An exponentially weighted moving average implementation that decays based on the elapsed time since the last update, approximating a time windowed moving average.

MovingAverage.java

/**
 * An exponentially weighted moving average implementation that decays based on the elapsed time since the last update,
 * approximating a time windowed moving average.
 */
public class MovingAverage {
  private final long windowNanos;

  // Mutable state
  private volatile long lastNanos;
  private volatile double average;

  /**
   * Creates a moving average of samples over a {@code window} for the {@code timeUnit}.
   */
  public MovingAverage(long window, TimeUnit timeUnit) {
    this.windowNanos = timeUnit.toNanos(window);
  }

  /**
   * Updates the average with the {@code sample}.
   */
  public synchronized void update(double sample) {
    long now = System.nanoTime();

    if (lastNanos == 0) {
      average = sample;
      lastNanos = now;
      return;
    }

    long elapsedNanos = now - lastNanos;
    double coeff = Math.exp(-1.0 * ((double) elapsedNanos / windowNanos));
    average = (1.0 - coeff) * sample + coeff * average;
    lastNanos = now;
  }

  /**
   * Returns the moving average.
   */
  public double get() {
    return average;
  }

  /**
   * Tick the internal moving average clock back by the {@code duration}. Useful for testing.
   */
  void tick(long duration, TimeUnit timeUnit) {
    if (lastNanos != 0)
      lastNanos -= timeUnit.toNanos(duration);
  }
}