Query plan optimization

It's very useful to to analyze execution plans for queries in order to detect and eliminate the causes of possible inefficiencies. YDB provides two types of query plans: logical plan and execution plan. Logical plan is better suited for analyzing complex queries with a large number of JOIN operators. Execution plan is more detailed: it additionally shows the stages of the distributed plan and connectors between them, which makes it more convenient for analyzing simple OLTP queries.

Logical Query Plan

You can get the logical plan via YDB CLI.

This plan allows you to compare the query optimizer's predictions with the execution statistics. If the predictions differ significantly from the actual data at the execution stage, this may indicate that the optimizer has not built the most efficient plan for the current query. In this case, you can use optimizer hints to create a more efficient plan.

Query Execution Plan

To illustrate how to work with the execution plan, consider the following OLTP query that searches for a series by name:

SELECT season_id, episode_id
  FROM episodes
  WHERE title = 'The Work Outing'

Schema of the episodes table:

Let's build a query execution plan for this query. You can do this via either UI or YDB CLI:

ydb -p <profile_name> table query explain \
  -q "SELECT season_id, episode_id
  FROM episodes
  WHERE title = 'The Work Outing'"

Query Plan:
ResultSet
└──Limit (Limit: 1001)
   └──<UnionAll>
      └──Limit (Limit: 1001)
      └──Filter (Predicate: item.title == "The Work Outing")
      └──TableFullScan (ReadRanges: ["series_id (-∞, +∞)","season_id (-∞, +∞)","episode_id (-∞, +∞)"], ReadColumns: ["episode_id","season_id","title"], Table: episodes)
         Tables: ["episodes"]

Both plan representations contain the result being returned to the client at the root, table operations at the leaves, and data transformations at the intermediate nodes. It is important to pay attention to the node containing the table reading operation. In this case, it is a TableFullScan for the episodes table. Full table scans consume time and resources proportional to the size of the table, so it is advisable to avoid them whenever possible in tables that tend to grow over time or are simply large.

One typical approach to avoid full scans is using a secondary index. In this case, it makes sense to add a secondary index for the column title using the following query:

ALTER TABLE episodes
  ADD INDEX title_index GLOBAL ON (title)

Please note that this example uses synchronous secondary index. Building an index in YDB is an asynchronous operation. Even if the index creation query is successful, it is advisable to wait for some time because the index may not be ready for use immediately. You can manage asynchronous operations through the CLI.

Let's build the query plan using the secondary index title_index. Secondary indexes to be used need to be explicitly specified in the VIEW clause.

ydb -p <profile_name> table query explain \
  -q "SELECT season_id, episode_id
  FROM episodes VIEW title_index
  WHERE title = 'The Work Outing'"

Query Plan:
ResultSet
└──Limit (Limit: 1001)
   └──<UnionAll>
      └──Limit (Limit: 1001)
      └──Filter (Predicate: Exist(item.title))
      └──TablePointLookup (ReadRange: ["title (The Work Outing)","series_id (-∞, +∞)","season_id (-∞, +∞)","episode_id (-∞, +∞)"], ReadLimit: 1001, ReadColumns: ["episode_id","season_id","title"], Table: episodes/title_index/indexImplTable)
         Tables: ["episodes/title_index/indexImplTable"]

The secondary index allowed the query to be executed without fully scanning the main table. Instead of a TableFullScan, we received a TablePointLookup—reading the index table by key. We no longer need to read the main table because all necessary columns are contained in the index table.