Query caching

    If you’re unfamiliar with Druid architecture, review the following topics before proceeding with caching:

    For instructions to configure query caching see Using query caching.

    Cache monitoring, including the hit rate and number of evictions, is available in .

    Query-level caching is in addition to data-level caching on Historicals.

    Druid supports two types of query caching:

    • Per-segment caching stores partial query results for a specific segment. It is enabled by default.
    • stores final query results.

    Druid invalidates any cache the moment any underlying data change to avoid returning stale results. This is especially important for datasources that have highly-variable underlying data segments, including real-time data segments.

    The primary form of caching in Druid is a per-segment results cache. This cache stores partial query results on a per-segment basis and is enabled on Historical services by default.

    Druid may potentially merge per-segment cached results with the results of later queries that use a similar basic shape with similar filters, aggregations, etc. For example, if the query is identical except that it covers a different time period.

    Per-segment caching is controlled by the parameters useCache and populateCache.

    Use per-segment caching with real-time data. For example, your queries request data actively arriving from Kafka alongside intervals in segments that are loaded on Historicals. Druid can merge cached results from Historical segments with real-time results from the stream. , on the other hand, is not helpful in this scenario because new data from real-time ingestion will continually invalidate the entire cached result.

    With whole-query caching, Druid caches the entire results of individual queries, meaning the Broker no longer needs to merge per-segment results from data processes.

    Use whole-query caching on the Broker to increase query efficiency when there is little risk of ingestion invalidating the cache at a segment level. This applies particularly, for example, when not using real-time ingestion. Perhaps your queries tend to use batch-ingested data, in which case per-segment caching would be less efficient since the underlying segments hardly ever change, yet Druid would continue to acquire per-segment results for each query.

    Per-segment cache is available as follows:

    • On Historicals, the default. Enable segment-level cache population on Historicals for larger production clusters to prevent Brokers from having to merge all query results. When you enable cache population on Historicals instead of Brokers, the Historicals merge their own local results and put less strain on the Brokers.

    • On ingestion tasks in the Peon or Indexer service. Larger production clusters should enable segment-level cache population on task services only to prevent Brokers from having to merge all query results. When you enable cache population on task execution services instead of Brokers, the task execution services to merge their own local results and put less strain on the Brokers.

      Task executor services only support caches that store data locally. For example the caffeine cache. This restriction exists because the cache stores results at the level of intermediate partial segments generated by the ingestion tasks. These intermediate partial segments may not be identical across task replicas. Therefore task executor services ignore remote cache types such as .

    Avoid using per-segment cache at the Broker for large production clusters. When the Broker cache is enabled (druid.broker.cache.populateCache is true) and populateCache is not in the query context, individual Historicals will not merge individual segment-level results, and instead pass these back to the lead Broker. The Broker must then carry out a large merge from all segments on its own.

    Whole-query cache is available exclusively on Brokers.

    Caching enables increased concurrency on the same system, therefore leading to noticeable performance improvements for queries on Druid clusters handling throughput for concurrent, mixed workloads.

    If you are looking to improve response time for a single query or page load, you should ignore caching. In general, response time for a single task should meet performance objectives even when the cache is cold.

    During query processing, the per-segment cache intercepts the query and sends the results directly to the Broker. This way the query bypasses the data server processing threads. For queries requiring minimal processing in the Broker, cached queries are very quick. If work done on the Broker causes a query bottleneck, enabling caching results in little noticeable query improvement.

    The largest performance gains from segment caching tend to apply to topN and time series queries. For groupBy queries, if the bottleneck is in the merging phase on the broker, the impact is less. The same applies to queries with or without joins.

    Caching does not solve all types of query performance issues. For each cache type there are scenarios where caching is likely to be of little benefit.

    Per-segment caching doesn’t work for the following:

    • queries containing a sub-query in them. However the output of sub-queries may be cached. See for more details on sub-queries execution.
    • queries with joins do not support any caching on the broker.
    • queries with bySegment set in the query context are not cached on the broker.

    Whole-query caching doesn’t work for the following:

    • queries that involve an inline datasource or a lookup datasource.
    • GroupBy v2 queries.
    • queries with joins.
    • queries with a union datasource.
    • to learn how to configure and use caching.
    • Druid Design to learn about Druid processes.
    • to learn how Druid stores data.