Presto 329 Documentation

12.154. Release 0.103

12.154. Release 0.103

Cluster Resource Management

There is a new cluster resource manager, which can be enabled via the experimental.cluster-memory-manager-enabled flag. Currently, the only resource that’s tracked is memory, and the cluster resource manager guarantees that the cluster will not deadlock waiting for memory. However, in a low memory situation it is possible that only one query will make progress. Memory limits can now be configured via query.max-memory which controls the total distributed memory a query may use and query.max-memory-per-node which limits the amount of memory a query may use on any one node. On each worker, the resources.reserved-system-memory flags controls how much memory is reserved for internal Presto data structures and temporary allocations.

Task Parallelism

Queries involving a large number of aggregations or a large hash table for a join can be slow due to single threaded execution in the intermediate stages. This release adds experimental configuration and session properties to execute this single threaded work in parallel. Depending on the exact query this may reduce wall time, but will likely increase CPU usage.

Use the configuration parameter task.default-concurrency or the session property task_default_concurrency to set the default number of parallel workers to use for join probes, hash builds and final aggregations. Additionally, the session properties task_join_concurrency, task_hash_build_concurrency and task_aggregation_concurrency can be used to control the parallelism for each type of work.

This is an experimental feature and will likely change in a future release. It is also expected that this will eventually be handled automatically by the query planner and these options will be removed entirely.

Hive Changes

  • Removed the hive.max-split-iterator-threads parameter and renamed hive.max-global-split-iterator-threads to hive.max-split-iterator-threads.
  • Fix excessive object creation when querying tables with a large number of partitions.
  • Do not retry requests when an S3 path is not found.

General Changes

  • Add array_remove().
  • Fix NPE in max_by() and min_by() caused when few rows were present in the aggregation.
  • Reduce memory usage of map_agg().
  • Change HTTP client defaults: 2 second idle timeout, 10 second request timeout and 250 connections per host.
  • Add SQL command autocompletion to CLI.
  • Increase CLI history file size.