How to tune internal RocksDB state stores ?

Internally, KafkaStreams relies on an embedded key-value store so-called RocksDB to provided persistent storage. Depending on the throughput of your application, you may want to tune internal RocksDB instances.

Kafka Streams allows you to customize the RocksDB settings for a given Store by implementing the interface org.apache.kafka.streams.state.RocksDBConfigSetter.

The custom implementation must then be configured using :

streamsConfig.put(StreamsConfig.ROCKSDB_CONFIG_SETTER_CLASS_CONFIG, CustomRocksDBConfig.class)

Azkarra Streams provides a built-in io.streamthoughts.azkarra.commons.rocksdb.AzkarraRocksDBConfigSetter which is automatically configured if no one is already provided.

AzkarraRocksDBConfigSetter allows overriding not only some default RocksDB options but also to enable log statistics for performance debugging and shared memory usage.

Available properties are :

Property Type Default Description
rocksdb.stats.enable boolean Enable RocksDB statistics
rocksdb.stats.dump.period.sec integer The RocksDB statistics dump period in seconds.
rocksdb.log.dir string The RocksDB log directory
rocksdb.log.level string The RocksDB log level (see org.rocksdb.InfoLogLevel).
rocksdb.log.max.file.size integer The RocksDB maximum log file size.
rocksdb.max.write.buffer.number integer The maximum number of memtables build up in memory, before they flush to SST files.
rocksdb.write.buffer.size long The size of a single memtable.
rocksdb.memory.managed boolean false Enable automatic memory management across all RocksDB instances.
rocksdb.memory.write.buffer.ratio double 0.5 The ratio of total cache memory which will be reserved for write buffer manager. This property is only used when ‘rocksdb.memory.managed’ is set to true.
rocksdb.memory.high.prio.pool.ratio double 0.1 The ratio of cache memory that is reserved for high priority blocks (e.g.: indexes, filters and compressions blocks).
rocksdb.memory.strict.capacity.limit boolean false Create a block cache with strict capacity limit, i.e., insert to the cache will fail when cache is full. This property is only used when ‘rocksdb.memory.managed’ is set to true or ‘rocksdb.block.cache.size’ is set.
rocksdb.block.cache.size long false The total size to be used for caching uncompressed data blocks. string The compaction style.
rocksdb.compression.type string The compression type. long The maximum number of open files that can be used per RocksDB instance.
rocksdb.background.thread.flush.pool integer The number of threads to be used for the background flush process.
rocksdb.background.compaction.flush.pool integer The number of threads to be used for the background compaction process.
rocksdb.max.background.flushes integer The maximum number of concurrent flush operations.

Note that all properties described above are optional.

RocksDB properties can be passed either using default configuration :

azrkarra {
  streams {
    rocksdb {
      stats.enable = false
      stats.dumpPeriodSec = 30
      log {
        dir = "/var/log/kafka-streams/rocksdb"
        file.size = 104857600

or programmatically through the method StreamsExecutionEnvironment#addConfiguration:

    .addConfiguration(new RocksDBConfig()
          .withBlockCacheSize(500 * 1024 * 1024)

Please read the official documentation for more information: RocksDB Tuning Guide

Last modified March 11, 2021: fix(site): fix RocksDB config table (a6f4f624)