The bitsets belonging to a query component are independent from the rest of the search request. This means that, once cached, a query can be reused in multiple search requests. It is not dependent on the "context" of the surrounding query. This allows caching to accelerate the most frequently used portions of your queries, without wasting overhead on the less frequent / more volatile portions.

Similarly, if a single search request reuses the same non-scoring query, its cached bitset can be reused for all instances inside the single search request.

Let’s look at this example query, which looks for emails that are either of the following:

GET /inbox/emails/_search
{
  "query": {
      "constant_score": {
          "filter": {
              "bool": {
                 "should": [
                    { "bool": {
                          "must": [
                             { "term": { "folder": "inbox" }}, 
                             { "term": { "read": false }}
                          ]
                    }},
                    { "bool": {
                          "must_not": {
                             "term": { "folder": "inbox" } 
                          },
                          "must": {
                             "term": { "important": true }
                          }
                    }}
                 ]
              }
            }
        }
    }
}

Even though one of the inbox clauses is a must clause and the other is a must_not clause, the two clauses themselves are identical. If this particular term query was previously cached, both instances would benefit from the cached representation despite being used in different styles of boolean logic.

This ties in nicely with the composability of the query DSL. It is easy to move filtering queries around, or reuse the same query in multiple places within the search request. This isn’t just convenient to the developer—​it has direct performance benefits.

In older versions of Elasticsearch, the default behavior was to cache everything that was cacheable. This often meant the system cached bitsets too aggressively and performance suffered due to thrashing the cache. In addition, many filters are very fast to evaluate, but substantially slower to cache (and reuse from cache). These filters don’t make sense to cache, since you’d be better off just re-executing the filter again.

Inspecting the inverted index is very fast and most query components are rare. Consider a term filter on a "user_id" field: if you have millions of users, any particular user ID will only occur rarely. It isn’t profitable to cache the bitsets for this filter, as the cached result will likely be evicted from the cache before it is used again.

This type of cache churn can have serious effects on performance. What’s worse, it is difficult for developers to identify which components exhibit good cache behavior and which are useless.

To address this, Elasticsearch caches queries automatically based on usage frequency. If a non-scoring query has been used a few times (dependent on the query type) in the last 256 queries, the query is a candidate for caching. However, not all segments are guaranteed to cache the bitset. Only segments that hold more than 10,000 documents (or 3% of the total documents, whichever is larger) will cache the bitset. Because small segments are fast to search and merged out quickly, it doesn’t make sense to cache bitsets here.

Once cached, a non-scoring bitset will remain in the cache until it is evicted. Eviction is done on an LRU basis: the least-recently used filter will be evicted once the cache is full.