techdocsmith commented on a change in pull request #10848: URL: https://github.com/apache/druid/pull/10848#discussion_r589841039
########## File path: docs/querying/caching.md ########## @@ -22,63 +23,87 @@ title: "Query caching" ~ under the License. --> +You can enable caching in Apache Druid to improve query times for frequently accessed data. This topic defines the different types of caching for Druid. It describes the default caching behavior and provides guidance and examples to help you hone your caching strategy. -Apache Druid supports query result caching at both the segment and whole-query result level. Cache data can be stored in the -local JVM heap or in an external distributed key/value store. In all cases, the Druid cache is a query result cache. -The only difference is whether the result is a _partial result_ for a particular segment, or the result for an entire -query. In both cases, the cache is invalidated as soon as any underlying data changes; it will never return a stale -result. +If you're unfamiliar with Druid architecture, review the following topics before proceeding with caching: +- [Druid Design](../design/architecture.md) +- [Segments](../design/segments.md) +- [Query execution](./query-execution) -Segment-level caching allows the cache to be leveraged even when some of the underling segments are mutable and -undergoing real-time ingestion. In this case, Druid will potentially cache query results for immutable historical -segments, while re-computing results for the real-time segments on each query. Whole-query result level caching is not -useful in this scenario, since it would be continuously invalidated. +For instructions to configure query caching see [Using query caching](./using-caching.md). -Segment-level caching does require Druid to merge the per-segment results on each query, even when they are served -from the cache. For this reason, whole-query result level caching can be more efficient if invalidation due to real-time -ingestion is not an issue. +## Cache types +Druid supports the following types of caches: -## Using and populating cache +- **Per-segment** caching which stores _partial results_ of a query for a specific segment. Per-segment caching is enabled on Historicals by default. +- **Whole-query** caching which stores all results for a query. -All caches have a pair of parameters that control the behavior of how individual queries interact with the cache, a 'use' cache parameter, and a 'populate' cache parameter. These settings must be enabled at the service level via [runtime properties](../configuration/index.md) to utilize cache, but can be controlled on a per query basis by setting them on the [query context](../querying/query-context.md). The 'use' parameter obviously controls if a query will utilize cached results. The 'populate' parameter controls if a query will update cached results. These are separate parameters to allow queries on uncommon data to utilize cached results without polluting the cache with results that are unlikely to be re-used by other queries, for example large reports or very old data. +To avoid returning stale results, Druid invalidates the cache the moment any underlying data changes for both types of cache. -## Query caching on Brokers +Druid can store cache data the local JVM heap or in an external distributed key/value store. The default is a local cache based upon [Caffeine](https://github.com/ben-manes/caffeine). Maximum cache storage defaults to the minimum value of 1 GiB or the ten percent of the maximum runtime memory for the JVM with no cache expiration. See [Cache configuration](../configuration/index.md#cache-configuration) for information on how to configure cache storage. -Brokers support both segment-level and whole-query result level caching. Segment-level caching is controlled by the -parameters `useCache` and `populateCache`. Whole-query result level caching is controlled by the parameters -`useResultLevelCache` and `populateResultLevelCache` and [runtime properties](../configuration/index.md) -`druid.broker.cache.*`. +### Per-segment caching -Enabling segment-level caching on the Broker can yield faster results than if query caches were enabled on Historicals for small -clusters. This is the recommended setup for smaller production clusters (< 5 servers). Populating segment-level caches on -the Broker is _not_ recommended for large production clusters, since when the property `druid.broker.cache.populateCache` is -set to `true` (and query context parameter `populateCache` is _not_ set to `false`), results from Historicals are returned -on a per segment basis, and Historicals will not be able to do any local result merging. This impairs the ability of the -Druid cluster to scale well. +The primary form of caching in Druid is the **per-segment cache** which stores query results on a per-segment basis. It is enabled on Historical services by default. -## Query caching on Historicals +When your queries include data from segments that are mutable and undergoing real-time ingestion, use a segment cache. In this case Druid caches query results for immutable historical segments when possible. It re-computes results for the real-time segments at query time. -Historicals only support segment-level caching. Segment-level caching is controlled by the query context -parameters `useCache` and `populateCache` and [runtime properties](../configuration/index.md) -`druid.historical.cache.*`. +For example, you have queries that frequently include incoming data from a Kafka or Kinesis stream alongside unchanging segments. Per-segment caching lets Druid cache results from older immutable segments and merge them with updated data. Whole-query caching would not be helpful in this scenario because the new data from real-time ingestion continually invalidates the cache. -Larger production clusters should enable segment-level cache population on Historicals only (not on Brokers) to avoid -having to use Brokers to merge all query results. Enabling cache population on the Historicals instead of the Brokers -enables the Historicals to do their own local result merging and puts less strain on the Brokers. +### Whole-query caching -## Query caching on Ingestion Tasks +If real-time ingestion invalidating the cache is not an issue for your queries, you can use **whole-query caching** on the Broker to increase query efficiency. The Broker performs whole-query caching operations before sending fan out queries to Historicals. Therefore Druid no longer needs to merge the per-segment results on the Broker. -Task executor processes such as the Peon or the experimental Indexer only support segment-level caching. Segment-level -caching is controlled by the query context parameters `useCache` and `populateCache` -and [runtime properties](../configuration/index.md) `druid.realtime.cache.*`. +For instance, whole-query caching is a good option when you have queries that include data from a batch ingestion task that runs every few hours or once a day. Per-segment caching would be less efficient in this case because it requires Druid to merge the per-segment results for each query, even when the results are cached. + +## Where to enable caching + +**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 execution 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 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 `memcached`. + +- On Brokers for small production clusters with less than five servers. + + Do not use per-segment caches on the Broker for large production clusters. When `druid.broker.cache.populateCache` is `true` and query context parameter `populateCache` _is not_ `false`, Historicals return results on a per-segment basis without merging results locally thus negatively impacting cluster scalability. + +**Whole-query cache** is only available on Brokers. + +## Performance considerations for caching +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. The impact is less for `groupBy` queries. The same applies to queries with or without joins. Review comment: ```suggestion The largest performance gains from segment caching tend to apply to `topN` and time series queries. The impact is less for `groupBy` queries. ``` I removed that sentence. I lifted it from a perf topic by @gianm and I can't find the source context now. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
