2 * Copyright 2013 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 #ifndef FOLLY_HISTOGRAM_H_
18 #define FOLLY_HISTOGRAM_H_
27 #include "folly/detail/Stats.h"
34 * A helper class to manage a set of histogram buckets.
36 template <typename T, typename BucketT>
37 class HistogramBuckets {
40 typedef BucketT BucketType;
43 * Create a set of histogram buckets.
45 * One bucket will be created for each bucketSize interval of values within
46 * the specified range. Additionally, one bucket will be created to track
47 * all values that fall below the specified minimum, and one bucket will be
48 * created for all values above the specified maximum.
50 * If (max - min) is not a multiple of bucketSize, the last bucket will cover
51 * a smaller range of values than the other buckets.
53 * (max - min) must be larger than or equal to bucketSize.
55 HistogramBuckets(ValueType bucketSize, ValueType min, ValueType max,
56 const BucketType& defaultBucket);
58 /* Returns the bucket size of each bucket in the histogram. */
59 ValueType getBucketSize() const {
63 /* Returns the min value at which bucketing begins. */
64 ValueType getMin() const {
68 /* Returns the max value at which bucketing ends. */
69 ValueType getMax() const {
74 * Returns the number of buckets.
76 * This includes the total number of buckets for the [min, max) range,
77 * plus 2 extra buckets, one for handling values less than min, and one for
78 * values greater than max.
80 unsigned int getNumBuckets() const {
81 return buckets_.size();
84 /* Returns the bucket index into which the given value would fall. */
85 unsigned int getBucketIdx(ValueType value) const;
87 /* Returns the bucket for the specified value */
88 BucketType& getByValue(ValueType value) {
89 return buckets_[getBucketIdx(value)];
92 /* Returns the bucket for the specified value */
93 const BucketType& getByValue(ValueType value) const {
94 return buckets_[getBucketIdx(value)];
98 * Returns the bucket at the specified index.
100 * Note that index 0 is the bucket for all values less than the specified
101 * minimum. Index 1 is the first bucket in the specified bucket range.
103 BucketType& getByIndex(unsigned int idx) {
104 return buckets_[idx];
107 /* Returns the bucket at the specified index. */
108 const BucketType& getByIndex(unsigned int idx) const {
109 return buckets_[idx];
113 * Returns the minimum threshold for the bucket at the given index.
115 * The bucket at the specified index will store values in the range
116 * [bucketMin, bucketMin + bucketSize), or [bucketMin, max), if the overall
117 * max is smaller than bucketMin + bucketSize.
119 ValueType getBucketMin(unsigned int idx) const {
121 return std::numeric_limits<ValueType>::min();
123 if (idx == buckets_.size() - 1) {
127 return min_ + ((idx - 1) * bucketSize_);
131 * Returns the maximum threshold for the bucket at the given index.
133 * The bucket at the specified index will store values in the range
134 * [bucketMin, bucketMin + bucketSize), or [bucketMin, max), if the overall
135 * max is smaller than bucketMin + bucketSize.
137 ValueType getBucketMax(unsigned int idx) const {
138 if (idx == buckets_.size() - 1) {
139 return std::numeric_limits<ValueType>::max();
142 return min_ + (idx * bucketSize_);
146 * Determine which bucket the specified percentile falls into.
148 * Looks for the bucket that contains the Nth percentile data point.
150 * @param pct The desired percentile to find, as a value from 0.0 to 1.0.
151 * @param countFn A function that takes a const BucketType&, and returns the
152 * number of values in that bucket.
153 * @param lowPct The lowest percentile stored in the selected bucket will be
154 * returned via this parameter.
155 * @param highPct The highest percentile stored in the selected bucket will
156 * be returned via this parameter.
158 * @return Returns the index of the bucket that contains the Nth percentile
161 template <typename CountFn>
162 unsigned int getPercentileBucketIdx(double pct,
163 CountFn countFromBucket,
164 double* lowPct = NULL,
165 double* highPct = NULL) const;
168 * Estimate the value at the specified percentile.
170 * @param pct The desired percentile to find, as a value from 0.0 to 1.0.
171 * @param countFn A function that takes a const BucketType&, and returns the
172 * number of values in that bucket.
173 * @param avgFn A function that takes a const BucketType&, and returns the
174 * average of all the values in that bucket.
176 * @return Returns an estimate for N, where N is the number where exactly pct
177 * percentage of the data points in the histogram are less than N.
179 template <typename CountFn, typename AvgFn>
180 ValueType getPercentileEstimate(double pct,
181 CountFn countFromBucket,
182 AvgFn avgFromBucket) const;
185 * Iterator access to the buckets.
187 * Note that the first bucket is for all values less than min, and the last
188 * bucket is for all values greater than max. The buckets tracking values in
189 * the [min, max) actually start at the second bucket.
191 typename std::vector<BucketType>::const_iterator begin() const {
192 return buckets_.begin();
194 typename std::vector<BucketType>::iterator begin() {
195 return buckets_.begin();
197 typename std::vector<BucketType>::const_iterator end() const {
198 return buckets_.end();
200 typename std::vector<BucketType>::iterator end() {
201 return buckets_.end();
205 const ValueType bucketSize_;
206 const ValueType min_;
207 const ValueType max_;
208 std::vector<BucketType> buckets_;
215 * A basic histogram class.
217 * Groups data points into equally-sized buckets, and stores the overall sum of
218 * the data points in each bucket, as well as the number of data points in the
221 * The caller must specify the minimum and maximum data points to expect ahead
222 * of time, as well as the bucket width.
224 template <typename T>
228 typedef detail::Bucket<T> Bucket;
230 Histogram(ValueType bucketSize, ValueType min, ValueType max)
231 : buckets_(bucketSize, min, max, Bucket()) {}
233 /* Add a data point to the histogram */
234 void addValue(ValueType value) {
235 Bucket& bucket = buckets_.getByValue(value);
236 // TODO: It would be nice to handle overflow here.
242 * Remove a data point to the histogram
244 * Note that this method does not actually verify that this exact data point
245 * had previously been added to the histogram; it merely subtracts the
246 * requested value from the appropriate bucket's sum.
248 void removeValue(ValueType value) {
249 Bucket& bucket = buckets_.getByValue(value);
250 // TODO: It would be nice to handle overflow here.
255 /* Remove all data points from the histogram */
257 for (int i = 0; i < buckets_.getNumBuckets(); i++) {
258 buckets_.getByIndex(i).clear();
262 /* Merge two histogram data together */
263 void merge(Histogram &hist) {
264 // the two histogram bucket definitions must match to support
266 if (getBucketSize() != hist.getBucketSize() ||
267 getMin() != hist.getMin() ||
268 getMax() != hist.getMax() ||
269 getNumBuckets() != hist.getNumBuckets() ) {
270 throw std::invalid_argument("Cannot merge from input histogram.");
273 for (int i = 0; i < buckets_.getNumBuckets(); i++) {
274 buckets_.getByIndex(i) += hist.buckets_.getByIndex(i);
278 /* Copy bucket values from another histogram */
279 void copy(Histogram &hist) {
280 // the two histogram bucket definition must match
281 if (getBucketSize() != hist.getBucketSize() ||
282 getMin() != hist.getMin() ||
283 getMax() != hist.getMax() ||
284 getNumBuckets() != hist.getNumBuckets() ) {
285 throw std::invalid_argument("Cannot copy from input histogram.");
288 for (int i = 0; i < buckets_.getNumBuckets(); i++) {
289 buckets_.getByIndex(i) = hist.buckets_.getByIndex(i);
293 /* Returns the bucket size of each bucket in the histogram. */
294 ValueType getBucketSize() const {
295 return buckets_.getBucketSize();
297 /* Returns the min value at which bucketing begins. */
298 ValueType getMin() const {
299 return buckets_.getMin();
301 /* Returns the max value at which bucketing ends. */
302 ValueType getMax() const {
303 return buckets_.getMax();
305 /* Returns the number of buckets */
306 unsigned int getNumBuckets() const {
307 return buckets_.getNumBuckets();
310 /* Returns the specified bucket (for reading only!) */
311 const Bucket& getBucketByIndex(int idx) const {
312 return buckets_.getByIndex(idx);
316 * Returns the minimum threshold for the bucket at the given index.
318 * The bucket at the specified index will store values in the range
319 * [bucketMin, bucketMin + bucketSize), or [bucketMin, max), if the overall
320 * max is smaller than bucketMin + bucketSize.
322 ValueType getBucketMin(unsigned int idx) const {
323 return buckets_.getBucketMin(idx);
327 * Returns the maximum threshold for the bucket at the given index.
329 * The bucket at the specified index will store values in the range
330 * [bucketMin, bucketMin + bucketSize), or [bucketMin, max), if the overall
331 * max is smaller than bucketMin + bucketSize.
333 ValueType getBucketMax(unsigned int idx) const {
334 return buckets_.getBucketMax(idx);
338 * Get the bucket that the specified percentile falls into
340 * The lowest and highest percentile data points in returned bucket will be
341 * returned in the lowPct and highPct arguments, if they are non-NULL.
343 unsigned int getPercentileBucketIdx(double pct,
344 double* lowPct = NULL,
345 double* highPct = NULL) const {
346 // We unfortunately can't use lambdas here yet;
347 // Some users of this code are still built with gcc-4.4.
348 CountFromBucket countFn;
349 return buckets_.getPercentileBucketIdx(pct, countFn, lowPct, highPct);
353 * Estimate the value at the specified percentile.
355 * @param pct The desired percentile to find, as a value from 0.0 to 1.0.
357 * @return Returns an estimate for N, where N is the number where exactly pct
358 * percentage of the data points in the histogram are less than N.
360 ValueType getPercentileEstimate(double pct) const {
361 CountFromBucket countFn;
363 return buckets_.getPercentileEstimate(pct, countFn, avgFn);
367 * Get a human-readable string describing the histogram contents
369 std::string debugString() const;
372 * Write the histogram contents in tab-separated values (TSV) format.
373 * Format is "min max count sum".
375 void toTSV(std::ostream& out, bool skipEmptyBuckets = true) const;
378 struct CountFromBucket {
379 uint64_t operator()(const Bucket& bucket) const {
383 struct AvgFromBucket {
384 ValueType operator()(const Bucket& bucket) const {
385 if (bucket.count == 0) {
388 // Cast bucket.count to a signed integer type. This ensures that we
389 // perform division properly here: If bucket.sum is a signed integer
390 // type but we divide by an unsigned number, unsigned division will be
391 // performed and bucket.sum will be converted to unsigned first.
392 // If bucket.sum is unsigned, the code will still do unsigned division
395 // The only downside is if bucket.count is large enough to be negative
396 // when treated as signed. That should be extremely unlikely, though.
397 return bucket.sum / static_cast<int64_t>(bucket.count);
401 detail::HistogramBuckets<ValueType, Bucket> buckets_;
406 #include "folly/Histogram-inl.h"
408 #endif // FOLLY_HISTOGRAM_H_