GridGain Developers Hub

Entry Compression

Entry Compression

GridGain Ultimate Edition supports dictionary-based cache entry compression using Zstandard library. You can enable it on per-cache basis by setting CacheConfiguration.setEntryCompressionConfiguration(new ZstdDictionaryCompressionConfiguration).

When any cache entry is created or updated, its key and value may be transparently compressed. When the entry is accessed, it is transparently decompressed. No change in the behavior of user code is expected when enabling or disabling entry compression.

<bean class="org.apache.ignite.configuration.IgniteConfiguration">
    <property name="cacheConfiguration">
        <bean class="org.apache.ignite.configuration.CacheConfiguration">
            <property name="name" value="compressedCache"/>
            <property name="entryCompressionConfiguration">
                <bean class="org.gridgain.grid.cache.compress.ZstdDictionaryCompressionConfiguration">
                    <!-- Default values for all properties listed below. -->
                    <property name="compressKeys" value="false"/>
                    <property name="requireDictionary" value="true"/>
                    <property name="dictionarySize" value="1024"/>
                    <property name="samplesBufferSize" value="#{4 * 1024 * 1024}"/>
                    <property name="compressionLevel" value="2"/>
CacheConfiguration cfg = new CacheConfiguration();

ZstdDictionaryCompressionConfiguration compressionCfg = new ZstdDictionaryCompressionConfiguration();
// Default values for all properties listed below:
compressionCfg.setSamplesBufferSize(4 * 1024 * 1024);

This API is not presently available for C#/.NET. You can use XML configuration.
This API is not presently available for C++. You can use XML configuration.

Performance Considerations

Entry compression allows saving RAM and disk space by storing compressed data in cache entries, at the cost of spending CPU time on compression and decompression.

Enabling entry compression will lead to reduced Off-Heap utilization and checkpoint directory size, as well as slightly shorter WAL. If native persistence is used, it is possible to save space while improving performance at the same time if the load pattern is I/O bound.

Configuring Dictionary

External dictionary use is based on the assumption that cache entries are small and have similar structure and content. Thus a pre-trained dictionary allows a decent compression ratio of 40% to 60% on real-world samples that otherwise see little to no benefit from compression.

With default settings, GridGain collects values as samples when they are added to the cache. When enough samples are accumulated, a dictionary will be trained based on these samples. After the dictionary is ready, it will be applied to every new value. The value will be stored in the compressed form if it yields sufficient benefit over plain serialized form. Some values will be collected as samples after a dictionary is trained, and a new dictionary will be trained periodically. If this new dictionary shows greater benefit than the existing one, it will be used onwards; otherwise, it is discarded. Dictionaries are kept alongside cache data.

A separate set of samples and dictionaries will be kept for each individual cache, unless it is a part of cache group, in which case they are shared between all caches in the group.

The default dictionary size is 1024 bytes. As a heuristic, dictionary size should be on par with the average value size in the cache. It can be specified by setting the dictionarySize property. Longer dictionaries lead to increased performance cost but potentially better compression ratio. The recommended range of dictionary size is 264 to 16384 bytes.

As per the underlying algorithm, it is possible to specify the compression level by setting the compressionLevel property. The default value is 2, which combines decent compression performance and ratio. Tuning compression level is usually less important than choosing the best dictionary size.

You may also change the length of buffer used to accumulate samples for dictionary training by modifying the samplesBufferSize property. The samples are held on a heap, with a total size of 4 megabytes by default. You may decrease it to have smaller heap usage and train the dictionary faster or increase it to have a slightly better compression ratio (especially in the case of larger dictionarySize values).

Compression Without Dictionary

The requireDictionary property is true by default, meaning that a dictionary is required to compress data. Suppose you set this property to false. In that case, when the dictionary is not ready, data that is added to the cache will be compressed without a dictionary and stored in the compressed form if it yields sufficient benefit after compression. As soon as a dictionary is ready, it will always be used.

If the values stored in the cache are large and self-contained, you may also set dictionarySize to 0. In this mode of operation, no samples are collected; the dictionary is never trained or used.


Using cache entry compression in a cluster that has configured snapshots is not supported yet.

Currently, enabling enry compression for a cache disables historical rebalancing of that cache. Full partition rebalance is used insted.

If the compressKeys property is set, the entry key and entry value are considered for compression separately. The key may end up compressed, but the value will not, or vice versa. The same dictionary is used for the compression of both keys and values.

Currently, only binary objects will be compressed. This covers most of the usage patterns: POJO classes, SQL tables, and BinaryObject. Primitive types will not be compressed. Therefore, if String or byte[] is desired to be used as a compressed key/value type, it needs to be stored as an object field.