Kafka Streams Memory Management

You can specify the total memory (RAM) size used for internal caching and compacting of records. This caching happens before the records are written to state stores or forwarded downstream to other nodes.

The record caches are implemented slightly different in the DSL and Processor API.

Record caches in the DSL

You can specify the total memory (RAM) size of the record cache for an instance of the processing topology. It is leveraged by the following KTable instances:

  • Source KTable: KTable instances that are created via StreamsBuilder#table() or StreamsBuilder#globalTable().
  • Aggregation KTable: instances of KTable that are created as a result of aggregations.

For such KTable instances, the record cache is used for:

  • Internal caching and compacting of output records before they are written by the underlying stateful processor node to its internal state stores.
  • Internal caching and compacting of output records before they are forwarded from the underlying stateful processor node to any of its downstream processor nodes.

Use the following example to understand the behaviors with and without record caching. In this example, the input is a KStream<String, Integer> with the records <K,V>: <A, 1>, <D, 5>, <A, 20>, <A, 300>. The focus in this example is on the records with key == A.

  • An aggregation computes the sum of record values, grouped by key, for the input and returns a KTable<String, Integer>.
    • Without caching: a sequence of output records is emitted for key A that represent changes in the resulting aggregation table. The parentheses (()) denote changes, the left number is the new aggregate value and the right number is the old aggregate value: <A, (1, null)>, <A, (21, 1)>, <A, (321, 21)>.
    • With caching: a single output record is emitted for key A that would likely be compacted in the cache, leading to a single output record of <A, (321, null)>. This record is written to the aggregation’s internal state store and forwarded to any downstream operations.

The cache size is specified through the cache.max.bytes.buffering parameter, which is a global setting per processing topology:

// Enable record cache of size 10 MB.
Properties streamsConfiguration = new Properties();
streamsConfiguration.put(StreamsConfig.CACHE_MAX_BYTES_BUFFERING_CONFIG, 10 * 1024 * 1024L);

This parameter controls the number of bytes allocated for caching. Specifically, for a processor topology instance with T threads and C bytes allocated for caching, each thread will have an even C/T bytes to construct its own cache and use as it sees fit among its tasks. This means that there are as many caches as there are threads, but no sharing of caches across threads happens.

The basic API for the cache is made of put() and get() calls. Records are evicted using a simple LRU scheme after the cache size is reached. The first time a keyed record R1 = <K1, V1> finishes processing at a node, it is marked as dirty in the cache. Any other keyed record R2 = <K1, V2> with the same key K1 that is processed on that node during that time will overwrite <K1, V1>, this is referred to as “being compacted”. This has the same effect as Kafka’s log compaction, but happens earlier, while the records are still in memory, and within your client-side application, rather than on the server-side (i.e. the Apache Kafka® broker). After flushing, R2 is forwarded to the next processing node and then written to the local state store.

The semantics of caching is that data is flushed to the state store and forwarded to the next downstream processor node whenever the earliest of commit.interval.ms or cache.max.bytes.buffering (cache pressure) hits. Both commit.interval.ms and cache.max.bytes.buffering are global parameters. As such, it is not possible to specify different parameters for individual nodes.

Here are example settings for both parameters based on desired scenarios.

  • To turn off caching the cache size can be set to zero:

    // Disable record cache
Properties streamsConfiguration = new Properties();
streamsConfiguration.put(StreamsConfig.CACHE_MAX_BYTES_BUFFERING_CONFIG, 0);

    With default settings, caching is enabled in Kafka Streams and RocksDB.

  • To enable caching but still have an upper bound on how long records will be cached, you can set the commit interval. In this example, it is set to 1000 milliseconds:

    Properties streamsConfiguration = new Properties();
// Enable record cache of size 10 MB.
streamsConfiguration.put(StreamsConfig.CACHE_MAX_BYTES_BUFFERING_CONFIG, 10 * 1024 * 1024L);
// Set commit interval to 1 second.
streamsConfiguration.put(StreamsConfig.COMMIT_INTERVAL_MS_CONFIG, 1000);

The effect of these two configurations is described in the figure below. The records are shown using 4 keys: blue, red, yellow, and green. Assume the cache has space for only 3 keys.

  • When the cache is disabled (a), all of the input records will be output.
  • When the cache is enabled (b):
    • Most records are output at the end of commit intervals (e.g., at t1 a single blue record is output, which is the final over-write of the blue key up to that time).
    • Some records are output because of cache pressure (i.e. before the end of a commit interval). For example, see the red record before t2. With smaller cache sizes we expect cache pressure to be the primary factor that dictates when records are output. With large cache sizes, the commit interval will be the primary factor.
    • The total number of records output has been reduced from 15 to 8.
../../_images/streams-cache-and-commit-interval.png

Record caches in the Processor API

You can specify the total memory (RAM) size of the record cache for an instance of the processing topology. It is used for internal caching and compacting of output records before they are written from a stateful processor node to its state stores.

The record cache in the Processor API does not cache or compact any output records that are being forwarded downstream. This means that all downstream processor nodes can see all records, whereas the state stores see a reduced number of records. This does not impact correctness of the system, but is a performance optimization for the state stores. For example, with the Processor API you can store a record in a state store while forwarding a different value downstream.

Following from the example first shown in section State Stores, to enable caching, you can add the withCachingEnabled call.

StoreBuilder countStoreBuilder =
  Stores.keyValueStoreBuilder(
    Stores.persistentKeyValueStore("Counts"),
    Serdes.String(),
    Serdes.Long())
  .withCachingEnabled()

RocksDB

Each instance of RocksDB allocates off-heap memory for a block cache, index and filter blocks, and memtable (write buffer). Critical configs (for RocksDB version 4.1.0) include block_cache_size, write_buffer_size and max_write_buffer_number. These can be specified through the rocksdb.config.setter configuration.

Also, we recommend changing RocksDB’s default memory allocator, because the default allocator may lead to increased memory consumption. To change the memory allocator to jemalloc, set the LD_PRELOAD environment variable before you start your Kafka Streams application:

# example: install jemalloc (on Debian)
$ apt install -y libjemalloc-dev

# set LD_PRELOAD before you start your Kafka Streams application
$ export LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libjemalloc.so"

The RocksDB block cache only serves reads and doesn’t soak up writes. Writes always go into the memtable. Write traffic to disk is determined by the memtable being flushed, which happens during a commit or when the memtables are “full”. By default, you need 3 16MB memtables to fill up before flushing. You may be able to determine the right size by leveraging the RocksDB statistics to determine which limit you hit first: the commit interval or the memtable size limit.

The default block cache size is 50MB per store, but the Kafka Streams default is 10MB for caching for the entire instance. If you have a large number of stores, this 50MB default can be too high.

As of 5.3.0, the memory usage across all instances can be bounded, limiting the total off-heap memory of your Kafka Streams application. To do so you must configure RocksDB to cache the index and filter blocks in the block cache, limit the memtable memory through a shared WriteBufferManager and count its memory against the block cache, and then pass the same Cache object to each instance. See RocksDB Memory Usage for details. An example RocksDBConfigSetter implementing this is shown below:

public static class BoundedMemoryRocksDBConfig implements RocksDBConfigSetter {

  // See #1 below
  private static org.rocksdb.Cache cache = new org.rocksdb.LRUCache(TOTAL_OFF_HEAP_MEMORY, -1, false, INDEX_FILTER_BLOCK_RATIO);
  private static org.rocksdb.WriteBufferManager writeBufferManager = new org.rocksdb.WriteBufferManager(TOTAL_MEMTABLE_MEMORY, cache);

  @Override
  public void setConfig(final String storeName, final Options options, final Map<String, Object> configs) {

    BlockBasedTableConfig tableConfig = (BlockBasedTableConfig) options.tableFormatConfig();

    // These three options in combination will limit the memory used by RocksDB to the size passed to the block cache (TOTAL_OFF_HEAP_MEMORY)
    tableConfig.setBlockCache(cache);
    tableConfig.setCacheIndexAndFilterBlocks(true);
    options.setWriteBufferManager(writeBufferManager);

    // These options are recommended to be set when bounding the total memory
    // See #2 below
    tableConfig.setCacheIndexAndFilterBlocksWithHighPriority(true);
    tableConfig.setPinTopLevelIndexAndFilter(true);
    // See #3 below
    tableConfig.setBlockSize(BLOCK_SIZE);
    options.setMaxWriteBufferNumber(N_MEMTABLES);
    options.setWriteBufferSize(MEMTABLE_SIZE);

    options.setTableFormatConfig(tableConfig);
  }

  @Override
  public void close(final String storeName, final Options options) {
    // Cache and WriteBufferManager should not be closed here, as the same objects are shared by every store instance.
  }
}
Footnotes:
  1. INDEX_FILTER_BLOCK_RATIO can be used to set a fraction of the block cache to set aside for “high priority” (aka index and filter) blocks, preventing them from being evicted by data blocks. See the full signature of the LRUCache constructor. NOTE: the boolean parameter in the cache constructor lets you control whether the cache should enforce a strict memory limit by failing the read or iteration in the rare cases where it might go larger than its capacity. Due to a bug in RocksDB, this option cannot be used if the write buffer memory is also counted against the cache. If you set this to true, you should NOT pass the cache in to the WriteBufferManager and just control the write buffer and cache memory separately.
  2. This must be set in order for INDEX_FILTER_BLOCK_RATIO to take effect (see footnote 1) as described in the RocksDB docs.
  3. You may want to modify the default block size per these instructions from the RocksDB GitHub. A larger block size means index blocks will be smaller, but the cached data blocks may contain more cold data that would otherwise be evicted.

Note: While we recommend setting at least the above configs, the specific options that yield the best performance are workload dependent and you should consider experimenting with these to determine the best choices for your specific use case. Keep in mind that the optimal configs for one app may not apply to one with a different topology or input topic. In addition to the recommended configs above, you may want to consider using partitioned index filters as described by the RocksDB docs.

Other memory usage

There are other modules inside Kafka that allocate memory during runtime. They include the following:

  • Producer buffering, managed by the producer config buffer.memory.
  • Consumer buffering, currently not strictly managed, but can be indirectly controlled by fetch size, i.e., fetch.max.bytes and fetch.max.wait.ms.
  • Both producer and consumer also have separate TCP send / receive buffers that are not counted as the buffering memory. These are controlled by the send.buffer.bytes / receive.buffer.bytes configs.
  • Deserialized objects buffering: after consumer.poll() returns records, they will be deserialized to extract timestamp and buffered in the streams space. Currently this is only indirectly controlled by buffered.records.per.partition.

Tip

Iterators should be closed explicitly to release resources: Store iterators (e.g., KeyValueIterator and WindowStoreIterator) must be closed explicitly upon completeness to release resources such as open file handlers and in-memory read buffers, or use try-with-resources statement (available since JDK7) for this Closeable class.

Otherwise, stream application’s memory usage keeps increasing when running until it hits an OOM.

Note

This website includes content developed at the Apache Software Foundation under the terms of the Apache License v2.