GridGain Developers Hub

SQL Performance Tuning

This article outlines basic and advanced optimization techniques for GridGain SQL queries. Some of the sections are also useful for debugging and troubleshooting.

Basic Considerations: GridGain vs RDBMS

GridGain and Ignite are frequently compared to relational databases for their SQL capabilities with an expectation that existing SQL queries, created for an RDBMS, will work out of the box and perform faster in GridGain without any changes. Usually, such an assumption is based on the fact that GridGain stores and processes data in-memory. However, it’s not enough just to put data in RAM and expect an order of magnitude increase in performance. Generally, extra tuning is required. Below you can see a standard checklist of best practices to consider before you benchmark GridGain against an RDBMS or do any performance testing:

  • Ignite/GridGain is optimized for multi-nodes deployments with RAM as a primary storage. Don’t try to compare a single-node GridGain cluster to a relational database. You should deploy a multi-node GridGain cluster with the whole copy of data in RAM.

  • Be ready to adjust your data model and existing SQL queries. Use the affinity colocation concept during the data modelling phase for proper data distribution. Remember, it’s not enough just to put data in RAM. If your data is properly colocated, you can run SQL queries with JOINs at massive scale and expect significant performance benefits.

  • Define secondary indexes and use other standard, and GridGain-specific, tuning techniques described below.

  • Keep in mind that relational databases leverage local caching techniques and, depending on the total data size, an RDBMS can complete some queries even faster than GridGain even in a multi-node configuration. If your data set is around 10-100GB and an RDBMS has enough RAM for caching data locally than it, for instance, can outperform a multi-node GridGain cluster because the latter will be utilizing the network. Store much more data in GridGain to see the difference.

Using the EXPLAIN Statement

GridGain supports the EXPLAIN statement which could be used to read the execution plan of a query. Use this command to analyse your queries for possible optimization. Note that the plan will contain multiple rows: the last one will contain a query for the reducing side (usually your application), others are for map nodes (usually server nodes). Read the Distributed Queries section to learn how queries are executed in GridGain.

EXPLAIN SELECT name FROM Person WHERE age = 26;

The execution plan is generated by H2 as described here.

OR Operator and Selectivity

If a query contains an OR operator, then indexes may not be used as expected depending on the complexity of the query. For example, for the query select name from Person where gender='M' and (age = 20 or age = 30), an index on the gender field will be used instead of an index on the age field, although the latter is a more selective index. As a workaround for this issue, you can rewrite the query with UNION ALL (notice that UNION without ALL will return DISTINCT rows, which will change the query semantics and will further penalize your query performance):

SELECT name FROM Person WHERE gender='M' and age = 20
SELECT name FROM Person WHERE gender='M' and age = 30

Avoid Having Too Many Columns

Avoid having too many columns in the result set of a SELECT query. Due to limitations of the H2 query parser, queries with 100+ columns may perform worse than expected.

Enforcing Join Order

When this flag is set, the query optimizer will not reorder tables in joins. In other words, the order in which joins are applied during query execution will be the same as specified in the query. Without this flag, the query optimizer can reorder joins to improve performance. However, sometimes it might make an incorrect decision. This flag helps to control and explicitly specify the order of joins instead of relying on the optimizer.

Consider the following example:

SELECT * FROM Person p
JOIN Company c ON = where = 'John Doe'
AND p.age > 20
AND > 5000
AND < 100000

This query contains a join between two tables: Person and Company. To get the best performance, we should understand which join will return the smallest result set. The table with the smaller result set size should be given first in the join pair. To get the size of each result set, let’s test each part.

SELECT count(*)
FROM Person p
where = 'John Doe'
AND p.age > 20
AND > 5000
AND < 100000;
SELECT count(*)
FROM Company c
where NOT LIKE 'O%';

After running Q1 and Q2, we can get two different outcomes:

Case 1:





Q2 returns more entries than Q1. In this case, we don’t need to modify the original query, because smaller subset has already been located on the left side of the join.

Case 2:





Q1 returns more entries than Q2. So we need to change the initial query as follows:

FROM Company c
JOIN Person p
ON =
where = 'John Doe'
AND p.age > 20
AND > 5000
AND < 100000

The force join order hint can be specified as follows:

Increasing Index Inline Size

Every entry in the index has a constant size which is calculated during index creation. This size is called index inline size. Ideally this size should be enough to store full indexed entry in serialized form. When values are not fully included in the index, GridGain may need to perform additional data page reads during index lookup, which can impair performance if persistence is enabled.

Here is how values are stored in the index:

0     1       5
| tag | value |
Total: 5 bytes

0     1       9
| tag | value |
Total: 9 bytes

0     1      3             N
| tag | size | UTF-8 value |
Total: 3 + string length

POJO (BinaryObject)
0     1         5
| tag | BO hash |
Total: 5

For primitive data types (bool, byte, short, int, etc.) and variable length data with specified precision (for example, varchar(20)), GridGain automatically calculates the index inline size so that the values are included in full. For example, for int fields, the inline size is 5 (1 byte for the tag and 4 bytes for the value itself). For long fields, the inline size is 9 (1 byte for the tag + 8 bytes for the value).

For binary objects, the index includes the hash of each object, which is enough to avoid collisions. The inline size is 5.

For variable length data without specified precision, indexes include only first several bytes of the value. Therefore, when indexing fields with variable-length data, we recommend that you estimate the length of your field values and set the inline size to a value that includes most (about 95%) or all values. For example, if you have a String field with 95% of the values containing 10 characters or fewer, you can set the inline size for the index on that field to 13.

The inline sizes explained above apply to single field indexes. However, when you define an index on a field in the value object or on a non-primary key column, GridGain creates a composite index by appending the primary key to the indexed value. Therefore, when calculating the inline size for composite indexes, add up the inline size of the primary key.

Below is an example of index inline size calculation for a cache where both key and value are complex objects.

public class Key {
    private long id;

    private long affinityKey;

public class Value {
    @QuerySqlField(index = true)
    private long longField;

    @QuerySqlField(index = true)
    private int intField;

    @QuerySqlField(index = true)
    private String stringField; // we suppose that 95% of the values are 10 symbols

The following table summarizes the inline index sizes for the indexes defined in the example above.

Index Kind Recommended Inline Size Comment


Primary key index


Inlined hash of a binary object (5)

(affinityKey, _key)

Affinity key index


Inlined long (9) + binary object’s hash (5)

(longField, _key)

Secondary index


Inlined long (9) + binary object’s hash (5)

(intField, _key)

Secondary index


Inlined int (5) + binary object up to hash (5)

(stringField, _key)

Secondary index


Inlined string (13) + binary object’s hash (5) (assuming that the string is ~10 symbols)

Note that you will only have to set the inline size for the index on stringField. For other indexes, GridGain will calculate the inline size automatically.

Refer to the Configuring Index Inline Size section for the information on how to change the inline size.

You can check the inline size of an existing index in the INDEXES system view.

Query Parallelism

By default, a SQL query is executed in a single thread on each participating GridGain node. This approach is optimal for queries returning small result sets involving index search. For example:


Certain queries might benefit from being executed in multiple threads. This relates to queries with table scans and aggregations, which is often the case for HTAP and OLAP workloads. For example:

SELECT SUM(salary) FROM Person;

The number of threads created on a single node for query execution is configured per cache and by default equals 1. You can change the value by setting the CacheConfiguration.queryParallelism parameter. If you create SQL tables using the CREATE TABLE command, you can use a cache template to set this parameter.

If a query contains JOINs, then all the participating caches must have the same degree of parallelism.

Index Hints

Index hints are useful in scenarios when you know that one index is more suitable for certain queries than another. You can use them to instruct the query optimizer to choose a more efficient execution plan. To do this, you can use USE INDEX(indexA,…​,indexN) statement as shown in the following example.

SELECT * FROM Person USE INDEX(index_age)
WHERE salary > 150000 AND age < 35;

Partition Pruning

Partition pruning is a technique that optimizes queries that use affinity keys in the WHERE condition. When executing such a query, GridGain will scan only those partitions where the requested data is stored. This will reduce query time because the query will be sent only to the nodes that store the requested partitions.

In the following example, the employee objects are colocated by the id field (if an affinity key is not set explicitly then the primary key is used as the affinity key):

CREATE TABLE employee (id BIGINT PRIMARY KEY, department_id INT, name VARCHAR)

/* This query is sent to the node where the requested key is stored */
SELECT * FROM employee WHERE id=10;

/* This query is sent to all nodes */
SELECT * FROM employee WHERE department_id=10;

In the next example, the affinity key is set explicitly and, therefore, will be used to colocate data and direct queries to the nodes that keep primary copies of the data:

CREATE TABLE employee (id BIGINT, department_id INT, name VARCHAR,  PRIMARY KEY(id, department_id)) WITH "AFFINITY_KEY=department_id"

/* This query is sent to all nodes */
SELECT * FROM employee WHERE id=10;

/* This query is sent to the node where the requested key is stored */
SELECT * FROM employee WHERE department_id=10;

Skip Reducer on Update

When GridGain executes a DML operation, it first fetches all the affected intermediate rows for analysis to the reducer node (usually your application), and only then prepares batches of updated values that will be sent to remote nodes.

This approach might affect performance and saturate the network if a DML operation has to move many entries.

Use this flag as a hint for the SQL engine to do all intermediate rows analysis and updates “in-place” on the server nodes. To add a hint in JDBC and ODBC, specify it in the connection string. In SQL query API, specify setSkipReducerOnUpdate:

//jdbc connection string
cache.query(new SqlFieldsQuery("UPDATE Person SET name = Fail(name)")

SQL On-heap Row Cache

GridGain stores data and indexes in its own memory space outside of Java heap. This means that with every data access, a part of the data will be copied from the off-heap space to Java heap, potentially deserialized, and kept in the heap as long as your application or server node references it.

The SQL on-heap row cache is intended to store hot rows (key-value objects) in Java heap, minimizing resources spent for data copying and deserialization. Each cached row refers to an entry in the off-heap region and can be invalidated when one of the following happens:

  • The master entry stored in the off-heap region is updated or removed.

  • The data page that stores the master entry is evicted from RAM.

The on-heap row cache can be enabled for a specific cache/table (if you use CREATE TABLE to create SQL tables and caches, then the parameter can be passed via a cache template):

<bean class="org.apache.ignite.configuration.IgniteConfiguration">
    <property name="cacheConfiguration">
        <bean class="org.apache.ignite.configuration.CacheConfiguration">
            <property name="name" value="myCache"/>
            <property name="sqlOnheapCacheEnabled" value="true"/>

If the row cache is enabled, you might be able to trade RAM for performance. You might get up to a 2x performance increase for some SQL queries and use cases by allocating more RAM for rows caching purposes.

Using TIMESTAMP instead of DATE

Use the TIMESTAMP type instead of DATE whenever possible. Presently, the DATE type is serialized/deserialized very inefficiently resulting in performance degradation.

Hash Joins

To boost performance of join queries, GridGain supports the hash join algorithm. Hash joins can be more efficient than nested loop joins for many scenarios, except when the probe side of the join is very small. However, hash joins can only be used with equi-joins, i.e. a type of join with equality comparison in the join-predicate.

To enforce the use of hash joins:

  1. Use the enforceJoinOrder option:

    SqlFieldsQuery query = new SqlFieldsQuery(
    // Open the JDBC connection.
    Connection conn = DriverManager.getConnection("jdbc:ignite:thin://");
    var query = new SqlFieldsQuery("SELECT * FROM TABLE_A, TABLE_B USE INDEX(HASH_JOIN_IDX) WHERE TABLE_A.column1 = TABLE_B.column2")
        EnforceJoinOrder = true
    SqlFieldsQuery query = SqlFieldsQuery("SELECT * FROM TABLE_A, TABLE_B USE INDEX(HASH_JOIN_IDX) WHERE TABLE_A.column1 = TABLE_B.column2");
  2. Specify USE INDEX(HASH_JOIN_IDX) on the table for which you want to create the hash-join index: