com.yammer.metrics.stats.ExponentiallyDecayingSample Maven / Gradle / Ivy
package com.yammer.metrics.stats;
import java.util.ArrayList;
import java.util.concurrent.ConcurrentSkipListMap;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicLong;
import java.util.concurrent.locks.ReentrantReadWriteLock;
import static java.lang.Math.*;
/**
* An exponentially-decaying random sample of {@code long}s. Uses Cormode et al's forward-decaying
* priority reservoir sampling method to produce a statistically representative sample,
* exponentially biased towards newer entries.
*
* @see
* Cormode et al. Forward Decay: A Practical Time Decay Model for Streaming Systems. ICDE '09:
* Proceedings of the 2009 IEEE International Conference on Data Engineering (2009)
*/
public class ExponentiallyDecayingSample implements Sample {
private static final long RESCALE_THRESHOLD = TimeUnit.HOURS.toNanos(1);
private final ConcurrentSkipListMap values;
private final ReentrantReadWriteLock lock;
private final double alpha;
private final int reservoirSize;
private final AtomicLong count = new AtomicLong(0);
private volatile long startTime;
private final AtomicLong nextScaleTime = new AtomicLong(0);
/**
* Creates a new {@link ExponentiallyDecayingSample}.
*
* @param reservoirSize the number of samples to keep in the sampling reservoir
* @param alpha the exponential decay factor; the higher this is, the more biased the
* sample will be towards newer values
*/
public ExponentiallyDecayingSample(int reservoirSize, double alpha) {
this.values = new ConcurrentSkipListMap();
this.lock = new ReentrantReadWriteLock();
this.alpha = alpha;
this.reservoirSize = reservoirSize;
clear();
}
@Override
public void clear() {
lockForRescale();
try {
values.clear();
count.set(0);
this.startTime = tick();
nextScaleTime.set(System.nanoTime() + RESCALE_THRESHOLD);
} finally {
unlockForRescale();
}
}
@Override
public int size() {
return (int) min(reservoirSize, count.get());
}
@Override
public void update(long value) {
update(value, tick());
}
/**
* Adds an old value with a fixed timestamp to the sample.
*
* @param value the value to be added
* @param timestamp the epoch timestamp of {@code value} in seconds
*/
public void update(long value, long timestamp) {
lockForRegularUsage();
try {
final double priority = weight(timestamp - startTime) / ThreadLocalRandom.current().nextDouble();
final long newCount = count.incrementAndGet();
if (newCount <= reservoirSize) {
values.put(priority, value);
} else {
Double first = values.firstKey();
if (first < priority) {
if (values.putIfAbsent(priority, value) == null) {
// ensure we always remove an item
while (values.remove(first) == null) {
first = values.firstKey();
}
}
}
}
} finally {
unlockForRegularUsage();
}
final long now = System.nanoTime();
final long next = nextScaleTime.get();
if (now >= next) {
rescale(now, next);
}
}
@Override
public Snapshot getSnapshot() {
lockForRegularUsage();
try {
return new Snapshot(values.values());
} finally {
unlockForRegularUsage();
}
}
private long tick() {
return TimeUnit.MILLISECONDS.toSeconds(System.currentTimeMillis());
}
private double weight(long t) {
return exp(alpha * t);
}
/* "A common feature of the above techniques—indeed, the key technique that
* allows us to track the decayed weights efficiently—is that they maintain
* counts and other quantities based on g(ti − L), and only scale by g(t − L)
* at query time. But while g(ti −L)/g(t−L) is guaranteed to lie between zero
* and one, the intermediate values of g(ti − L) could become very large. For
* polynomial functions, these values should not grow too large, and should be
* effectively represented in practice by floating point values without loss of
* precision. For exponential functions, these values could grow quite large as
* new values of (ti − L) become large, and potentially exceed the capacity of
* common floating point types. However, since the values stored by the
* algorithms are linear combinations of g values (scaled sums), they can be
* rescaled relative to a new landmark. That is, by the analysis of exponential
* decay in Section III-A, the choice of L does not affect the final result. We
* can therefore multiply each value based on L by a factor of exp(−α(L′ − L)),
* and obtain the correct value as if we had instead computed relative to a new
* landmark L′ (and then use this new L′ at query time). This can be done with
* a linear pass over whatever data structure is being used."
*/
private void rescale(long now, long next) {
if (nextScaleTime.compareAndSet(next, now + RESCALE_THRESHOLD)) {
lockForRescale();
try {
final long oldStartTime = startTime;
this.startTime = tick();
final ArrayList keys = new ArrayList(values.keySet());
for (Double key : keys) {
final Long value = values.remove(key);
values.put(key * exp(-alpha * (startTime - oldStartTime)), value);
}
} finally {
unlockForRescale();
}
}
}
private void unlockForRescale() {
lock.writeLock().unlock();
}
private void lockForRescale() {
lock.writeLock().lock();
}
private void lockForRegularUsage() {
lock.readLock().lock();
}
private void unlockForRegularUsage() {
lock.readLock().unlock();
}
}