2 * Copyright 2017 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.
27 #include <folly/CPortability.h>
28 #include <folly/detail/Stats.h>
35 * A helper class to manage a set of histogram buckets.
37 template <typename T, typename BucketT>
38 class HistogramBuckets {
41 typedef BucketT BucketType;
44 * Create a set of histogram buckets.
46 * One bucket will be created for each bucketSize interval of values within
47 * the specified range. Additionally, one bucket will be created to track
48 * all values that fall below the specified minimum, and one bucket will be
49 * created for all values above the specified maximum.
51 * If (max - min) is not a multiple of bucketSize, the last bucket will cover
52 * a smaller range of values than the other buckets.
54 * (max - min) must be larger than or equal to bucketSize.
56 HistogramBuckets(ValueType bucketSize, ValueType min, ValueType max,
57 const BucketType& defaultBucket);
59 /* Returns the bucket size of each bucket in the histogram. */
60 ValueType getBucketSize() const {
64 /* Returns the min value at which bucketing begins. */
65 ValueType getMin() const {
69 /* Returns the max value at which bucketing ends. */
70 ValueType getMax() const {
75 * Returns the number of buckets.
77 * This includes the total number of buckets for the [min, max) range,
78 * plus 2 extra buckets, one for handling values less than min, and one for
79 * values greater than max.
81 size_t getNumBuckets() const {
82 return buckets_.size();
85 /* Returns the bucket index into which the given value would fall. */
86 size_t getBucketIdx(ValueType value) const;
88 /* Returns the bucket for the specified value */
89 BucketType& getByValue(ValueType value) {
90 return buckets_[getBucketIdx(value)];
93 /* Returns the bucket for the specified value */
94 const BucketType& getByValue(ValueType value) const {
95 return buckets_[getBucketIdx(value)];
99 * Returns the bucket at the specified index.
101 * Note that index 0 is the bucket for all values less than the specified
102 * minimum. Index 1 is the first bucket in the specified bucket range.
104 BucketType& getByIndex(size_t idx) {
105 return buckets_[idx];
108 /* Returns the bucket at the specified index. */
109 const BucketType& getByIndex(size_t idx) const {
110 return buckets_[idx];
114 * Returns the minimum threshold for the bucket at the given index.
116 * The bucket at the specified index will store values in the range
117 * [bucketMin, bucketMin + bucketSize), or [bucketMin, max), if the overall
118 * max is smaller than bucketMin + bucketSize.
120 ValueType getBucketMin(size_t idx) const {
122 return std::numeric_limits<ValueType>::min();
124 if (idx == buckets_.size() - 1) {
128 return min_ + ((idx - 1) * bucketSize_);
132 * Returns the maximum threshold for the bucket at the given index.
134 * The bucket at the specified index will store values in the range
135 * [bucketMin, bucketMin + bucketSize), or [bucketMin, max), if the overall
136 * max is smaller than bucketMin + bucketSize.
138 ValueType getBucketMax(size_t idx) const {
139 if (idx == buckets_.size() - 1) {
140 return std::numeric_limits<ValueType>::max();
143 return min_ + (idx * bucketSize_);
147 * Computes the total number of values stored across all buckets.
149 * Runs in O(numBuckets)
151 * @param countFn A function that takes a const BucketType&, and returns the
152 * number of values in that bucket
153 * @return Returns the total number of values stored across all buckets
155 template <typename CountFn>
156 uint64_t computeTotalCount(CountFn countFromBucket) const;
159 * Determine which bucket the specified percentile falls into.
161 * Looks for the bucket that contains the Nth percentile data point.
163 * @param pct The desired percentile to find, as a value from 0.0 to 1.0.
164 * @param countFn A function that takes a const BucketType&, and returns the
165 * number of values in that bucket.
166 * @param lowPct The lowest percentile stored in the selected bucket will be
167 * returned via this parameter.
168 * @param highPct The highest percentile stored in the selected bucket will
169 * be returned via this parameter.
171 * @return Returns the index of the bucket that contains the Nth percentile
174 template <typename CountFn>
175 size_t getPercentileBucketIdx(
177 CountFn countFromBucket,
178 double* lowPct = nullptr,
179 double* highPct = nullptr) const;
182 * Estimate the value at the specified percentile.
184 * @param pct The desired percentile to find, as a value from 0.0 to 1.0.
185 * @param countFn A function that takes a const BucketType&, and returns the
186 * number of values in that bucket.
187 * @param avgFn A function that takes a const BucketType&, and returns the
188 * average of all the values in that bucket.
190 * @return Returns an estimate for N, where N is the number where exactly pct
191 * percentage of the data points in the histogram are less than N.
193 template <typename CountFn, typename AvgFn>
194 ValueType getPercentileEstimate(double pct,
195 CountFn countFromBucket,
196 AvgFn avgFromBucket) const;
199 * Iterator access to the buckets.
201 * Note that the first bucket is for all values less than min, and the last
202 * bucket is for all values greater than max. The buckets tracking values in
203 * the [min, max) actually start at the second bucket.
205 typename std::vector<BucketType>::const_iterator begin() const {
206 return buckets_.begin();
208 typename std::vector<BucketType>::iterator begin() {
209 return buckets_.begin();
211 typename std::vector<BucketType>::const_iterator end() const {
212 return buckets_.end();
214 typename std::vector<BucketType>::iterator end() {
215 return buckets_.end();
219 ValueType bucketSize_;
222 std::vector<BucketType> buckets_;
229 * A basic histogram class.
231 * Groups data points into equally-sized buckets, and stores the overall sum of
232 * the data points in each bucket, as well as the number of data points in the
235 * The caller must specify the minimum and maximum data points to expect ahead
236 * of time, as well as the bucket width.
238 template <typename T>
242 typedef detail::Bucket<T> Bucket;
244 Histogram(ValueType bucketSize, ValueType min, ValueType max)
245 : buckets_(bucketSize, min, max, Bucket()) {}
247 /* Add a data point to the histogram */
248 void addValue(ValueType value) UBSAN_DISABLE("signed-integer-overflow")
249 UBSAN_DISABLE("unsigned-integer-overflow") {
250 Bucket& bucket = buckets_.getByValue(value);
251 // NOTE: Overflow is handled elsewhere and tests check this
252 // behavior (see HistogramTest.cpp TestOverflow* tests).
253 // TODO: It would be nice to handle overflow here and redesign this class.
258 /* Add multiple same data points to the histogram */
259 void addRepeatedValue(ValueType value, uint64_t nSamples)
260 UBSAN_DISABLE("signed-integer-overflow")
261 UBSAN_DISABLE("unsigned-integer-overflow") {
262 Bucket& bucket = buckets_.getByValue(value);
263 // NOTE: Overflow is handled elsewhere and tests check this
264 // behavior (see HistogramTest.cpp TestOverflow* tests).
265 // TODO: It would be nice to handle overflow here and redesign this class.
266 bucket.sum += value * nSamples;
267 bucket.count += nSamples;
271 * Remove a data point to the histogram
273 * Note that this method does not actually verify that this exact data point
274 * had previously been added to the histogram; it merely subtracts the
275 * requested value from the appropriate bucket's sum.
277 void removeValue(ValueType value) UBSAN_DISABLE("signed-integer-overflow")
278 UBSAN_DISABLE("unsigned-integer-overflow") {
279 Bucket& bucket = buckets_.getByValue(value);
280 // NOTE: Overflow is handled elsewhere and tests check this
281 // behavior (see HistogramTest.cpp TestOverflow* tests).
282 // TODO: It would be nice to handle overflow here and redesign this class.
283 if (bucket.count > 0) {
287 bucket.sum = ValueType();
292 /* Remove multiple same data points from the histogram */
293 void removeRepeatedValue(ValueType value, uint64_t nSamples) {
294 Bucket& bucket = buckets_.getByValue(value);
295 // TODO: It would be nice to handle overflow here.
296 if (bucket.count >= nSamples) {
297 bucket.sum -= value * nSamples;
298 bucket.count -= nSamples;
300 bucket.sum = ValueType();
305 /* Remove all data points from the histogram */
307 for (size_t i = 0; i < buckets_.getNumBuckets(); i++) {
308 buckets_.getByIndex(i).clear();
312 /* Subtract another histogram data from the histogram */
313 void subtract(const Histogram &hist) {
314 // the two histogram bucket definitions must match to support
316 if (getBucketSize() != hist.getBucketSize() ||
317 getMin() != hist.getMin() ||
318 getMax() != hist.getMax() ||
319 getNumBuckets() != hist.getNumBuckets() ) {
320 throw std::invalid_argument("Cannot subtract input histogram.");
323 for (size_t i = 0; i < buckets_.getNumBuckets(); i++) {
324 buckets_.getByIndex(i) -= hist.buckets_.getByIndex(i);
328 /* Merge two histogram data together */
329 void merge(const Histogram &hist) {
330 // the two histogram bucket definitions must match to support
332 if (getBucketSize() != hist.getBucketSize() ||
333 getMin() != hist.getMin() ||
334 getMax() != hist.getMax() ||
335 getNumBuckets() != hist.getNumBuckets() ) {
336 throw std::invalid_argument("Cannot merge from input histogram.");
339 for (size_t i = 0; i < buckets_.getNumBuckets(); i++) {
340 buckets_.getByIndex(i) += hist.buckets_.getByIndex(i);
344 /* Copy bucket values from another histogram */
345 void copy(const Histogram &hist) {
346 // the two histogram bucket definition must match
347 if (getBucketSize() != hist.getBucketSize() ||
348 getMin() != hist.getMin() ||
349 getMax() != hist.getMax() ||
350 getNumBuckets() != hist.getNumBuckets() ) {
351 throw std::invalid_argument("Cannot copy from input histogram.");
354 for (size_t i = 0; i < buckets_.getNumBuckets(); i++) {
355 buckets_.getByIndex(i) = hist.buckets_.getByIndex(i);
359 /* Returns the bucket size of each bucket in the histogram. */
360 ValueType getBucketSize() const {
361 return buckets_.getBucketSize();
363 /* Returns the min value at which bucketing begins. */
364 ValueType getMin() const {
365 return buckets_.getMin();
367 /* Returns the max value at which bucketing ends. */
368 ValueType getMax() const {
369 return buckets_.getMax();
371 /* Returns the number of buckets */
372 size_t getNumBuckets() const {
373 return buckets_.getNumBuckets();
376 /* Returns the specified bucket (for reading only!) */
377 const Bucket& getBucketByIndex(size_t idx) const {
378 return buckets_.getByIndex(idx);
382 * Returns the minimum threshold for the bucket at the given index.
384 * The bucket at the specified index will store values in the range
385 * [bucketMin, bucketMin + bucketSize), or [bucketMin, max), if the overall
386 * max is smaller than bucketMin + bucketSize.
388 ValueType getBucketMin(size_t idx) const {
389 return buckets_.getBucketMin(idx);
393 * Returns the maximum threshold for the bucket at the given index.
395 * The bucket at the specified index will store values in the range
396 * [bucketMin, bucketMin + bucketSize), or [bucketMin, max), if the overall
397 * max is smaller than bucketMin + bucketSize.
399 ValueType getBucketMax(size_t idx) const {
400 return buckets_.getBucketMax(idx);
404 * Computes the total number of values stored across all buckets.
406 * Runs in O(numBuckets)
408 uint64_t computeTotalCount() const {
409 CountFromBucket countFn;
410 return buckets_.computeTotalCount(countFn);
414 * Get the bucket that the specified percentile falls into
416 * The lowest and highest percentile data points in returned bucket will be
417 * returned in the lowPct and highPct arguments, if they are non-NULL.
419 size_t getPercentileBucketIdx(
421 double* lowPct = nullptr,
422 double* highPct = nullptr) const {
423 // We unfortunately can't use lambdas here yet;
424 // Some users of this code are still built with gcc-4.4.
425 CountFromBucket countFn;
426 return buckets_.getPercentileBucketIdx(pct, countFn, lowPct, highPct);
430 * Estimate the value at the specified percentile.
432 * @param pct The desired percentile to find, as a value from 0.0 to 1.0.
434 * @return Returns an estimate for N, where N is the number where exactly pct
435 * percentage of the data points in the histogram are less than N.
437 ValueType getPercentileEstimate(double pct) const {
438 CountFromBucket countFn;
440 return buckets_.getPercentileEstimate(pct, countFn, avgFn);
444 * Get a human-readable string describing the histogram contents
446 std::string debugString() const;
449 * Write the histogram contents in tab-separated values (TSV) format.
450 * Format is "min max count sum".
452 void toTSV(std::ostream& out, bool skipEmptyBuckets = true) const;
454 struct CountFromBucket {
455 uint64_t operator()(const Bucket& bucket) const {
459 struct AvgFromBucket {
460 ValueType operator()(const Bucket& bucket) const {
461 if (bucket.count == 0) {
464 // Cast bucket.count to a signed integer type. This ensures that we
465 // perform division properly here: If bucket.sum is a signed integer
466 // type but we divide by an unsigned number, unsigned division will be
467 // performed and bucket.sum will be converted to unsigned first.
468 // If bucket.sum is unsigned, the code will still do unsigned division
471 // The only downside is if bucket.count is large enough to be negative
472 // when treated as signed. That should be extremely unlikely, though.
473 return bucket.sum / static_cast<int64_t>(bucket.count);
478 detail::HistogramBuckets<ValueType, Bucket> buckets_;