Package org.apache.lucene.search

Class LRUQueryCache

java.lang.Object
org.apache.lucene.search.LRUQueryCache
All Implemented Interfaces:
QueryCache, Accountable
public class LRUQueryCache extends Object implements QueryCache, Accountable
A QueryCache that evicts queries using a LRU (least-recently-used) eviction policy in order to remain under a given maximum size and number of bytes used.

This class is thread-safe.

Note that query eviction runs in linear time with the total number of segments that have cache entries so this cache works best with caching policies that only cache on "large" segments, and it is advised to not share this cache across too many indices.

A default query cache and policy instance is used in IndexSearcher. If you want to replace those defaults it is typically done like this: 

   final int maxNumberOfCachedQueries = 256;
   final long maxRamBytesUsed = 50 * 1024L * 1024L; // 50MB
   // these cache and policy instances can be shared across several queries and readers
   // it is fine to eg. store them into static variables
   final QueryCache queryCache = new LRUQueryCache(maxNumberOfCachedQueries, maxRamBytesUsed);
   final QueryCachingPolicy defaultCachingPolicy = new UsageTrackingQueryCachingPolicy();
   indexSearcher.setQueryCache(queryCache);
   indexSearcher.setQueryCachingPolicy(defaultCachingPolicy);
This cache exposes some global statistics (hit count, miss count, number of cache entries, total number of DocIdSets that have ever been cached, number of evicted entries). In case you would like to have more fine-grained statistics, such as per-index or per-query-class statistics, it is possible to override various callbacks: onHit(java.lang.Object, org.apache.lucene.search.Query), onMiss(java.lang.Object, org.apache.lucene.search.Query), onQueryCache(org.apache.lucene.search.Query, long), onQueryEviction(org.apache.lucene.search.Query, long), onDocIdSetCache(java.lang.Object, long), onDocIdSetEviction(java.lang.Object, int, long) and onClear(). It is better to not perform heavy computations in these methods though since they are called synchronously and under a lock.
See Also:
WARNING: This API is experimental and might change in incompatible ways in the next release.

  • Constructor Details

    • LRUQueryCache

      public LRUQueryCache(int maxSize, long maxRamBytesUsed, Predicate<LeafReaderContext> leavesToCache, float skipCacheFactor)
      Expert: Create a new instance that will cache at most maxSize queries with at most maxRamBytesUsed bytes of memory, only on leaves that satisfy leavesToCache.

      Also, clauses whose cost is skipCacheFactor times more than the cost of the top-level query will not be cached in order to not slow down queries too much.

    • LRUQueryCache

      public LRUQueryCache(int maxSize, long maxRamBytesUsed)
      Create a new instance that will cache at most maxSize queries with at most maxRamBytesUsed bytes of memory. Queries will only be cached on leaves that have more than 10k documents and have more than half of the average documents per leave of the index. This should guarantee that all leaves from the upper tier will be cached. Only clauses whose cost is at most 100x the cost of the top-level query will be cached in order to not hurt latency too much because of caching.

  • Method Details

    • getSkipCacheFactor

      public float getSkipCacheFactor()
      Get the skip cache factor
      Returns:
      #setSkipCacheFactor

    • setSkipCacheFactor

      public void setSkipCacheFactor(float skipCacheFactor)
      This setter enables the skipCacheFactor to be updated dynamically.
      Parameters:
      skipCacheFactor - clauses whose cost is skipCacheFactor times more than the cost of the top-level query will not be cached in order to not slow down queries too much.

    • onHit

      protected void onHit(Object readerCoreKey, Query query)
      Expert: callback when there is a cache hit on a given query. Implementing this method is typically useful in order to compute more fine-grained statistics about the query cache.
      See Also:
      WARNING: This API is experimental and might change in incompatible ways in the next release.

    • onMiss

      protected void onMiss(Object readerCoreKey, Query query)
      Expert: callback when there is a cache miss on a given query.
      See Also:
      WARNING: This API is experimental and might change in incompatible ways in the next release.

    • onQueryCache

      protected void onQueryCache(Query query, long ramBytesUsed)
      Expert: callback when a query is added to this cache. Implementing this method is typically useful in order to compute more fine-grained statistics about the query cache.
      See Also:
      WARNING: This API is experimental and might change in incompatible ways in the next release.

    • onQueryEviction

      protected void onQueryEviction(Query query, long ramBytesUsed)
      Expert: callback when a query is evicted from this cache.
      See Also:
      WARNING: This API is experimental and might change in incompatible ways in the next release.

    • onDocIdSetCache

      protected void onDocIdSetCache(Object readerCoreKey, long ramBytesUsed)
      Expert: callback when a DocIdSet is added to this cache. Implementing this method is typically useful in order to compute more fine-grained statistics about the query cache.
      See Also:
      WARNING: This API is experimental and might change in incompatible ways in the next release.

    • onDocIdSetEviction

      protected void onDocIdSetEviction(Object readerCoreKey, int numEntries, long sumRamBytesUsed)
      Expert: callback when one or more DocIdSets are removed from this cache.
      See Also:
      WARNING: This API is experimental and might change in incompatible ways in the next release.

    • onClear

      protected void onClear()
      Expert: callback when the cache is completely cleared.
      WARNING: This API is experimental and might change in incompatible ways in the next release.

    • get

      protected LRUQueryCache.CacheAndCount get(Query key, IndexReader.CacheHelper cacheHelper)

    • clearCoreCacheKey

      public void clearCoreCacheKey(Object coreKey)
      Remove all cache entries for the given core cache key.

    • clearQuery

      public void clearQuery(Query query)
      Remove all cache entries for the given query.

    • clear

      public void clear()
      Clear the content of this cache.

    • doCache

      public Weight doCache(Weight weight, QueryCachingPolicy policy)
      Description copied from interface: QueryCache
      Return a wrapper around the provided weight that will cache matching docs per-segment accordingly to the given policy. NOTE: The returned weight will only be equivalent if scores are not needed.
      Specified by:
      doCache in interface QueryCache
      See Also:

    • ramBytesUsed

      public long ramBytesUsed()
      Description copied from interface: Accountable
      Returns an estimate of the JVM heap memory used by this object in bytes. The method name uses "ram" for historical reasons; only JVM heap memory should be reported. Off-heap resources such as memory-mapped files or native direct buffers should not be included. Negative values are illegal.
      Specified by:
      ramBytesUsed in interface Accountable

    • getChildResources

      public Collection<Accountable> getChildResources()
      Description copied from interface: Accountable
      Returns nested resources of this class. The result should be a point-in-time snapshot (to avoid race conditions).
      Specified by:
      getChildResources in interface Accountable
      See Also:

    • cacheImpl

      protected LRUQueryCache.CacheAndCount cacheImpl(BulkScorer scorer, int maxDoc) throws IOException
      Default cache implementation: uses RoaringDocIdSet for sets that have a density < 1% and a BitDocIdSet over a FixedBitSet otherwise.
      Throws:
      IOException

    • tryPopulateCache

      protected LRUQueryCache.CacheAndCount tryPopulateCache(IndexReader.CacheHelper cacheKey, Weight weight, ScorerSupplier scorerSupplier, LeafReaderContext context) throws IOException
      Populates the cache for the given scorer and leaf reader context.

      Subclasses can override this method to implement custom caching strategies, such as deferring cache population to a background thread or skipping it when other threads are already populating the cache. This can be useful for intra-segment searches.

      Returns:
      the cached result if available; otherwise null
      Throws:
      IOException

    • getTotalCount

      public final long getTotalCount()
      Return the total number of times that a Query has been looked up in this QueryCache. Note that this number is incremented once per segment so running a cached query only once will increment this counter by the number of segments that are wrapped by the searcher. Note that by definition, getTotalCount() is the sum of getHitCount() and getMissCount().
      See Also:

    • getHitCount

      public final long getHitCount()
      Over the total number of times that a query has been looked up, return how many times a cached DocIdSet has been found and returned.
      See Also:

    • getMissCount

      public final long getMissCount()
      Over the total number of times that a query has been looked up, return how many times this query was not contained in the cache.
      See Also:

    • getCacheSize

      public final long getCacheSize()
      Return the total number of DocIdSets which are currently stored in the cache.
      See Also:

    • getCacheCount

      public final long getCacheCount()
      Return the total number of cache entries that have been generated and put in the cache. It is highly desirable to have a hit count that is much higher than the cache count as the opposite would indicate that the query cache makes efforts in order to cache queries but then they do not get reused.
      See Also:

    • getEvictionCount

      public final long getEvictionCount()
      Return the number of cache entries that have been removed from the cache either in order to stay under the maximum configured size/ram usage, or because a segment has been closed. High numbers of evictions might mean that queries are not reused or that the caching policy caches too aggressively on NRT segments which get merged early.
      See Also: