Cost-Based Optimizer

The cost-based optimizer seeks the lowest cost for a query, usually related to time.

How is cost calculated?

A given SQL query can have thousands of equivalent query plans with vastly different execution times. The cost-based optimizer enumerates these plans and chooses the lowest cost plan.

Cost is roughly calculated by:

• Estimating how much time each node in the query plan will use to process all results
• Modeling how data flows through the query plan

The most important factor in determining the quality of a plan is cardinality (i.e., the number of rows); the fewer rows each SQL operator needs to process, the faster the query will run.

Table statistics

The cost-based optimizer can often find more performant query plans if it has access to statistical data on the contents of your tables. This data needs to be generated from scratch for new tables, and regenerated periodically for existing tables.

By default, CockroachDB automatically generates table statistics when tables are created, and as they are updated. It does this using a background job that automatically determines which columns to get statistics on — specifically, it chooses:

• Columns that are part of the primary key or an index (in other words, all indexed columns).
• Up to 100 non-indexed columns.

By default, CockroachDB also automatically collects multi-column statistics on columns that prefix an index.

Controlling automatic statistics

For best query performance, most users should leave automatic statistics enabled with the default settings. The information provided in this section is useful for troubleshooting or performance tuning by advanced users.

Controlling statistics refresh rate

Statistics are refreshed in the following cases:

1. When there are no statistics.
2. When it's been a long time since the last refresh, where "long time" is defined according to a moving average of the time across the last several refreshes.
3. After each mutation operation (INSERT, UPDATE, or DELETE), the probability of a refresh is calculated using a formula that takes the cluster settings shown below as inputs. These settings define the target number of rows in a table that should be stale before statistics on that table are refreshed. Increasing either setting will reduce the frequency of refreshes. In particular, min_stale_rows impacts the frequency of refreshes for small tables, while fraction_stale_rows has more of an impact on larger tables.

Turning off statistics

If you need to turn off automatic statistics collection, follow the steps below:

1. Run the following statement to disable the automatic statistics cluster setting:

   > SET CLUSTER SETTING sql.stats.automatic_collection.enabled = false;
   

2. Use the SHOW STATISTICS statement to view automatically generated statistics.

3. Delete the automatically generated statistics using the following statement:

   > DELETE FROM system.table_statistics WHERE true;
   

4. Restart the nodes in your cluster to clear the statistics caches.

For instructions showing how to manually generate statistics, see the examples in the CREATE STATISTICS documentation.

Controlling histogram collection

By default, the optimizer collects histograms for all index columns (specifically the first column in each index) during automatic statistics collection. If a single column statistic is explicitly requested using manual invocation of CREATE STATISTICS, a histogram will be collected, regardless of whether or not the column is part of an index.

If you are an advanced user and need to disable histogram collection for troubleshooting or performance tuning reasons, change the sql.stats.histogram_collection.enabled cluster setting by running SET CLUSTER SETTING as follows:

SET CLUSTER SETTING sql.stats.histogram_collection.enabled = false;

Query plan cache

CockroachDB uses a cache for the query plans generated by the optimizer. This can lead to faster query execution since the database can reuse a query plan that was previously calculated, rather than computing a new plan each time a query is executed.

The query plan cache is enabled by default. To disable it, execute the following statement:

> SET CLUSTER SETTING sql.query_cache.enabled = false;

Finally, note that only the following statements use the plan cache:

• SELECT
• INSERT
• UPDATE
• UPSERT
• DELETE

Join reordering

The cost-based optimizer will explore additional join orderings in an attempt to find the lowest-cost execution plan for a query involving multiple joins, which can lead to significantly better performance in some cases.

Because this process leads to an exponential increase in the number of possible execution plans for such queries, it's only used to reorder subtrees containing 4 or fewer joins by default.

To change this setting, which is controlled by the reorder_joins_limit session variable, run the statement shown below. To disable this feature, set the variable to 0.

> SET reorder_joins_limit = 6;

For more information about the difficulty of selecting an optimal join ordering, see our blog post An Introduction to Join Ordering.

Join hints

The optimizer supports hint syntax to force the use of a specific join algorithm even if the optimizer determines that a different plan would have a lower cost. The algorithm is specified between the join type (INNER, LEFT, etc.) and the JOIN keyword, for example:

• INNER HASH JOIN
• OUTER MERGE JOIN
• LEFT LOOKUP JOIN
• CROSS MERGE JOIN
• INNER INVERTED JOIN
• LEFT INVERTED JOIN

Note that the hint cannot be specified with a bare hint keyword (e.g., MERGE) - in that case, the INNER keyword must be added. For example, a INNER MERGE JOIN b will work, but a MERGE JOIN b will not work.

Supported join algorithms

• HASH: Forces a hash join; in other words, it disables merge and lookup joins. A hash join is always possible, even if there are no equality columns - CockroachDB considers the nested loop join with no index a degenerate case of the hash join (i.e., a hash table with one bucket).

• MERGE: Forces a merge join, even if it requires re-sorting both sides of the join.

• LOOKUP: Forces a lookup join into the right side; the right side must be a table with a suitable index. Note that LOOKUP can only be used with INNER and LEFT joins.

• INVERTED: Forces an inverted join into the right side; the right side must be a table with a suitable inverted index. Note that INVERTED can only be used with INNER and LEFT joins.

If it is not possible to use the algorithm specified in the hint, an error is signaled.

Additional considerations

• This syntax is consistent with the SQL Server syntax for join hints, except that:

  • SQL Server uses LOOP instead of LOOKUP.
  • CockroachDB does not support LOOP and instead supports LOOKUP for the specific case of nested loop joins with an index.

• When a join hint is specified, the two tables will not be reordered by the optimizer. The reordering behavior has the following characteristics, which can be affected by hints:

  • Given a JOIN b, CockroachDB will not try to commute to b JOIN a. This means that you will need to pay attention to this ordering, which is especially important for lookup joins. Without a hint, a JOIN b might be executed as b INNER LOOKUP JOIN a using an index into a, whereas a INNER LOOKUP JOIN b requires an index into b.
  • (a JOIN b) JOIN c might be changed to a JOIN (b JOIN c), but this does not happen if a JOIN b uses a hint; the hint forces that particular join to happen as written in the query.

• Hint usage should be reconsidered with each new release of CockroachDB. Due to improvements in the optimizer, hints specified to work with an older version may cause decreased performance in a newer version.

Examples

Inverted join examples

To run these examples, initialize a demo cluster with the MovR workload.

Start the MovR database on a 3-node CockroachDB demo cluster with a larger data set.

cockroach demo movr --num-histories 250000 --num-promo-codes 250000 --num-rides 125000 --num-users 12500 --num-vehicles 3750 --nodes 3

Create an inverted index on the vehicles table's ext column.

CREATE INVERTED INDEX idx_vehicle_details ON vehicles(ext);

Check the statement plan for a SELECT statement that uses an inner inverted join.

EXPLAIN SELECT * FROM vehicles@primary AS v2 INNER INVERTED JOIN vehicles@idx_vehicle_details AS v1 ON v1.ext @> v2.ext;

                                           info
-------------------------------------------------------------------------------------------
  distribution: full
  vectorized: true

  • lookup join
  │ table: vehicles@primary
  │ equality: (city, id) = (city,id)
  │ equality cols are key
  │ pred: ext @> ext
  │
  └── • inverted join
      │ table: vehicles@idx_vehicle_details
      │
      └── • scan
            estimated row count: 3,750 (100% of the table; stats collected 3 minutes ago)
            table: vehicles@primary
            spans: FULL SCAN
(16 rows)

Time: 1ms total (execution 1ms / network 0ms)

You can omit the INNER INVERTED JOIN statement by putting v1.ext on the left side of a @> join condition in a WHERE clause and using an index hint for the inverted index.

EXPLAIN SELECT * FROM vehicles@idx_vehicle_details AS v1, vehicles AS v2 WHERE v1.ext @> v2.ext;

                                            info
--------------------------------------------------------------------------------------------
  distribution: full
  vectorized: true

  • lookup join
  │ table: vehicles@primary
  │ equality: (city, id) = (city,id)
  │ equality cols are key
  │ pred: ext @> ext
  │
  └── • inverted join
      │ table: vehicles@idx_vehicle_details
      │
      └── • scan
            estimated row count: 3,750 (100% of the table; stats collected 12 minutes ago)
            table: vehicles@primary
            spans: FULL SCAN
(16 rows)

Time: 1ms total (execution 1ms / network 0ms)

Use the LEFT INVERTED JOIN hint to perform a left inverted join.

EXPLAIN SELECT * FROM vehicles AS v2 LEFT INVERTED JOIN vehicles AS v1 ON v1.ext @> v2.ext;

                                            info
--------------------------------------------------------------------------------------------
  distribution: full
  vectorized: true

  • lookup join (left outer)
  │ table: vehicles@primary
  │ equality: (city, id) = (city,id)
  │ equality cols are key
  │ pred: ext @> ext
  │
  └── • inverted join (left outer)
      │ table: vehicles@idx_vehicle_details
      │
      └── • scan
            estimated row count: 3,750 (100% of the table; stats collected 16 minutes ago)
            table: vehicles@primary
            spans: FULL SCAN
(16 rows)

Time: 2ms total (execution 2ms / network 0ms)

See also

• JOIN expressions
• SET (session variable)
• SET CLUSTER SETTING
• RESET CLUSTER SETTING
• SHOW (session variable)
• CREATE STATISTICS
• SHOW STATISTICS
• EXPLAIN

Was this page helpful?

Yes No