CMSStore.scala

/*
 * Copyright Confluent Inc.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *    http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
package io.confluent.examples.streams.algebird

import com.twitter.algebird.{CMSHasher, TopCMS, TopPctCMS}
import org.apache.kafka.common.serialization.Serdes
import org.apache.kafka.streams.processor.{StateStore, StateStoreContext}
import org.apache.kafka.streams.state.StateSerdes

/**
  * An in-memory store that leverages the Count-Min Sketch implementation of
  * [[https://github.com/twitter/algebird Twitter Algebird]].
  *
  * This store allows you to probabilistically count items of type T with a
  * [[https://en.wikipedia.org/wiki/Count%E2%80%93min_sketch Count-Min Sketch]] data structure.
  * Here, the counts returned by the store will be approximate counts, i.e. estimations, because a
  * Count-Min Sketch trades slightly inaccurate counts for greatly reduced space utilization
  * (however, the estimation error is mathematically proven to be bounded).
  * With probability at least `1 - delta`, this estimate is within `eps * N` of the true frequency
  * (i.e., `true frequency <= estimate <= true frequency + eps * N`), where `N` is the total number
  * of items counted ("seen" in the input) so far (cf. [[CMSStore#totalCount]]).
  *
  * A traditional Count-Min Sketch is a fixed-size data structure that is essentially an array of
  * counters of a particular width (derived from the parameter `eps`) and depth (derived from the
  * parameter `delta`).  The CMS variant used in this store, [[TopPctCMS]], additionally tracks the
  * so-called "heavy hitters" among the counted items (i.e. the items with the largest counts) based
  * on a percentage threshold; the size of heavy hitters is still bounded, however, hence the total
  * size of the [[TopPctCMS]] data structure is still fixed.
  *
  * =Fault-tolerance=
  *
  * This store supports changelogging its state to Kafka and is thus fault-tolerant.  Every time the
  * store is flushed (cf. [[org.apache.kafka.streams.StreamsConfig.COMMIT_INTERVAL_MS_CONFIG]]) the
  * underlying CMS data structure is written to the store's changelog topic.  For many use cases
  * this approach should be sufficiently efficient because the absolute size of a CMS is typically
  * rather small (a few KBs up to a megabyte, depending on the CMS settings, which are determined by
  * e.g. your error bound requirements for approximate counts).
  *
  * =Usage=
  *
  * Note: Twitter Algebird is best used with Scala, so all the examples below are in Scala, too.
  *
  * In a Kafka Streams application, you'd typically create this store as such:
  *
  * {{{
  * val builder: StreamsBuilder = new StreamsBuilder()
  *
  * // In this example, we create a store for type [[String]].
  * // It's recommended to reduce Kafka's log segment size for the changelogs of CMS stores, which
  * // you can do by passing the respective Kafka setting to the CMSStoreBuilder via `withLoggingEnabled()`.
  * builder.addStateStore(new CMSStoreBuilder[String]("my-cms-store-name", Serdes.String()))
  * }}}
  *
  * Then you'd use the store within a [[org.apache.kafka.streams.processor.Processor]] or a
  * [[org.apache.kafka.streams.kstream.Transformer]] similar to:
  *
  * {{{
  * class ProbabilisticCounter extends Transformer[Array[Byte], String, KeyValue[String, Long]] {
  *
  *   private var cmsState: CMSStore[String] = _
  *   private var processorContext: ProcessorContext = _
  *
  *   override def init(processorContext: ProcessorContext): Unit = {
  *     this.processorContext = processorContext
  *     cmsState = this.processorContext.getStateStore("my-cms-store-name").asInstanceOf[CMSStore[String]]
  *   }
  *
  *   override def transform(key: Array[Byte], value: String): KeyValue[String, Long] = {
  *     // Count the record value, think: "+ 1"
  *     cmsState.put(value)
  *
  *     // Emit the latest count estimate for the record value
  *     KeyValue.pair[String, Long](value, cmsState.get(value))
  *   }
  *
  *   override def punctuate(l: Long): KeyValue[String, Long] = null
  *
  *   override def close(): Unit = {}
  * }
  * }}}
  *
  * @param name            The name of this store instance
  * @param loggingEnabled  Whether or not changelogging (fault-tolerance) is enabled for this store.
  * @param delta           CMS parameter: A bound on the probability that a query estimate does not
  *                        lie within some small interval (an interval that depends on `eps`) around
  *                        the truth.
  *                        See [[TopPctCMS]] and [[com.twitter.algebird.CMSMonoid]].
  * @param eps             CMS parameter: One-sided error bound on the error of each point query,
  *                        i.e. frequency estimate.
  *                        See [[TopPctCMS]] and [[com.twitter.algebird.CMSMonoid]].
  * @param seed            CMS parameter: A seed to initialize the random number generator used to
  *                        create the pairwise independent hash functions.  Typically you do not
  *                        need to change this.
  *                        See [[TopPctCMS]] and [[com.twitter.algebird.CMSMonoid]].
  * @param heavyHittersPct CMS parameter: A threshold for finding heavy hitters, i.e., items that
  *                        appear at least (heavyHittersPct * totalCount) times in the stream.
  *                        Every item that appears at least `(heavyHittersPct * totalCount)` times
  *                        is included, and with probability `p >= 1 - delta`, no item whose count
  *                        is less than `(heavyHittersPct - eps) * totalCount` is included.
  *                        This also means that this parameter is an upper bound on the number of
  *                        heavy hitters that will be tracked: the set of heavy hitters contains at
  *                        most `1 / heavyHittersPct` elements.  For example, if
  *                        `heavyHittersPct=0.01` (or 0.25), then at most `1 / 0.01 = 100` items
  *                        or `1 / 0.25 = 4` items) will be tracked/returned as heavy hitters.
  *                        This parameter can thus control the memory footprint required for
  *                        tracking heavy hitters.
  *                        See [[TopPctCMS]] and [[com.twitter.algebird.TopPctCMSMonoid]].
  * @tparam T The type used to identify the items to be counted with the CMS.  For example, if
  *           you want to count the occurrence of user names, you could use count user names
  *           directly with `T=String`;  alternatively, you could map each username to a unique
  *           numeric ID expressed as a `Long`, and then count the occurrences of those `Long`s with
  *           a CMS of type `T=Long`.  Note that such a mapping between the items of your problem
  *           domain and their identifiers used for counting via CMS should be bijective.
  *           We require a [[CMSHasher]] context bound for `K`, see [[CMSHasher]] for available
  *           implicits that can be imported.
  *           See [[com.twitter.algebird.CMSMonoid]] for further information.
  */
class CMSStore[T: CMSHasher](override val name: String,
                             val loggingEnabled: Boolean = true,
                             val delta: Double = 1E-10,
                             val eps: Double = 0.001,
                             val seed: Int = 1,
                             val heavyHittersPct: Double = 0.01)
    extends StateStore {

  private val cmsMonoid = TopPctCMS.monoid[T](eps, delta, seed, heavyHittersPct)

  /**
    * The "storage backend" of this store.
    *
    * Needs proper initializing in case the store's changelog is empty.
    */
  private var cms: TopCMS[T] = cmsMonoid.zero

  private var timestampOfLastStateStoreUpdate: Long = 0L

  private var changeLogger: CMSStoreChangeLogger[Integer, TopCMS[T]] = _

  /**
    * The record key used to write to the state's changelog.
    *
    * This key can be a constant because:
    *
    * 1. We always write the full CMS when writing to the changelog.
    * 2. A CMS does not retain information about which items were counted, i.e. it does not track
    * information about the keyspace (in the case of this store, the only information about the
    * keyspace are the heavy hitters); so, unless we opted for a different approach than (1)
    * above, we cannot leverage keyspace information anyways.
    * 3. We use a [[CMSStoreChangeLogger]] that uses a stream task's
    * [[org.apache.kafka.streams.processor.TaskId]] to identify the changelog partition to write to.
    * Thus only one particular stream task will ever be writing to that changelog partition.
    * 4. When restoring from the changelog, a stream task will read only its own (one) changelog
    * partition.
    *
    * In other words, we can hardcode the changelog key because only the "right" stream task will be
    * (a) writing to AND (b) reading from the respective partition of the changelog.
    */
  private[algebird] val changelogKey = 42

  /**
    * For unit testing
    */
  private[algebird] def cmsFrom(items: Seq[T]): TopCMS[T] = cmsMonoid.create(items)

  /**
    * For unit testing
    */
  private[algebird] def cmsFrom(item: T): TopCMS[T] = cmsMonoid.create(item)

  @volatile private var open: Boolean = false

  /**
    * Initializes this store, including restoring the store's state from its changelog.
    */
  override def init(context: StateStoreContext, root: StateStore): Unit = {
    val serdes = new StateSerdes[Integer, TopCMS[T]](
      name,
      Serdes.Integer(),
      TopCMSSerde[T])
    changeLogger = new CMSStoreChangeLogger[Integer, TopCMS[T]](name, context, serdes, name)

    // Note: We must manually guard with `loggingEnabled` here because `context.register()` ignores
    // that parameter.
    if (root != null && loggingEnabled) {
      context.register(root, (_, value) => {
        if (value == null) {
          cms = cmsMonoid.zero
        }
        else {
          cms = serdes.valueFrom(value)
        }
      })
    }

    open = true
  }

  /**
    * Returns the estimated count of the item.
    *
    * @param item item to be counted
    * @return estimated count
    */
  def get(item: T): Long = cms.frequency(item).estimate

  /**
    * Counts the item.
    *
    * @param item item to be counted
    */
  def put(item: T, timestamp: Long): Unit = {
    cms = cms + item
    timestampOfLastStateStoreUpdate = timestamp
  }

  /**
    * The top items counted so far, with the percentage-based cut-off being defined by the CMS
    * parameter `heavyHittersPct`.
    *
    * @return the top items counted so far
    */
  def heavyHitters: Set[T] = cms.heavyHitters

  /**
    * Returns the total number of items counted ("seen" in the input) so far.
    *
    * This number is not the same as the total number of <em>unique</em> items counted so far, i.e.
    * it is not the cardinality of the set of items.
    *
    * Example: After having counted the input "foo", "bar", "foo", the return value would be 3.
    *
    * @return number of count operations so far
    */
  def totalCount: Long = cms.totalCount

  override val persistent: Boolean = false

  override def isOpen: Boolean = open

  /**
    * Periodically saves the latest CMS state to Kafka.
    *
    * =Implementation detail=
    *
    * The changelog records have the form: (hardcodedKey, CMS).  That is, we are backing up the
    * underlying CMS data structure in its entirety to Kafka.
    */
  override def flush(): Unit = {
    if (loggingEnabled) {
      changeLogger.logChange(changelogKey, cms, timestampOfLastStateStoreUpdate)
    }
  }

  override def close(): Unit = {
    open = false
  }

}