Cost-based оптимизация#

Примечание

Ниже приведена оригинальная документация Trino. Скоро мы ее переведем на русский язык и дополним полезными примерами.

CedrusData поддерживает несколько видов cost-based оптимизации запросов, описанных ниже.

Режим оптимизации Cascades#

Выполнение ряда операций, таких как Join, Aggregation или Window, зачастую требует передачи данных между узлами. В процессе планирования запроса оптимизатор Trino вставляет рядом с такими операторами специальный оператор Exchange, который моделирует передачу данных между узлами. В дальнейшем Trino разбивает план запроса на несколько фрагментов по границам Exchange.

По умолчанию оптимизатор Trino определяет местоположение операторов Exchange эвристически на основе требований конкретного оператора, что во многих случаях приводит к неоптимальным планам.

В режиме оптимизации Cascades оптимизатор рассматривает всю совокупность операторов запроса для выбора наиболее оптимальной расстановки операторов Exchange. Это позволяет оптимизатору во многих случаях снизить количество фрагментов, необходимых для выполнения запроса, а так же уменьшить количество передаваемых по сети данных. Совместно это приводит к ускорению запросов с большим количеством операций Join, Aggregation и Window.

Для включения данного режима воспользуйтесь параметром конфигурации cedrusdata.optimizer.cascades-enabled или своейства сессии cedrusdata_cascades_enabled.

Join enumeration#

The order in which joins are executed in a query can have a significant impact on the query’s performance. The aspect of join ordering that has the largest impact on performance is the size of the data being processed and transferred over the network. If a join which produces a lot of data is performed early in the query’s execution, then subsequent stages need to process large amounts of data for longer than necessary, increasing the time and resources needed for processing the query.

With cost-based join enumeration, Trino uses Table statistics provided by connectors to estimate the costs for different join orders and automatically picks the join order with the lowest computed costs.

The join enumeration strategy is governed by the join_reordering_strategy session property, with the optimizer.join-reordering-strategy configuration property providing the default value.

The possible values are:

  • AUTOMATIC (default) - enable full automatic join enumeration

  • ELIMINATE_CROSS_JOINS - eliminate unnecessary cross joins

  • NONE - purely syntactic join order

If you are using AUTOMATIC join enumeration and statistics are not available or a cost can not be computed for any other reason, the ELIMINATE_CROSS_JOINS strategy is used instead.

Join distribution selection#

Trino uses a hash-based join algorithm. For each join operator, a hash table must be created from one join input, referred to as the build side. The other input, called the probe side, is then iterated on. For each row, the hash table is queried to find matching rows.

There are two types of join distributions:

  • Partitioned: each node participating in the query builds a hash table from only a fraction of the data

  • Broadcast: each node participating in the query builds a hash table from all of the data. The data is replicated to each node.

Each type has advantages and disadvantages. Partitioned joins require redistributing both tables using a hash of the join key. These joins can be much slower than broadcast joins, but they allow much larger joins overall. Broadcast joins are faster if the build side is much smaller than the probe side. However, broadcast joins require that the tables on the build side of the join after filtering fit in memory on each node, whereas distributed joins only need to fit in distributed memory across all nodes.

With cost-based join distribution selection, Trino automatically chooses whether to use a partitioned or broadcast join. With cost-based join enumeration, Trino automatically chooses which sides are probe and build.

The join distribution strategy is governed by the join_distribution_type session property, with the join-distribution-type configuration property providing the default value.

The valid values are:

  • AUTOMATIC (default) - join distribution type is determined automatically for each join

  • BROADCAST - broadcast join distribution is used for all joins

  • PARTITIONED - partitioned join distribution is used for all join

Capping replicated table size#

The join distribution type is automatically chosen when the join reordering strategy is set to AUTOMATIC or when the join distribution type is set to AUTOMATIC. In both cases, it is possible to cap the maximum size of the replicated table with the join-max-broadcast-table-size configuration property or with the join_max_broadcast_table_size session property. This allows you to improve cluster concurrency and prevent bad plans when the cost-based optimizer misestimates the size of the joined tables.

By default, the replicated table size is capped to 100MB.

Syntactic join order#

If not using cost-based optimization, Trino defaults to syntactic join ordering. While there is no formal way to optimize queries for this case, it is possible to take advantage of how Trino implements joins to make them more performant.

Trino uses in-memory hash joins. When processing a join statement, Trino loads the right-most table of the join into memory as the build side, then streams the next right-most table as the probe side to execute the join. If a query has multiple joins, the result of this first join stays in memory as the build side, and the third right-most table is then used as the probe side, and so on for additional joins. In the case where join order is made more complex, such as when using parentheses to specify specific parents for joins, Trino may execute multiple lower-level joins at once, but each step of that process follows the same logic, and the same applies when the results are ultimately joined together.

Because of this behavior, it is optimal to syntactically order joins in your SQL queries from the largest tables to the smallest, as this minimizes memory usage.

As an example, if you have a small, medium, and large table and are using left joins:

SELECT
  *
FROM
  large_table l
  LEFT JOIN medium_table m ON l.user_id = m.user_id
  LEFT JOIN small_table s ON s.user_id = l.user_id

Предупреждение

This means of optimization is not a feature of Trino. It is an artifact of how joins are implemented, and therefore this behavior may change without notice.

Connector implementations#

In order for the Trino optimizer to use the cost based strategies, the connector implementation must provide Table statistics.