package org.apache.lucene.facet.sampling;
import java.io.IOException;
import java.util.Arrays;
import java.util.logging.Level;
import java.util.logging.Logger;
import org.apache.lucene.facet.old.ScoredDocIDs;
import org.apache.lucene.facet.old.ScoredDocIDsIterator;
import org.apache.lucene.facet.old.ScoredDocIdsUtils;
import org.apache.lucene.util.PriorityQueue;
/*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You 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.
*/
/**
* Take random samples of large collections.
* @lucene.experimental
*/
public class RepeatableSampler extends Sampler {
private static final Logger logger = Logger.getLogger(RepeatableSampler.class.getName());
public RepeatableSampler(SamplingParams params) {
super(params);
}
@Override
protected SampleResult createSample(ScoredDocIDs docids, int actualSize,
int sampleSetSize) throws IOException {
int[] sampleSet = null;
try {
sampleSet = repeatableSample(docids, actualSize,
sampleSetSize);
} catch (IOException e) {
if (logger.isLoggable(Level.WARNING)) {
logger.log(Level.WARNING, "sampling failed: "+e.getMessage()+" - falling back to no sampling!", e);
}
return new SampleResult(docids, 1d);
}
ScoredDocIDs sampled = ScoredDocIdsUtils.createScoredDocIDsSubset(docids,
sampleSet);
if (logger.isLoggable(Level.FINEST)) {
logger.finest("******************** " + sampled.size());
}
return new SampleResult(sampled, sampled.size()/(double)docids.size());
}
/**
* Returns <code>sampleSize</code> values from the first <code>collectionSize</code>
* locations of <code>collection</code>, chosen using
* the <code>TRAVERSAL</code> algorithm. The sample values are not sorted.
* @param collection The values from which a sample is wanted.
* @param collectionSize The number of values (from the first) from which to draw the sample.
* @param sampleSize The number of values to return.
* @return An array of values chosen from the collection.
* @see Algorithm#TRAVERSAL
*/
private static int[] repeatableSample(ScoredDocIDs collection,
int collectionSize, int sampleSize)
throws IOException {
return repeatableSample(collection, collectionSize,
sampleSize, Algorithm.HASHING, Sorted.NO);
}
/**
* Returns <code>sampleSize</code> values from the first <code>collectionSize</code>
* locations of <code>collection</code>, chosen using <code>algorithm</code>.
* @param collection The values from which a sample is wanted.
* @param collectionSize The number of values (from the first) from which to draw the sample.
* @param sampleSize The number of values to return.
* @param algorithm Which algorithm to use.
* @param sorted Sorted.YES to sort the sample values in ascending order before returning;
* Sorted.NO to return them in essentially random order.
* @return An array of values chosen from the collection.
*/
private static int[] repeatableSample(ScoredDocIDs collection,
int collectionSize, int sampleSize,
Algorithm algorithm, Sorted sorted)
throws IOException {
if (collection == null) {
throw new IOException("docIdSet is null");
}
if (sampleSize < 1) {
throw new IOException("sampleSize < 1 (" + sampleSize + ")");
}
if (collectionSize < sampleSize) {
throw new IOException("collectionSize (" + collectionSize + ") less than sampleSize (" + sampleSize + ")");
}
int[] sample = new int[sampleSize];
long[] times = new long[4];
if (algorithm == Algorithm.TRAVERSAL) {
sample1(collection, collectionSize, sample, times);
} else if (algorithm == Algorithm.HASHING) {
sample2(collection, collectionSize, sample, times);
} else {
throw new IllegalArgumentException("Invalid algorithm selection");
}
if (sorted == Sorted.YES) {
Arrays.sort(sample);
}
if (returnTimings) {
times[3] = System.currentTimeMillis();
if (logger.isLoggable(Level.FINEST)) {
logger.finest("Times: " + (times[1] - times[0]) + "ms, "
+ (times[2] - times[1]) + "ms, " + (times[3] - times[2])+"ms");
}
}
return sample;
}
/**
* Returns <code>sample</code>.length values chosen from the first <code>collectionSize</code>
* locations of <code>collection</code>, using the TRAVERSAL algorithm. The sample is
* pseudorandom: no subset of the original collection
* is in principle more likely to occur than any other, but for a given collection
* and sample size, the same sample will always be returned. This algorithm walks the
* original collection in a methodical way that is guaranteed not to visit any location
* more than once, which makes sampling without replacement faster because removals don't
* have to be tracked, and the number of operations is proportional to the sample size,
* not the collection size.
* Times for performance measurement
* are returned in <code>times</code>, which must be an array of at least three longs, containing
* nanosecond event times. The first
* is set when the algorithm starts; the second, when the step size has been calculated;
* and the third when the sample has been taken.
* @param collection The set to be sampled.
* @param collectionSize The number of values to use (starting from first).
* @param sample The array in which to return the sample.
* @param times The times of three events, for measuring performance.
*/
private static void sample1(ScoredDocIDs collection, int collectionSize, int[] sample, long[] times)
throws IOException {
ScoredDocIDsIterator it = collection.iterator();
if (returnTimings) {
times[0] = System.currentTimeMillis();
}
int sampleSize = sample.length;
int prime = findGoodStepSize(collectionSize, sampleSize);
int mod = prime % collectionSize;
if (returnTimings) {
times[1] = System.currentTimeMillis();
}
int sampleCount = 0;
int index = 0;
for (; sampleCount < sampleSize;) {
if (index + mod < collectionSize) {
for (int i = 0; i < mod; i++, index++) {
it.next();
}
} else {
index = index + mod - collectionSize;
it = collection.iterator();
for (int i = 0; i < index; i++) {
it.next();
}
}
sample[sampleCount++] = it.getDocID();
}
if (returnTimings) {
times[2] = System.currentTimeMillis();
}
}
/**
* Returns a value which will allow the caller to walk
* a collection of <code>collectionSize</code> values, without repeating or missing
* any, and spanning the collection from beginning to end at least once with
* <code>sampleSize</code> visited locations. Choosing a value
* that is relatively prime to the collection size ensures that stepping by that size (modulo
* the collection size) will hit all locations without repeating, eliminating the need to
* track previously visited locations for a "without replacement" sample. Starting with the
* square root of the collection size ensures that either the first or second prime tried will
* work (they can't both divide the collection size). It also has the property that N steps of
* size N will span a collection of N**2 elements once. If the sample is bigger than N, it will
* wrap multiple times (without repeating). If the sample is smaller, a step size is chosen
* that will result in at least one spanning of the collection.
*
* @param collectionSize The number of values in the collection to be sampled.
* @param sampleSize The number of values wanted in the sample.
* @return A good increment value for walking the collection.
*/
private static int findGoodStepSize(int collectionSize, int sampleSize) {
int i = (int) Math.sqrt(collectionSize);
if (sampleSize < i) {
i = collectionSize / sampleSize;
}
do {
i = findNextPrimeAfter(i);
} while (collectionSize % i == 0);
return i;
}
/**
* Returns the first prime number that is larger than <code>n</code>.
* @param n A number less than the prime to be returned.
* @return The smallest prime larger than <code>n</code>.
*/
private static int findNextPrimeAfter(int n) {
n += (n % 2 == 0) ? 1 : 2; // next odd
foundFactor: for (;; n += 2) { //TODO labels??!!
int sri = (int) (Math.sqrt(n));
for (int primeIndex = 0; primeIndex < N_PRIMES; primeIndex++) {
int p = primes[primeIndex];
if (p > sri) {
return n;
}
if (n % p == 0) {
continue foundFactor;
}
}
for (int p = primes[N_PRIMES - 1] + 2;; p += 2) {
if (p > sri) {
return n;
}
if (n % p == 0) {
continue foundFactor;
}
}
}
}
/**
* The first N_PRIMES primes, after 2.
*/
private static final int N_PRIMES = 4000;
private static int[] primes = new int[N_PRIMES];
static {
primes[0] = 3;
for (int count = 1; count < N_PRIMES; count++) {
primes[count] = findNextPrimeAfter(primes[count - 1]);
}
}
/**
* Returns <code>sample</code>.length values chosen from the first <code>collectionSize</code>
* locations of <code>collection</code>, using the HASHING algorithm. Performance measurements
* are returned in <code>times</code>, which must be an array of at least three longs. The first
* will be set when the algorithm starts; the second, when a hash key has been calculated and
* inserted into the priority queue for every element in the collection; and the third when the
* original elements associated with the keys remaining in the PQ have been stored in the sample
* array for return.
* <P>
* This algorithm slows as the sample size becomes a significant fraction of the collection
* size, because the PQ is as large as the sample set, and will not do early rejection of values
* below the minimum until it fills up, and a larger PQ contains more small values to be purged,
* resulting in less early rejection and more logN insertions.
*
* @param collection The set to be sampled.
* @param collectionSize The number of values to use (starting from first).
* @param sample The array in which to return the sample.
* @param times The times of three events, for measuring performance.
*/
private static void sample2(ScoredDocIDs collection, int collectionSize, int[] sample, long[] times)
throws IOException {
if (returnTimings) {
times[0] = System.currentTimeMillis();
}
int sampleSize = sample.length;
IntPriorityQueue pq = new IntPriorityQueue(sampleSize);
/*
* Convert every value in the collection to a hashed "weight" value, and insert
* into a bounded PQ (retains only sampleSize highest weights).
*/
ScoredDocIDsIterator it = collection.iterator();
MI mi = null;
while (it.next()) {
if (mi == null) {
mi = new MI();
}
mi.value = (int) (it.getDocID() * PHI_32) & 0x7FFFFFFF;
mi = pq.insertWithOverflow(mi);
}
if (returnTimings) {
times[1] = System.currentTimeMillis();
}
/*
* Extract heap, convert weights back to original values, and return as integers.
*/
Object[] heap = pq.getHeap();
for (int si = 0; si < sampleSize; si++) {
sample[si] = (int)(((MI) heap[si+1]).value * PHI_32I) & 0x7FFFFFFF;
}
if (returnTimings) {
times[2] = System.currentTimeMillis();
}
}
/**
* A mutable integer that lets queue objects be reused once they start overflowing.
*/
private static class MI {
MI() { }
public int value;
}
/**
* A bounded priority queue for Integers, to retain a specified number of
* the highest-weighted values for return as a random sample.
*/
private static class IntPriorityQueue extends PriorityQueue<MI> {
/**
* Creates a bounded PQ of size <code>size</code>.
* @param size The number of elements to retain.
*/
public IntPriorityQueue(int size) {
super(size);
}
/**
* Returns the underlying data structure for faster access. Extracting elements
* one at a time would require N logN time, and since we want the elements sorted
* in ascending order by value (not weight), the array is useful as-is.
* @return The underlying heap array.
*/
public Object[] getHeap() {
return getHeapArray();
}
/**
* Returns true if <code>o1<code>'s weight is less than that of <code>o2</code>, for
* ordering in the PQ.
* @return True if <code>o1</code> weighs less than <code>o2</code>.
*/
@Override
public boolean lessThan(MI o1, MI o2) {
return o1.value < o2.value;
}
}
/**
* For specifying which sampling algorithm to use.
*/
private enum Algorithm {
/**
* Specifies a methodical traversal algorithm, which is guaranteed to span the collection
* at least once, and never to return duplicates. Faster than the hashing algorithm and
* uses much less space, but the randomness of the sample may be affected by systematic
* variations in the collection. Requires only an array for the sample, and visits only
* the number of elements in the sample set, not the full set.
*/
// TODO (Facet): This one produces a bimodal distribution (very flat around
// each peak!) for collection size 10M and sample sizes 10k and 10544.
// Figure out why.
TRAVERSAL,
/**
* Specifies a Fibonacci-style hash algorithm (see Knuth, S&S), which generates a less
* systematically distributed subset of the sampled collection than the traversal method,
* but requires a bounded priority queue the size of the sample, and creates an object
* containing a sampled value and its hash, for every element in the full set.
*/
HASHING
}
/**
* For specifying whether to sort the sample.
*/
private enum Sorted {
/**
* Sort resulting sample before returning.
*/
YES,
/**
*Do not sort the resulting sample.
*/
NO
}
/**
* Magic number 1: prime closest to phi, in 32 bits.
*/
private static final long PHI_32 = 2654435769L;
/**
* Magic number 2: multiplicative inverse of PHI_32, modulo 2**32.
*/
private static final long PHI_32I = 340573321L;
/**
* Switch to cause methods to return timings.
*/
private static boolean returnTimings = false;
}