSub endpoint, and there must be listeners on the cache servers
40 * that subscribe to the endpoint and update the caches.
42 * Broadcasted operations like delete() and touchCheckKey() are done
43 * synchronously in the local cluster, but are relayed asynchronously.
44 * This means that callers in other datacenters will see older values
45 * for a however many milliseconds the datacenters are apart. As with
46 * any cache, this should not be relied on for cases where reads are
47 * used to determine writes to source (e.g. non-cache) data stores.
49 * All values are wrapped in metadata arrays. Keys use a "WANCache:" prefix
50 * to avoid collisions with keys that are not wrapped as metadata arrays. The
51 * prefixes are as follows:
52 * - a) "WANCache:v" : used for regular value keys
53 * - b) "WANCache:s" : used for temporarily storing values of tombstoned keys
54 * - c) "WANCache:t" : used for storing timestamp "check" keys
59 class WANObjectCache
{
60 /** @var BagOStuff The local cluster cache */
62 /** @var string Cache pool name */
64 /** @var EventRelayer */
68 protected $lastRelayError = self
::ERR_NONE
;
70 /** Seconds to tombstone keys on delete() */
71 const HOLDOFF_TTL
= 10;
72 /** Seconds to keep dependency purge keys around */
73 const CHECK_KEY_TTL
= 31536000; // 1 year
74 /** Seconds to keep lock keys around */
77 /** Cache format version number */
80 /** Fields of value holder arrays */
81 const FLD_VERSION
= 0;
86 /** Possible values for getLastError() */
87 const ERR_NONE
= 0; // no error
88 const ERR_NO_RESPONSE
= 1; // no response
89 const ERR_UNREACHABLE
= 2; // can't connect
90 const ERR_UNEXPECTED
= 3; // response gave some error
91 const ERR_RELAY
= 4; // relay broadcast failed
93 const VALUE_KEY_PREFIX
= 'WANCache:v:';
94 const STASH_KEY_PREFIX
= 'WANCache:s:';
95 const TIME_KEY_PREFIX
= 'WANCache:t:';
97 const PURGE_VAL_PREFIX
= 'PURGED:';
100 * @param array $params
101 * - cache : BagOStuff object
103 * - relayer : EventRelayer object
105 public function __construct( array $params ) {
106 $this->cache
= $params['cache'];
107 $this->pool
= $params['pool'];
108 $this->relayer
= $params['relayer'];
112 * Fetch the value of a key from cache
114 * If passed in, $curTTL is set to the remaining TTL (current time left):
115 * - a) INF; if the key exists and has no TTL
116 * - b) float (>=0); if the key exists and has a TTL
117 * - c) float (<0); if the key is tombstoned or expired by $checkKeys
118 * - d) null; if the key does not exist and is not tombstoned
120 * If a key is tombstoned, $curTTL will reflect the time since delete().
122 * The timestamp of $key will be checked against the last-purge timestamp
123 * of each of $checkKeys. Those $checkKeys not in cache will have the last-purge
124 * initialized to the current timestamp. If any of $checkKeys have a timestamp
125 * greater than that of $key, then $curTTL will reflect how long ago $key
126 * became invalid. Callers can use $curTTL to know when the value is stale.
127 * The $checkKeys parameter allow mass invalidations by updating a single key:
128 * - a) Each "check" key represents "last purged" of some source data
129 * - b) Callers pass in relevant "check" keys as $checkKeys in get()
130 * - c) When the source data that "check" keys represent changes,
131 * the touchCheckKey() method is called on them
133 * For keys that are hot/expensive, consider using getWithSetCallback() instead.
135 * @param string $key Cache key
136 * @param mixed $curTTL Approximate TTL left on the key if present [returned]
137 * @param array $checkKeys List of "check" keys
138 * @return mixed Value of cache key or false on failure
140 final public function get( $key, &$curTTL = null, array $checkKeys = array() ) {
142 $values = $this->getMulti( array( $key ), $curTTLs, $checkKeys );
143 $curTTL = isset( $curTTLs[$key] ) ?
$curTTLs[$key] : null;
145 return isset( $values[$key] ) ?
$values[$key] : false;
149 * Fetch the value of several keys from cache
151 * @see WANObjectCache::get()
153 * @param array $keys List of cache keys
154 * @param array $curTTLs Map of (key => approximate TTL left) for existing keys [returned]
155 * @param array $checkKeys List of "check" keys
156 * @return array Map of (key => value) for keys that exist
158 final public function getMulti(
159 array $keys, &$curTTLs = array(), array $checkKeys = array()
164 $vPrefixLen = strlen( self
::VALUE_KEY_PREFIX
);
165 $valueKeys = self
::prefixCacheKeys( $keys, self
::VALUE_KEY_PREFIX
);
166 $checkKeys = self
::prefixCacheKeys( $checkKeys, self
::TIME_KEY_PREFIX
);
168 // Fetch all of the raw values
169 $wrappedValues = $this->cache
->getMulti( array_merge( $valueKeys, $checkKeys ) );
170 $now = microtime( true );
172 // Get/initialize the timestamp of all the "check" keys
173 $checkKeyTimes = array();
174 foreach ( $checkKeys as $checkKey ) {
175 $timestamp = isset( $wrappedValues[$checkKey] )
176 ? self
::parsePurgeValue( $wrappedValues[$checkKey] )
178 if ( !is_float( $timestamp ) ) {
179 // Key is not set or invalid; regenerate
180 $this->cache
->add( $checkKey,
181 self
::PURGE_VAL_PREFIX
. $now, self
::CHECK_KEY_TTL
);
185 $checkKeyTimes[] = $timestamp;
188 // Get the main cache value for each key and validate them
189 foreach ( $valueKeys as $vKey ) {
190 if ( !isset( $wrappedValues[$vKey] ) ) {
191 continue; // not found
194 $key = substr( $vKey, $vPrefixLen ); // unprefix
196 list( $value, $curTTL ) = $this->unwrap( $wrappedValues[$vKey], $now );
197 if ( $value !== false ) {
198 $result[$key] = $value;
199 foreach ( $checkKeyTimes as $checkKeyTime ) {
200 // Force dependant keys to be invalid for a while after purging
201 // to reduce race conditions involving stale data getting cached
202 $safeTimestamp = $checkKeyTime + self
::HOLDOFF_TTL
;
203 if ( $safeTimestamp >= $wrappedValues[$vKey][self
::FLD_TIME
] ) {
204 $curTTL = min( $curTTL, $checkKeyTime - $now );
209 $curTTLs[$key] = $curTTL;
216 * Set the value of a key from cache
218 * Simply calling this method when source data changes is not valid because
219 * the changes do not replicate to the other WAN sites. In that case, delete()
220 * should be used instead. This method is intended for use on cache misses.
222 * @param string $key Cache key
223 * @param mixed $value
224 * @param integer $ttl Seconds to live [0=forever]
225 * @return bool Success
227 final public function set( $key, $value, $ttl = 0 ) {
228 $key = self
::VALUE_KEY_PREFIX
. $key;
229 $wrapped = $this->wrap( $value, $ttl );
231 $func = function ( $cache, $key, $cWrapped ) use ( $wrapped ) {
232 return ( is_string( $cWrapped ) )
233 ?
false // key is tombstoned; do nothing
237 return $this->cache
->merge( $key, $func, $ttl, 1 );
241 * Purge a key from all clusters
243 * This instantiates a hold-off period where the key cannot be
244 * written to avoid race conditions where dependent keys get updated
245 * with a stale value (e.g. from a DB slave).
247 * This should only be called when the underlying data (being cached)
248 * changes in a significant way. If called twice on the same key, then
249 * the last TTL takes precedence.
251 * @param string $key Cache key
252 * @param integer $ttl How long to block writes to the key [seconds]
253 * @return bool True if the item was purged or not found, false on failure
255 final public function delete( $key, $ttl = self
::HOLDOFF_TTL
) {
256 $key = self
::VALUE_KEY_PREFIX
. $key;
257 // Update the local cluster immediately
258 $ok = $this->cache
->set( $key, self
::PURGE_VAL_PREFIX
. microtime( true ), $ttl );
259 // Publish the purge to all clusters
260 return $this->relayPurge( $key, $ttl ) && $ok;
264 * Fetch the value of a timestamp "check" key
266 * Note that "check" keys won't collide with other regular keys
269 * @return float|bool TS_UNIX timestamp of the key; false if not present
271 final public function getCheckKeyTime( $key ) {
272 return self
::parsePurgeValue( $this->cache
->get( self
::TIME_KEY_PREFIX
. $key ) );
276 * Purge a "check" key from all clusters, invalidating keys that use it
278 * This should only be called when the underlying data (being cached)
279 * changes in a significant way, and it is impractical to call delete()
280 * on all keys that should be changed. When get() is called on those
281 * keys, the relevant "check" keys must be supplied for this to work.
283 * The "check" key essentially represents a last-modified field.
284 * It is set in the future a few seconds when this is called, to
285 * avoid race conditions where dependent keys get updated with a
286 * stale value (e.g. from a DB slave).
288 * Note that "check" keys won't collide with other regular keys
290 * @see WANObjectCache::get()
292 * @param string $key Cache key
293 * @return bool True if the item was purged or not found, false on failure
295 final public function touchCheckKey( $key ) {
296 $key = self
::TIME_KEY_PREFIX
. $key;
297 // Update the local cluster immediately
298 $ok = $this->cache
->set( $key,
299 self
::PURGE_VAL_PREFIX
. microtime( true ), self
::CHECK_KEY_TTL
);
300 // Publish the purge to all clusters
301 return $this->relayPurge( $key, self
::CHECK_KEY_TTL
) && $ok;
305 * Method to fetch/regenerate cache keys
307 * On cache miss, the key will be set to the callback result.
308 * The callback function returns the new value given the current
309 * value (false if not present). If false is returned, then nothing
310 * will be saved to cache.
312 * Usually, callbacks ignore the current value, but it can be used
313 * to maintain "most recent X" values that come from time or sequence
314 * based source data, provided that the "as of" id/time is tracked.
316 * Usage of $checkKeys is similar to get()/getMulti(). However,
317 * rather than the caller having to inspect a "current time left"
318 * variable (e.g. $curTTL, $curTTLs), a cache regeneration will be
319 * triggered using the callback.
321 * The simplest way to avoid stampedes for hot keys is to use
322 * the 'lockTSE' option in $opts. If cache purges are needed, also:
323 * a) Pass $key into $checkKeys
324 * b) Use touchCheckKey( $key ) instead of delete( $key )
325 * Following this pattern lets the old cache be used until a
326 * single thread updates it as needed. Also consider tweaking
327 * the 'lowTTL' parameter.
331 * $key = wfMemcKey( 'cat-recent-actions', $catId );
332 * // Function that derives the new key value given the old value
333 * $callback = function( $cValue ) { ... };
334 * // Get the key value from cache or from source on cache miss;
335 * // try to only let one cluster thread manage doing cache updates
336 * $opts = array( 'lockTSE' => 5, 'lowTTL' => 10 );
337 * $value = $cache->getWithSetCallback( $key, $callback, 60, array(), $opts );
342 * $key = wfMemcKey( 'cat-state', $catId );
343 * // The "check" keys that represent things the value depends on;
344 * // Calling touchCheckKey() on them invalidates "cat-state"
345 * $checkKeys = array(
346 * wfMemcKey( 'water-bowls', $houseId ),
347 * wfMemcKey( 'food-bowls', $houseId ),
348 * wfMemcKey( 'people-present', $houseId )
350 * // Function that derives the new key value
351 * $callback = function() { ... };
352 * // Get the key value from cache or from source on cache miss;
353 * // try to only let one cluster thread manage doing cache updates
354 * $opts = array( 'lockTSE' => 5, 'lowTTL' => 10 );
355 * $value = $cache->getWithSetCallback( $key, $callback, 60, $checkKeys, $opts );
358 * @see WANObjectCache::get()
360 * @param string $key Cache key
361 * @param callable $callback Value generation function
362 * @param integer $ttl Seconds to live when the key is updated [0=forever]
363 * @param array $checkKeys List of "check" keys
364 * @param array $opts Options map:
365 * - lowTTL : consider pre-emptive updates when the current TTL (sec)
366 * of the key is less than this. It becomes more likely
367 * over time, becoming a certainty once the key is expired.
368 * - lockTSE : if the key is tombstoned or expired (by $checkKeys) less
369 * than this many seconds ago, then try to have a single
370 * thread handle cache regeneration at any given time.
371 * Other threads will try to use stale values if possible.
372 * If, on miss, the time since expiration is low, the assumption
373 * is that the key is hot and that a stampede is worth avoiding.
374 * - tempTTL : when 'lockTSE' is set, this determines the TTL of the temp
375 * key used to cache values while a key is tombstoned.
376 * This avoids excessive regeneration of hot keys on delete() but
377 * may result in stale values.
378 * @return mixed Value to use for the key
380 final public function getWithSetCallback(
381 $key, $callback, $ttl, array $checkKeys = array(), array $opts = array()
383 $lowTTL = isset( $opts['lowTTL'] ) ?
$opts['lowTTL'] : min( 10, $ttl );
384 $lockTSE = isset( $opts['lockTSE'] ) ?
$opts['lockTSE'] : -1;
385 $tempTTL = isset( $opts['tempTTL'] ) ?
$opts['tempTTL'] : 5;
387 // Get the current key value
389 $cValue = $this->get( $key, $curTTL, $checkKeys ); // current value
390 $value = $cValue; // return value
392 // Determine if a regeneration is desired
393 if ( $value !== false && $curTTL > 0 && !$this->worthRefresh( $curTTL, $lowTTL ) ) {
397 if ( !is_callable( $callback ) ) {
398 throw new InvalidArgumentException( "Invalid cache miss callback provided." );
401 // Assume a key is hot if requested soon after invalidation
402 $isHot = ( $curTTL !== null && $curTTL <= 0 && abs( $curTTL ) <= $lockTSE );
403 $isTombstone = ( $curTTL !== null && $value === false );
406 if ( $isHot ||
$isTombstone ) {
407 // Acquire a cluster-local non-blocking lock
408 if ( $this->cache
->lock( $key, 0, self
::LOCK_TTL
) ) {
409 // Lock acquired; this thread should update the key
411 } elseif ( $value !== false ) {
412 // If it cannot be acquired; then the stale value can be used
415 // Either another thread has the lock or the lock failed.
416 // Use the stash value, which is likely from the prior thread.
417 $value = $this->cache
->get( self
::STASH_KEY_PREFIX
. $key );
418 // Regenerate on timeout or if the other thread failed
419 if ( $value !== false ) {
425 // Generate the new value from the callback...
426 $value = call_user_func( $callback, $cValue );
427 // When delete() is called, writes are write-holed by the tombstone,
428 // so use a special stash key to pass the new value around threads.
429 if ( $value !== false && ( $isHot ||
$isTombstone ) ) {
430 $this->cache
->set( self
::STASH_KEY_PREFIX
. $key, $value, $tempTTL );
434 $this->cache
->unlock( $key );
437 if ( $value !== false ) {
438 // Update the cache; this will fail if the key is tombstoned
439 $this->set( $key, $value, $ttl );
446 * Get the "last error" registered; clearLastError() should be called manually
447 * @return int ERR_* constant for the "last error" registry
449 final public function getLastError() {
450 if ( $this->lastRelayError
) {
451 // If the cache and the relayer failed, focus on the later.
452 // An update not making it to the relayer means it won't show up
453 // in other DCs (nor will consistent re-hashing see up-to-date values).
454 // On the other hand, if just the cache update failed, then it should
455 // eventually be applied by the relayer.
456 return $this->lastRelayError
;
459 $code = $this->cache
->getLastError();
461 case BagOStuff
::ERR_NONE
:
462 return self
::ERR_NONE
;
463 case BagOStuff
::ERR_NO_RESPONSE
:
464 return self
::ERR_NO_RESPONSE
;
465 case BagOStuff
::ERR_UNREACHABLE
:
466 return self
::ERR_UNREACHABLE
;
468 return self
::ERR_UNEXPECTED
;
473 * Clear the "last error" registry
475 final public function clearLastError() {
476 $this->cache
->clearLastError();
477 $this->lastRelayError
= self
::ERR_NONE
;
481 * Do the actual async bus purge of a key
483 * This must set the key to "PURGED:<UNIX timestamp>"
485 * @param string $key Cache key
486 * @param integer $ttl How long to keep the tombstone [seconds]
487 * @return bool Success
489 protected function relayPurge( $key, $ttl ) {
490 $event = $this->cache
->modifySimpleRelayEvent( array(
493 'val' => 'PURGED:$UNIXTIME$',
494 'ttl' => max( $ttl, 1 ),
495 'sbt' => true, // substitute $UNIXTIME$ with actual microtime
498 $ok = $this->relayer
->notify( "{$this->pool}:purge", $event );
500 $this->lastRelayError
= self
::ERR_RELAY
;
507 * Check if a key should be regenerated (using random probability)
509 * This returns false if $curTTL >= $lowTTL. Otherwise, the chance
510 * of returning true increases steadily from 0% to 100% as the $curTTL
511 * moves from $lowTTL to 0 seconds. This handles widely varying
512 * levels of cache access traffic.
514 * @param float|INF $curTTL Approximate TTL left on the key if present
515 * @param float $lowTTL Consider a refresh when $curTTL is less than this
518 protected function worthRefresh( $curTTL, $lowTTL ) {
519 if ( $curTTL >= $lowTTL ) {
521 } elseif ( $curTTL <= 0 ) {
525 $chance = ( 1 - $curTTL / $lowTTL );
527 return mt_rand( 1, 1e9
) <= 1e9
* $chance;
531 * Do not use this method outside WANObjectCache
533 * @param mixed $value
534 * @param integer $ttl [0=forever]
537 protected function wrap( $value, $ttl ) {
539 self
::FLD_VERSION
=> self
::VERSION
,
540 self
::FLD_VALUE
=> $value,
541 self
::FLD_TTL
=> $ttl,
542 self
::FLD_TIME
=> microtime( true )
547 * Do not use this method outside WANObjectCache
549 * @param array|string|bool $wrapped
550 * @param float $now Unix Current timestamp (preferrable pre-query)
551 * @return array (mixed; false if absent/invalid, current time left)
553 protected function unwrap( $wrapped, $now ) {
554 // Check if the value is a tombstone
555 $purgeTimestamp = self
::parsePurgeValue( $wrapped );
556 if ( is_float( $purgeTimestamp ) ) {
557 // Purged values should always have a negative current $ttl
558 $curTTL = min( -0.000001, $purgeTimestamp - $now );
559 return array( false, $curTTL );
562 if ( !is_array( $wrapped ) // not found
563 ||
!isset( $wrapped[self
::FLD_VERSION
] ) // wrong format
564 ||
$wrapped[self
::FLD_VERSION
] !== self
::VERSION
// wrong version
566 return array( false, null );
569 if ( $wrapped[self
::FLD_TTL
] > 0 ) {
570 // Get the approximate time left on the key
571 $age = $now - $wrapped[self
::FLD_TIME
];
572 $curTTL = max( $wrapped[self
::FLD_TTL
] - $age, 0.0 );
574 // Key had no TTL, so the time left is unbounded
578 return array( $wrapped[self
::FLD_VALUE
], $curTTL );
583 * @param string $prefix
586 protected static function prefixCacheKeys( array $keys, $prefix ) {
588 foreach ( $keys as $key ) {
589 $res[] = $prefix . $key;
596 * @param string $value String like "PURGED:<timestamp>"
597 * @return float|bool UNIX timestamp or false on failure
599 protected static function parsePurgeValue( $value ) {
601 if ( is_string( $value ) &&
602 preg_match( '/^' . self
::PURGE_VAL_PREFIX
. '([^:]+)$/', $value, $m )