A newer version of Hazelcast Platform is available.

View latest

Indexing Maps

When members run queries, they iterate through all their owned entries and find the ones that match the query. Indexes allow members to find matching entries faster by reducing the number of entries to check.

If you often query a map, make sure to add indexes to its most frequently queried fields. For example, if you do active AND age < 30 query, make sure you add an index for the active and age fields. The following example code does that by getting the map from the Hazelcast instance and adding indexes to the map with the map.addIndex() method.

IMap map = hazelcastInstance.getMap( "employees" );
// ordered, since we have ranged queries for this field
map.addIndex(new IndexConfig(IndexType.SORTED, "age"));
// not ordered, because boolean field cannot have range
map.addIndex(new IndexConfig(IndexType.HASH, "active"));
When you add new entries to maps that use indexes, it takes longer because the index must be added along with the entry.

You only need to create an index once. Although it is safe to call the map.addIndex() method more than once, doing so has a negative affect on performance due to the redundant index creation.

For example, when you call the map.addIndex("fieldName", true) method, each partition iterates over its records and adds each entry to the index. The previously created index entry will be recreated and replaced with the new entry. The performance penalty will be proportional to the number of entries. If you have maps with a large number of entries, then synchronizing index addition process is recommended.

Other than using the map. addIndex() method, you can define your index declaratively or programmatically as described in the Configuring IMap Indexes section.

Indexing Ranged Queries

The map.addIndex(IndexConfig) method is used for adding index. For each indexed field, if you have ranged queries such as age > 30, age BETWEEN 40 AND 60, then use a IndexType.SORTED index Otherwise, use a IndexType.HASH index.

Configuring Map Indexes

You can define map indexes in configuration. An example is shown below.

  • XML

  • YAML

  • Spring

<hazelcast>
    ...
    <map name="default">
        <indexes>
            <index type="HASH">
                <attributes>
                    <attribute>name</attribute>
                </attributes>
            </index>
            <index>
                <attributes>
                    <attribute>age</attribute>
                </attributes>
            </index>
        </indexes>
    </map>
    ...
</hazelcast>
hazelcast:
  map:
    default:
      indexes:
        - type: HASH
            attributes:
              - "name"
        - attributes:
            - "age"
<hz:map name="default">
    <hz:indexes>
        <hz:index type="HASH">
            <hz:attributes>
                <hz:attribute>name</hz:attribute>
            </hz:attributes>
        </hz:index>
        <hz:index>
            <hz:attributes>
                <hz:attribute>age</hz:attribute>
            </hz:attributes>
        </hz:index>
    </hz:indexes>
</hz:map>

You can also define map indexes using programmatic configuration, as in the example below.

mapConfig.addIndexConfig(new IndexConfig(IndexType.HASH, "name"));
mapConfig.addIndexConfig(new IndexConfig(IndexType.SORTED, "age"));
Non-primitive types to be indexed should implement Comparable.

If you configure the data structure to use High-Density Memory Store and indexes, the indexes are automatically stored in the High-Density Memory Store as well. This prevents from running into full garbage collections when doing a lot of updates to index.

Global and Partitioned Indexes

The on-heap indexes are always global, where one index covers all map entries stored on the partitions owned by a cluster member. Such indexes are beneficial for lookup and range queries because only one lookup operation is needed to execute a query. A drawback of global indexes is a potentially high contention on the index concurrent data structure that might cause performance degradation.

High-Density Memory Store supports partitioned indexes. Each partition owned by a cluster member has its own index. All operations on the partitioned index are performed on the partitioned thread, thus eliminating the contention issue of the global indexes. However, lookup and range queries have to perform lookup operations on every partition and combine the results. Normally, these partition and combine executions yield poorer performance results compared to the global indexes.

Global concurrent indexes (based on our own off-heap B+ Tree implementation) bring all the benefits of global indexes to IMap backed by High-Density Memory Store.

The global High-Density Memory Store indexes are enabled by default and controlled by the hazelcast.hd.global.index.enabled property. You can disable these indexes by setting this property to false.

Composite Indexes

Composite indexes, also known as compound indexes, are special kind of indexes that are built on top of the multiple map entry attributes and therefore may be used to significantly speed up the queries involving those attributes simultaneously.

There are two distinct composite index types used for two different purposes: unordered composite indexes and ordered ones.

Unordered Composite Indexes

The unordered indexes are used to perform equality queries, also known as the point queries, e.g., name = 'Alice'. These are specifically optimized for equality queries and don’t support other comparison operators like > or <=.

Additionally, the composite unordered indexes allow speeding up the equality queries involving multiple attributes simultaneously, e.g., name = 'Alice' and age = 33. This example query results in a single composite index lookup operation which can be performed very efficiently.

The unordered composite index on the name and age attributes may be configured for a map as follows:

  • XML

  • YAML

<hazelcast>
    ...
    <map name="persons">
        <indexes>
            <index type="HASH">
                <attributes>
                    <attribute>name</attribute>
                    <attribute>age</attribute>
                </attributes>
            </index>
        </indexes>
    </map>
    ...
</hazelcast>
hazelcast:
  map:
    default:
      - type: HASH
          attributes:
            - "name"
            - "age"

The attributes indexed by the unordered composite indexes can’t be matched partially: the name = 'Alice' query can’t utilize the composite index configured above.

Ordered Composite Indexes

The ordered indexes are specifically designed to perform efficient order comparison queries, also known as the range queries, e.g., age > 33. The equality queries, like age = 33, are still supported by the ordered indexes, but they are handled in a slightly less efficient manner comparing to the unordered indexes.

The composite ordered indexes extend the concept by allowing multiple equality predicates and a single order comparison predicate to be combined into a single index query operation. For instance, the name = 'Alice' and age > 33 and name = 'Bob' and age = 33 and balance > 0.0 queries are good candidates to be covered by an ordered composite index configured as follows:

  • XML

  • YAML

<hazelcast>
    ...
    <map name="persons">
        <indexes>
            <index>
                <attributes>
                    <attribute>name</attribute>
                    <attribute>age</attribute>
                    <attribute>balance</attribute>
                </attributes>
            </index>
        </indexes>
    </map>
    ...
</hazelcast>
hazelcast:
  map:
    persons:
      indexes:
        - attributes:
          - "name"
          - "age"
          - "balance"

Unlike the unordered composite indexes, partial attribute prefixes may be matched for the ordered composite indexes. In general, a valid non-empty attribute prefix is formed as a sequence of zero or more equality predicates followed by a zero or exactly one order comparison predicate. Given the index definition above, the following queries may be served by the index: name = 'Alice', name > 'Alice', name = 'Alice' and age > 33, name = 'Alice' and age = 33 and balance = 5.0. The following queries can’t be served the index: age = 33, age > 33 and balance = 0.0, balance > 0.0.

While matching the ordered composite indexes, multiple order comparison predicates acting on the same attribute are treated as a single range predicate acting on that attribute. Given the index definition above, the following queries may be served by the index: name > 'Alice' and name < 'Bob', name = 'Alice' and age > 33 and age < 55, name = 'Alice' and age = 33 and balance > 0.0 and balance < 100.0.

Composite Index Matching and Selection

The order of attributes involved in a query plays no role in the selection of the matching composite index: name = 'Alice' and age = 33 and age = 33 and name = 'Alice' queries are equivalent from the point of view of the index matching procedure.

The attributes involved in a query can be matched partially by the composite index matcher: name = 'Alice' and age = 33 and balance > 0.0 can be partially matched by the name, age composite index, the name = 'Alice' and age = 33 predicates are served by the matched index, while the balance > 0.0 predicate is processed by other means.

Bitmap Indexes

Bitmap indexes provide capabilities similar to unordered/hash indexes. The same set of predicates is supported:

  • equal

  • notEqual

  • in,

  • and

  • or

  • not

But, unlike hash indexes, bitmap indexes are able to achieve a much higher memory efficiency for low cardinality attributes at the cost of reduced query performance. In practice, the query performance is comparable to the performance of hash indexes, while memory footprint reduction is high, usually around an order of magnitude.

Bitmap indexes are specifically designed for indexing of collection and array attributes since a single IMap entry produces many index entries in that case. A single hash index entry costs a few tens of bytes, while a single bitmap index entry usually costs just a few bytes.

It’s also possible to improve the memory footprint while indexing regular single-value attributes, but the improvement is usually minor, depending on the data layout and total number of indexes.

Currently, bitmap indexes are not supported by off-heap High-Density Memory Stores (HD).

Configuring Bitmap Indexes

In the simplest form, bitmap index for an IMap entry attribute can be declaratively configured as follows:

  • XML

  • YAML

<hazelcast>
    ...
    <map name="persons">
        <indexes>
            <index type="BITMAP">
                <attributes>
                    <attribute>age</attribute>
                </attributes>
            </index>
        </indexes>
    </map>
    ...
</hazelcast>
hazelcast:
  map:
    persons:
      indexes:
        - type: BITMAP
          attributes:
            - "age"

Internally, a unique non-negative long ID is assigned to every indexed IMap entry based on the entry key. That unique ID is required for bitmap indexes to distinguish one indexed IMap entry from another.

The mapping between IMap entries and long IDs is not free and its performance and memory footprint can be improved in certain cases. For instance, if IMap entries already have a unique integer-valued attribute, the attribute values can be used as unique long IDs directly without any additional transformations. That can be configured as follows:

  • XML

  • YAML

<index type="BITMAP">
    <attributes>
        <attribute>age</attribute>
    </attributes>
    <bitmap-index-options>
        <unique-key>uniqueId</unique-key>
        <unique-key-transformation>RAW</unique-key-transformation>
    </bitmap-index-options>
</index>
      indexes:
        - type: BITMAP
          attributes:
            - "age"
          bitmap-index-options:
            unique-key: uniqueId
            unique-key-transformation: RAW

The index definition above instructs Hazelcast to create a bitmap index on the age attribute, extract the unique key values from uniqueId attribute and use the raw (RAW) extracted values directly as long IDs. If the extracted unique key value is not of long type, the widening conversion is performed for the following types: byte, short and int; boxed variants are also supported.

In certain cases, the extracted raw IDs might be randomly distributed. This causes increased memory usage in bitmap indexes since the best case scenario for them is sequential contiguous IDs. That can be countered by applying the renumbering technique:

  • XML

  • YAML

<index type="BITMAP">
    <attributes>
        <attribute>age</attribute>
    </attributes>
    <bitmap-index-options>
        <unique-key>uniqueId</unique-key>
        <unique-key-transformation>LONG</unique-key-transformation>
    </bitmap-index-options>
</index>
      indexes:
        - type: BITMAP
          attributes:
            - "age"
          bitmap-index-options:
            unique-key: uniqueId
            unique-key-transformation: LONG

The index definition above instructs the bitmap index to extract the unique keys from uniqueId attribute, convert every extracted non-negative value to long (LONG) and assign an internal sequential unique long ID based on that extracted and then converted unique value. The widening conversion is applied to the extracted values, if necessary.

This long-to-long mapping is performed more efficiently than the general object-to-long mapping done for the simple index definitions. Basically, the following simple bitmap index definition:

  • XML

  • YAML

<index type="BITMAP">
    <attributes>
        <attribute>age</attribute>
    </attributes>
</index>
      indexes:
        - type: BITMAP
          attributes:
            - "age"

is equivalent to the following full-form definition:

  • XML

  • YAML

<index type="BITMAP">
    <attributes>
        <attribute>age</attribute>
    </attributes>
    <bitmap-index-options>
        <unique-key>__key</unique-key>
        <unique-key-transformation>OBJECT</unique-key-transformation>
    </bitmap-index-options>
</index>
      indexes:
        - type: BITMAP
          attributes:
            - "age"
          bitmap-index-options:
            unique-key: __key
            unique-key-transformation: OBJECT

Which indexes age attribute, uses IMap entry keys (__key) interpreted as Java objects (OBJECT) to assign internal unique long IDs.

The full-form definition syntax is defined as follows:

  • XML

  • YAML

<index type="BITMAP">
    <attributes>
        <attribute><attr></attribute>
    </attributes>
    <bitmap-index-options>
        <unique-key><key></unique-key>
        <unique-key-transformation><transformation></unique-key-transformation>
    </bitmap-index-options>
</index>
      indexes:
        - type: BITMAP
          attributes:
            - <attribute>
          bitmap-index-options:
            unique-key: <key>
            unique-key-transformation: <transformation>

The following are the parameter descriptions:

  • <attr>: Specifies the attribute index.

  • <key>: Specifies the attribute to use as a unique key source for internal unique long ID assignment.

  • <transformation>: Specifies the transformation to be applied to unique keys to generate unique long IDs from them. The following transformations are supported:

    • OBJECT: Object-to-long transformation. Each extracted unique key value is interpreted as a Java object instance. Internally, an object-to-long hash table is used to establish the mapping from unique keys to unique IDs. Good as a general-purpose transformation.

    • LONG: Long-to-long transformation. Each extracted unique key value is interpreted as a non-negative long value, the widening conversion from byte, short and int is performed, if necessary. Internally, a long-to-long hash table is used to establish the mapping from unique keys to unique IDs, which is more efficient than the object-to-long hash table. It is good for sparse/random unique integer-valued keys renumbering to raise the IDs density and to make the bitmap index more memory-efficient as a result.

    • RAW: Raw transformation. Each extracted unique key value is interpreted as a non-negative long value, the widening conversion from byte, short and int is performed, if necessary. Internally, no hash table of any kind is used to establish the mapping from unique keys to unique IDs, the raw extracted keys are used directly as IDs. It is good for dense unique integer-valued keys, and it has the best performance in terms of time and memory.

The regular dotted attribute path syntax is supported for <attr> and <key>:

  • XML

  • YAML

<index type="BITMAP">
    <attributes>
        <attribute>name.first</attribute>
    </attributes>
</index>
<index type="BITMAP">
    <attributes>
        <attribute>name.first</attribute>
    </attributes>
    <bitmap-index-options>
        <unique-key>__key.id</unique-key>
    </bitmap-index-options>
</index>
<index type="BITMAP">
    <attributes>
        <attribute>name.first</attribute>
    </attributes>
    <bitmap-index-options>
        <unique-key>id.external</unique-key>
    </bitmap-index-options>
</index>
      indexes:
        - type: BITMAP
          attributes:
            - name.first
        - type: BITMAP
          attributes:
            - name.first
          bitmap-index-options:
            unique-key: __key.id
        - type: BITMAP
          attributes:
            - name.first
          bitmap-index-options:
            unique-key: id.external

Collection and array indexing is also possible using the regular syntax:

  • XML

  • YAML

<index type="BITMAP">
    <attributes>
        <attribute>habits[any]</attribute>
    </attributes>
</index>
<index type="BITMAP">
    <attributes>
        <attribute>habits[0]</attribute>
    </attributes>
</index>
      indexes:
        - type: BITMAP
          attributes:
            - habits[any]
        - type: BITMAP
          attributes:
            - habits[0]

Bitmap Index Querying

Bitmap index matching and selection for queries are performed automatically. No special treatment is required. The querying can be performed using the regular IMap querying methods: IMap.values(Predicate), IMap.entrySet(Predicate), etc.

Copying Indexes

The underlying data structures used by the indexes need to copy the query results to make sure that the results are correct. This copying process is performed either when reading the index from the data structure (on-read) or writing to it (on-write).

On-read copying means that, for each index-read operation, the result of the query is copied before it is sent to the caller. Depending on the query result’s size, this type of index copying may be slower since the result is stored in a map, i.e., all entries need to have the hash calculated before being stored. Unlike the index-read operations, each index-write operation is fast, since there is no copying. So, this option can be preferred in index-write intensive cases.

On-write copying means that each index-write operation completely copies the underlying map to provide the copy-on-write semantics and this may be a slow operation depending on the index size. Unlike index-write operations, each index-read operation is fast since the operation only includes accessing the map that stores the results and returning them to the caller.

Another option is never copying the results of a query to a separate map. This means the results backed by the underlying index-map can change after the query has been executed (such as an entry might have been added or removed from an index, or it might have been remapped). This option can be preferred if you expect "mostly correct" results, i.e., if it is not a problem when some entries returned in the query result set do not match the initial query criteria. This is the fastest option since there is no copying.

You can set one of these options using the system property hazelcast.index.copy.behavior. The following values, which are explained in the above paragraphs, can be set:

  • COPY_ON_READ (the default value)

  • COPY_ON_WRITE

  • NEVER

The following is an example configuration snippet:

  • XML

  • YAML

<hazelcast>
    <cluster-name>dev</cluster-name>
    ...
    <properties>
        <property name="hazelcast.index.copy.behavior">NEVER</property>
    </properties>
    ...
</hazelcast>
hazelcast:
  cluster-name: dev
  ...
  properties:
    hazelcast.index.copy.behavior: NEVER
  ...

See also the Configuring with System Properties section for reference.

Usage of this system property is supported for BINARY and OBJECT in-memory formats. Only in Hazelcast 3.8.7, it is also supported for NATIVE in-memory format.

Indexing Attributes with ValueExtractor

You can also define custom attributes that may be referenced in predicates, queries and indexes. Custom attributes can be defined by implementing a ValueExtractor. See the Custom Attributes section for details.

Using "this" as an Attribute

You can use the keyword this as an attribute name while adding an index or creating a predicate. A basic usage is shown below.

map.addIndex(new IndexConfig(IndexType.SORTED, "this"));
Predicate<Integer, Integer> lessEqual = Predicates.between("this", 12, 20);

Another basic example using SQL predicate is shown below.

Predicates.sql("this = 'jones'")
Predicates.sql("this.age > 33")

The special attribute this acts on the value of a map entry. Typically, you do not need to specify it while accessing a property of an entry’s value, since its presence is implicitly assumed if the special attribute __key is not specified.