Statistic collection with the IOStatistics API

@InterfaceAudience.Public
@InterfaceStability.Unstable

The IOStatistics API is intended to provide statistics on individual IO classes -such as input and output streams, in a standard way which applications can query

Many filesystem-related classes have implemented statistics gathering and provided private/unstable ways to query this, but as they were not common across implementations it was unsafe for applications to reference these values. Example: S3AInputStream and its statistics API. This is used in internal tests, but cannot be used downstream in applications such as Apache Hive or Apache HBase.

The IOStatistics API is intended to

  1. Be instance specific:, rather than shared across multiple instances of a class, or thread local.
  2. Be public and stable enough to be used by applications.
  3. Be easy to use in applications written in Java, Scala, and, via libhdfs, C/C++
  4. Have foundational interfaces and classes in the hadoop-common JAR.

Core Model

Any class may implement IOStatisticsSource in order to provide statistics.

Wrapper I/O Classes such as FSDataInputStream anc FSDataOutputStream should implement the interface and forward it to the wrapped class, if they also implement it -and return null if they do not.

IOStatisticsSource implementations getIOStatistics() return an instance of IOStatistics enumerating the statistics of that specific instance.

The IOStatistics Interface exports five kinds of statistic:

Category Type Description
counter long a counter which may increase in value; SHOULD BE >= 0
gauge long an arbitrary value which can down as well as up; SHOULD BE >= 0
minimum long an minimum value; MAY BE negative
maximum long a maximum value; MAY BE negative
meanStatistic MeanStatistic an arithmetic mean and sample size; mean MAY BE negative

Four are simple long values, with the variations how they are likely to change and how they are aggregated.

Aggregation of Statistic Values

For the different statistic category, the result of aggregate(x, y) is

Category Aggregation
counter max(0, x) + max(0, y)
gauge max(0, x) + max(0, y)
minimum min(x, y)
maximum max(x, y)
meanStatistic calculation of the mean of x and y )

Class MeanStatistic

package org.apache.hadoop.fs.statistics

This package contains the public statistics APIs intended for use by applications.

MeanStatistic is a tuple of (mean, samples) to support aggregation.

A MeanStatistic with a sample of 0 is considered an empty statistic.

All MeanStatistic instances where sample = 0 are considered equal, irrespective of the mean value.

Algorithm to calculate the mean :

if x.samples = 0:
    y
else if y.samples = 0 :
    x
else:
    samples' = x.samples + y.samples
    mean' = (x.mean * x.samples) + (y.mean * y.samples) / samples'
    (samples', mean')

Implicitly, this means that if both samples are empty, then the aggregate value is also empty.

public final class MeanStatistic implements Serializable, Cloneable {
  /**
   * Arithmetic mean.
   */
  private double mean;

  /**
   * Number of samples used to calculate
   * the mean.
   */
  private long samples;

  /**
   * Get the mean value.
   * @return the mean
   */
  public double getMean() {
    return mean;
  }

  /**
   * Get the sample count.
   * @return the sample count; 0 means empty
   */
  public long getSamples() {
    return samples;
  }

  /**
   * Is a statistic empty?
   * @return true if the sample count is 0
   */
  public boolean isEmpty() {
    return samples == 0;
  }
   /**
   * Add another mean statistic to create a new statistic.
   * When adding two statistics, if either is empty then
   * a copy of the non-empty statistic is returned.
   * If both are empty then a new empty statistic is returned.
   *
   * @param other other value
   * @return the aggregate mean
   */
  public MeanStatistic add(final MeanStatistic other) {
    /* Implementation elided. */
  }
  @Override
  public int hashCode() {
    return Objects.hash(mean, samples);
  }

  @Override
  public boolean equals(final Object o) {
    if (this == o) { return true; }
    if (o == null || getClass() != o.getClass()) { return false; }
    MeanStatistic that = (MeanStatistic) o;
    if (this.isEmpty()) {
      return that.isEmpty();
    }
    return Double.compare(that.mean, mean) == 0 &&
        samples == that.samples;
  }

  @Override
  public MeanStatistic clone() {
    return new MeanStatistic(this);
  }

  public MeanStatistic copy() {
    return new MeanStatistic(this);
  }

}

class org.apache.hadoop.fs.statistics.IOStatisticsSource

/**
 * A source of IO statistics.
 * These statistics MUST be instance specific, not thread local.
 */
@InterfaceStability.Unstable
public interface IOStatisticsSource {

  /**
   * Return a statistics instance.
   * It is not a requirement that the same instance is returned every time.
   * {@link IOStatisticsSource}.
   * If the object implementing this is Closeable, this method
   * may return null if invoked on a closed object, even if
   * it returns a valid instance when called earlier.
   * @return an IOStatistics instance or null
   */
  IOStatistics getIOStatistics();
}

This is the interface which an object instance MUST implement if they are a source of IOStatistics information.

Invariants

The result of getIOStatistics() must be one of

  • null
  • an immutable IOStatistics for which each map of entries is an empty map.
  • an instance of an IOStatistics whose statistics MUST BE unique to that instance of the class implementing IOStatisticsSource.

Less formally: if the statistics maps returned are non-empty, all the statistics must be collected from the current instance, and not from any other instances, the way some of the FileSystem statistics are collected.

The result of getIOStatistics(), if non-null, MAY be a different instance on every invocation.

class org.apache.hadoop.fs.statistics.IOStatistics

These are per-instance statistics provided by an object which implements IOStatisticsSource.

@InterfaceAudience.Public
@InterfaceStability.Unstable
public interface IOStatistics {

  /**
   * Map of counters.
   * @return the current map of counters.
   */
  Map<String, Long> counters();

  /**
   * Map of gauges.
   * @return the current map of gauges.
   */
  Map<String, Long> gauges();

  /**
   * Map of minumums.
   * @return the current map of minumums.
   */
  Map<String, Long> minumums();

  /**
   * Map of maximums.
   * @return the current map of maximums.
   */
  Map<String, Long> maximums();

  /**
   * Map of meanStatistics.
   * @return the current map of MeanStatistic statistics.
   */
  Map<String, MeanStatistic> meanStatistics();

}

Statistic Naming

The naming policy of statistics is designed to be readable, shareable and ideally consistent across IOStatisticSource implementations.

  • Characters in key names MUST match the regular expression [a-z|0-9|_] with the exception of the first character, which MUST be in the range [a-z]. Thus the full regular expression for a valid statistic name is:

    [a-z][a-z|0-9|_]+

  • Where possible, the names of statistics SHOULD be those defined with common names.

    org.apache.hadoop.fs.statistics.StreamStatisticNames
org.apache.hadoop.fs.statistics.StoreStatisticNames

Note 1.: these are evolving; for clients to safely reference their statistics by name they SHOULD be copied to the application. (i.e. for an application compiled hadoop 3.4.2 to link against hadoop 3.4.1, copy the strings).

Note 2: keys defined in these classes SHALL NOT be removed from subsequent Hadoop releases.

  • A common statistic name MUST NOT be used to report any other statistic and MUST use the pre-defined unit of measurement.

  • A statistic name in one of the maps SHOULD NOT be re-used in another map. This aids diagnostics of logged statistics.

Statistic Maps

For each map of statistics returned:

  • The operations to add/remove entries are unsupported: the map returned MAY be mutable by the source of statistics.

  • The map MAY be empty.

  • The map keys each represent a measured statistic.

  • The set of keys in a map SHOULD remain unchanged, and MUST NOT remove keys.

  • The statistics SHOULD be dynamic: every lookup of an entry SHOULD return the latest value.

  • The values MAY change across invocations of Map.values() and Map.entries()

  • The update MAY be in the iterable() calls of the iterators returned, or MAY be in the actual iterable.next() operation. That is: there is no guarantee as to when the evaluation takes place.

  • The returned Map.Entry instances MUST return the same value on repeated getValue() calls. (i.e once you have the entry, it is immutable).

  • Queries of statistics SHOULD be fast and non-blocking to the extent that if invoked during a long operation, they will prioritize returning fast over most timely values.

  • The statistics MAY lag; especially for statistics collected in separate operations (e.g stream IO statistics as provided by a filesystem instance).

  • Statistics which represent time SHOULD use milliseconds as their unit.

  • Statistics which represent time and use a different unit MUST document the unit used.

Thread Model

  1. An instance of IOStatistics can be shared across threads;

  2. Read access to the supplied statistics maps MUST be thread safe.

  3. Iterators returned from the maps MUST NOT be shared across threads.

  4. The statistics collected MUST include all operations which took place across all threads performing work for the monitored object.

  5. The statistics reported MUST NOT be local to the active thread.

This is different from the FileSystem.Statistics behavior where per-thread statistics are collected and reported.

That mechanism supports collecting limited read/write statistics for different worker threads sharing the same FS instance, but as the collection is thread local, it invariably under-reports IO performed in other threads on behalf of a worker thread.

Statisic Snapshot

A snapshot of the current statistic values MAY be obtained by calling IOStatisticsSupport.snapshotIOStatistics()

  public static <X extends IOStatistics & Serializable> X
      snapshotIOStatistics(IOStatistics statistics)

This snapshot is serializable through Java serialization and through Jackson to/from JSON.

Helper Classes

class org.apache.hadoop.fs.statistics.IOStatisticsSupport

This provides helper methods to work with IOStatistics sources and instances.

Consult the javadocs for its operations.

class org.apache.hadoop.fs.statistics.IOStatisticsLogging

Support for efficiently logging IOStatistics/IOStatisticsSource instances.

These are intended for assisting logging, including only enumerating the state of an IOStatistics instance when the log level needs it.

LOG.info("IOStatistics after upload: {}", demandStringify(iostats));

// or even better, as it results in only a single object creations
Object latest = demandStringify(iostats);
LOG.info("IOStatistics : {}", latest);
/* do some work. */
LOG.info("IOStatistics : {}", latest);

Package org.apache.hadoop.fs.statistics.impl

This contains implementation classes to support providing statistics to applications.

These MUST NOT BE used by applications. If a feature is needed from this package then the provisioning of a public implementation MAY BE raised via the Hadoop development channels.

These MAY be used by those implementations of the Hadoop FileSystem, AbstractFileSystem and related classes which are not in the hadoop source tree. Implementors MUST BE aware that the implementation this code is unstable and may change across minor point releases of Hadoop.