diff --git a/examples/random/main.go b/examples/random/main.go index 9b910faec..334fd8a47 100644 --- a/examples/random/main.go +++ b/examples/random/main.go @@ -54,10 +54,10 @@ var ( // normal distribution, with 20 buckets centered on the mean, each // half-sigma wide. rpcDurationsHistogram = prometheus.NewHistogram(prometheus.HistogramOpts{ - Name: "rpc_durations_histogram_seconds", - Help: "RPC latency distributions.", - Buckets: prometheus.LinearBuckets(*normMean-5**normDomain, .5**normDomain, 20), - SparseBucketsResolution: 20, + Name: "rpc_durations_histogram_seconds", + Help: "RPC latency distributions.", + Buckets: prometheus.LinearBuckets(*normMean-5**normDomain, .5**normDomain, 20), + SparseBucketsFactor: 1.1, }) ) diff --git a/go.mod b/go.mod index 82fca34f6..813ae76ce 100644 --- a/go.mod +++ b/go.mod @@ -5,7 +5,7 @@ require ( github.com/cespare/xxhash/v2 v2.1.1 github.com/golang/protobuf v1.4.3 github.com/json-iterator/go v1.1.11 - github.com/prometheus/client_model v0.2.1-0.20210403151606-24db95a3d5d6 + github.com/prometheus/client_model v0.2.1-0.20210611125623-bbaf1cc17b15 github.com/prometheus/common v0.26.0 github.com/prometheus/procfs v0.6.0 golang.org/x/sys v0.0.0-20210603081109-ebe580a85c40 diff --git a/go.sum b/go.sum index 9bd013484..fc2a05666 100644 --- a/go.sum +++ b/go.sum @@ -24,6 +24,7 @@ github.com/gogo/protobuf v1.1.1/go.mod h1:r8qH/GZQm5c6nD/R0oafs1akxWv10x8SbQlK7a github.com/golang/protobuf v1.2.0/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U= github.com/golang/protobuf v1.3.1/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U= github.com/golang/protobuf v1.3.2/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U= +github.com/golang/protobuf v1.3.5/go.mod h1:6O5/vntMXwX2lRkT1hjjk0nAC1IDOTvTlVgjlRvqsdk= github.com/golang/protobuf v1.4.0-rc.1/go.mod h1:ceaxUfeHdC40wWswd/P6IGgMaK3YpKi5j83Wpe3EHw8= github.com/golang/protobuf v1.4.0-rc.1.0.20200221234624-67d41d38c208/go.mod h1:xKAWHe0F5eneWXFV3EuXVDTCmh+JuBKY0li0aMyXATA= github.com/golang/protobuf v1.4.0-rc.2/go.mod h1:LlEzMj4AhA7rCAGe4KMBDvJI+AwstrUpVNzEA03Pprs= @@ -77,8 +78,8 @@ github.com/prometheus/client_golang v1.7.1/go.mod h1:PY5Wy2awLA44sXw4AOSfFBetzPP github.com/prometheus/client_model v0.0.0-20180712105110-5c3871d89910/go.mod h1:MbSGuTsp3dbXC40dX6PRTWyKYBIrTGTE9sqQNg2J8bo= github.com/prometheus/client_model v0.0.0-20190129233127-fd36f4220a90/go.mod h1:xMI15A0UPsDsEKsMN9yxemIoYk6Tm2C1GtYGdfGttqA= github.com/prometheus/client_model v0.2.0/go.mod h1:xMI15A0UPsDsEKsMN9yxemIoYk6Tm2C1GtYGdfGttqA= -github.com/prometheus/client_model v0.2.1-0.20210403151606-24db95a3d5d6 h1:wlZYx9ITBsvMO/wVoi30A36fAdRlBC130JksGGfaYl8= -github.com/prometheus/client_model v0.2.1-0.20210403151606-24db95a3d5d6/go.mod h1:xMI15A0UPsDsEKsMN9yxemIoYk6Tm2C1GtYGdfGttqA= +github.com/prometheus/client_model v0.2.1-0.20210611125623-bbaf1cc17b15 h1:l+7cw41KLeOScRk7f9Tg//xT8LAz55Kg+Fg9i0i0Cyw= +github.com/prometheus/client_model v0.2.1-0.20210611125623-bbaf1cc17b15/go.mod h1:LDGWKZIo7rky3hgvBe+caln+Dr3dPggB5dvjtD7w9+w= github.com/prometheus/common v0.4.1/go.mod h1:TNfzLD0ON7rHzMJeJkieUDPYmFC7Snx/y86RQel1bk4= github.com/prometheus/common v0.10.0/go.mod h1:Tlit/dnDKsSWFlCLTWaA1cyBgKHSMdTB80sz/V91rCo= github.com/prometheus/common v0.26.0 h1:iMAkS2TDoNWnKM+Kopnx/8tnEStIfpYA0ur0xQzzhMQ= diff --git a/prometheus/examples_test.go b/prometheus/examples_test.go index f97a3e2ac..bdcdfb4aa 100644 --- a/prometheus/examples_test.go +++ b/prometheus/examples_test.go @@ -538,8 +538,8 @@ func ExampleHistogram() { // cumulative_count: 816 // upper_bound: 40 // > - // sb_resolution: 0 - // sb_zero_threshold: 1e-128 + // sb_schema: 0 + // sb_zero_threshold: 0 // > } diff --git a/prometheus/histogram.go b/prometheus/histogram.go index 1c1112861..a0e4b4e13 100644 --- a/prometheus/histogram.go +++ b/prometheus/histogram.go @@ -28,6 +28,176 @@ import ( dto "github.com/prometheus/client_model/go" ) +// sparseBounds for the frac of observed values. Only relevant for schema > 0. +// Position in the slice is the schema. (0 is never used, just here for +// convenience of using the schema directly as the index.) +var sparseBounds = [][]float64{ + // Schema "0": + []float64{0.5}, + // Schema 1: + []float64{0.5, 0.7071067811865475}, + // Schema 2: + []float64{0.5, 0.5946035575013605, 0.7071067811865475, 0.8408964152537144}, + // Schema 3: + []float64{0.5, 0.5452538663326288, 0.5946035575013605, 0.6484197773255048, + 0.7071067811865475, 0.7711054127039704, 0.8408964152537144, 0.9170040432046711}, + // Schema 4: + []float64{0.5, 0.5221368912137069, 0.5452538663326288, 0.5693943173783458, + 0.5946035575013605, 0.620928906036742, 0.6484197773255048, 0.6771277734684463, + 0.7071067811865475, 0.7384130729697496, 0.7711054127039704, 0.805245165974627, + 0.8408964152537144, 0.8781260801866495, 0.9170040432046711, 0.9576032806985735}, + // Schema 5: + []float64{0.5, 0.5109485743270583, 0.5221368912137069, 0.5335702003384117, + 0.5452538663326288, 0.5571933712979462, 0.5693943173783458, 0.5818624293887887, + 0.5946035575013605, 0.6076236799902344, 0.620928906036742, 0.6345254785958666, + 0.6484197773255048, 0.6626183215798706, 0.6771277734684463, 0.6919549409819159, + 0.7071067811865475, 0.7225904034885232, 0.7384130729697496, 0.7545822137967112, + 0.7711054127039704, 0.7879904225539431, 0.805245165974627, 0.8228777390769823, + 0.8408964152537144, 0.8593096490612387, 0.8781260801866495, 0.8973545375015533, + 0.9170040432046711, 0.9370838170551498, 0.9576032806985735, 0.9785720620876999}, + // Schema 6: + []float64{0.5, 0.5054446430258502, 0.5109485743270583, 0.5165124395106142, + 0.5221368912137069, 0.5278225891802786, 0.5335702003384117, 0.5393803988785598, + 0.5452538663326288, 0.5511912916539204, 0.5571933712979462, 0.5632608093041209, + 0.5693943173783458, 0.5755946149764913, 0.5818624293887887, 0.5881984958251406, + 0.5946035575013605, 0.6010783657263515, 0.6076236799902344, 0.6142402680534349, + 0.620928906036742, 0.6276903785123455, 0.6345254785958666, 0.6414350080393891, + 0.6484197773255048, 0.6554806057623822, 0.6626183215798706, 0.6698337620266515, + 0.6771277734684463, 0.6845012114872953, 0.6919549409819159, 0.6994898362691555, + 0.7071067811865475, 0.7148066691959849, 0.7225904034885232, 0.7304588970903234, + 0.7384130729697496, 0.7464538641456323, 0.7545822137967112, 0.762799075372269, + 0.7711054127039704, 0.7795022001189185, 0.7879904225539431, 0.7965710756711334, + 0.805245165974627, 0.8140137109286738, 0.8228777390769823, 0.8318382901633681, + 0.8408964152537144, 0.8500531768592616, 0.8593096490612387, 0.8686669176368529, + 0.8781260801866495, 0.8876882462632604, 0.8973545375015533, 0.9071260877501991, + 0.9170040432046711, 0.9269895625416926, 0.9370838170551498, 0.9472879907934827, + 0.9576032806985735, 0.9680308967461471, 0.9785720620876999, 0.9892280131939752}, + // Schema 7: + []float64{0.5, 0.5027149505564014, 0.5054446430258502, 0.5081891574554764, + 0.5109485743270583, 0.5137229745593818, 0.5165124395106142, 0.5193170509806894, + 0.5221368912137069, 0.5249720429003435, 0.5278225891802786, 0.5306886136446309, + 0.5335702003384117, 0.5364674337629877, 0.5393803988785598, 0.5423091811066545, + 0.5452538663326288, 0.5482145409081883, 0.5511912916539204, 0.5541842058618393, + 0.5571933712979462, 0.5602188762048033, 0.5632608093041209, 0.5663192597993595, + 0.5693943173783458, 0.572486072215902, 0.5755946149764913, 0.5787200368168754, + 0.5818624293887887, 0.585021884841625, 0.5881984958251406, 0.5913923554921704, + 0.5946035575013605, 0.5978321960199137, 0.6010783657263515, 0.6043421618132907, + 0.6076236799902344, 0.6109230164863786, 0.6142402680534349, 0.6175755319684665, + 0.620928906036742, 0.6243004885946023, 0.6276903785123455, 0.6310986751971253, + 0.6345254785958666, 0.637970889198196, 0.6414350080393891, 0.6449179367033329, + 0.6484197773255048, 0.6519406325959679, 0.6554806057623822, 0.659039800633032, + 0.6626183215798706, 0.6662162735415805, 0.6698337620266515, 0.6734708931164728, + 0.6771277734684463, 0.6808045103191123, 0.6845012114872953, 0.688217985377265, + 0.6919549409819159, 0.6957121878859629, 0.6994898362691555, 0.7032879969095076, + 0.7071067811865475, 0.7109463010845827, 0.7148066691959849, 0.718687998724491, + 0.7225904034885232, 0.7265139979245261, 0.7304588970903234, 0.7344252166684908, + 0.7384130729697496, 0.7424225829363761, 0.7464538641456323, 0.7505070348132126, + 0.7545822137967112, 0.7586795205991071, 0.762799075372269, 0.7669409989204777, + 0.7711054127039704, 0.7752924388424999, 0.7795022001189185, 0.7837348199827764, + 0.7879904225539431, 0.7922691326262467, 0.7965710756711334, 0.8008963778413465, + 0.805245165974627, 0.8096175675974316, 0.8140137109286738, 0.8184337248834821, + 0.8228777390769823, 0.8273458838280969, 0.8318382901633681, 0.8363550898207981, + 0.8408964152537144, 0.8454623996346523, 0.8500531768592616, 0.8546688815502312, + 0.8593096490612387, 0.8639756154809185, 0.8686669176368529, 0.8733836930995842, + 0.8781260801866495, 0.8828942179666361, 0.8876882462632604, 0.8925083056594671, + 0.8973545375015533, 0.9022270839033115, 0.9071260877501991, 0.9120516927035263, + 0.9170040432046711, 0.9219832844793128, 0.9269895625416926, 0.9320230241988943, + 0.9370838170551498, 0.9421720895161669, 0.9472879907934827, 0.9524316709088368, + 0.9576032806985735, 0.9628029718180622, 0.9680308967461471, 0.9732872087896164, + 0.9785720620876999, 0.9838856116165875, 0.9892280131939752, 0.9945994234836328}, + // Schema 8: + []float64{0.5, 0.5013556375251013, 0.5027149505564014, 0.5040779490592088, + 0.5054446430258502, 0.5068150424757447, 0.5081891574554764, 0.509566998038869, + 0.5109485743270583, 0.5123338964485679, 0.5137229745593818, 0.5151158188430205, + 0.5165124395106142, 0.5179128468009786, 0.5193170509806894, 0.520725062344158, + 0.5221368912137069, 0.5235525479396449, 0.5249720429003435, 0.526395386502313, + 0.5278225891802786, 0.5292536613972564, 0.5306886136446309, 0.5321274564422321, + 0.5335702003384117, 0.5350168559101208, 0.5364674337629877, 0.5379219445313954, + 0.5393803988785598, 0.5408428074966075, 0.5423091811066545, 0.5437795304588847, + 0.5452538663326288, 0.5467321995364429, 0.5482145409081883, 0.549700901315111, + 0.5511912916539204, 0.5526857228508706, 0.5541842058618393, 0.5556867516724088, + 0.5571933712979462, 0.5587040757836845, 0.5602188762048033, 0.5617377836665098, + 0.5632608093041209, 0.564787964283144, 0.5663192597993595, 0.5678547070789026, + 0.5693943173783458, 0.5709381019847808, 0.572486072215902, 0.5740382394200894, + 0.5755946149764913, 0.5771552102951081, 0.5787200368168754, 0.5802891060137493, + 0.5818624293887887, 0.5834400184762408, 0.585021884841625, 0.5866080400818185, + 0.5881984958251406, 0.5897932637314379, 0.5913923554921704, 0.5929957828304968, + 0.5946035575013605, 0.5962156912915756, 0.5978321960199137, 0.5994530835371903, + 0.6010783657263515, 0.6027080545025619, 0.6043421618132907, 0.6059806996384005, + 0.6076236799902344, 0.6092711149137041, 0.6109230164863786, 0.6125793968185725, + 0.6142402680534349, 0.6159056423670379, 0.6175755319684665, 0.6192499490999082, + 0.620928906036742, 0.622612415087629, 0.6243004885946023, 0.6259931389331581, + 0.6276903785123455, 0.6293922197748583, 0.6310986751971253, 0.6328097572894031, + 0.6345254785958666, 0.6362458516947014, 0.637970889198196, 0.6397006037528346, + 0.6414350080393891, 0.6431741147730128, 0.6449179367033329, 0.6466664866145447, + 0.6484197773255048, 0.6501778216898253, 0.6519406325959679, 0.6537082229673385, + 0.6554806057623822, 0.6572577939746774, 0.659039800633032, 0.6608266388015788, + 0.6626183215798706, 0.6644148621029772, 0.6662162735415805, 0.6680225691020727, + 0.6698337620266515, 0.6716498655934177, 0.6734708931164728, 0.6752968579460171, + 0.6771277734684463, 0.6789636531064505, 0.6808045103191123, 0.6826503586020058, + 0.6845012114872953, 0.6863570825438342, 0.688217985377265, 0.690083933630119, + 0.6919549409819159, 0.6938310211492645, 0.6957121878859629, 0.6975984549830999, + 0.6994898362691555, 0.7013863456101023, 0.7032879969095076, 0.7051948041086352, + 0.7071067811865475, 0.7090239421602076, 0.7109463010845827, 0.7128738720527471, + 0.7148066691959849, 0.7167447066838943, 0.718687998724491, 0.7206365595643126, + 0.7225904034885232, 0.7245495448210174, 0.7265139979245261, 0.7284837772007218, + 0.7304588970903234, 0.7324393720732029, 0.7344252166684908, 0.7364164454346837, + 0.7384130729697496, 0.7404151139112358, 0.7424225829363761, 0.7444354947621984, + 0.7464538641456323, 0.7484777058836176, 0.7505070348132126, 0.7525418658117031, + 0.7545822137967112, 0.7566280937263048, 0.7586795205991071, 0.7607365094544071, + 0.762799075372269, 0.7648672334736434, 0.7669409989204777, 0.7690203869158282, + 0.7711054127039704, 0.7731960915705107, 0.7752924388424999, 0.7773944698885442, + 0.7795022001189185, 0.7816156449856788, 0.7837348199827764, 0.7858597406461707, + 0.7879904225539431, 0.7901268813264122, 0.7922691326262467, 0.7944171921585818, + 0.7965710756711334, 0.7987307989543135, 0.8008963778413465, 0.8030678282083853, + 0.805245165974627, 0.8074284071024302, 0.8096175675974316, 0.8118126635086642, + 0.8140137109286738, 0.8162207259936375, 0.8184337248834821, 0.820652723822003, + 0.8228777390769823, 0.8251087869603088, 0.8273458838280969, 0.8295890460808079, + 0.8318382901633681, 0.8340936325652911, 0.8363550898207981, 0.8386226785089391, + 0.8408964152537144, 0.8431763167241966, 0.8454623996346523, 0.8477546807446661, + 0.8500531768592616, 0.8523579048290255, 0.8546688815502312, 0.8569861239649629, + 0.8593096490612387, 0.8616394738731368, 0.8639756154809185, 0.8663180910111553, + 0.8686669176368529, 0.871022112577578, 0.8733836930995842, 0.8757516765159389, + 0.8781260801866495, 0.8805069215187917, 0.8828942179666361, 0.8852879870317771, + 0.8876882462632604, 0.890095013257712, 0.8925083056594671, 0.8949281411607002, + 0.8973545375015533, 0.8997875124702672, 0.9022270839033115, 0.9046732696855155, + 0.9071260877501991, 0.909585556079304, 0.9120516927035263, 0.9145245157024483, + 0.9170040432046711, 0.9194902933879467, 0.9219832844793128, 0.9244830347552253, + 0.9269895625416926, 0.92950288621441, 0.9320230241988943, 0.9345499949706191, + 0.9370838170551498, 0.93962450902828, 0.9421720895161669, 0.9447265771954693, + 0.9472879907934827, 0.9498563490882775, 0.9524316709088368, 0.9550139751351947, + 0.9576032806985735, 0.9601996065815236, 0.9628029718180622, 0.9654133954938133, + 0.9680308967461471, 0.9706554947643201, 0.9732872087896164, 0.9759260581154889, + 0.9785720620876999, 0.9812252401044634, 0.9838856116165875, 0.9865531961276168, + 0.9892280131939752, 0.9919100824251095, 0.9945994234836328, 0.9972960560854698}, +} + +// The sparseBounds above can be generated with the code below. +// TODO(beorn7): Actually do it via go generate. +// +// var sparseBounds [][]float64 = make([][]float64, 9) +// +// func init() { +// // Populate sparseBounds. +// numBuckets := 1 +// for i := range sparseBounds { +// bounds := []float64{0.5} +// factor := math.Exp2(math.Exp2(float64(-i))) +// for j := 0; j < numBuckets-1; j++ { +// var bound float64 +// if (j+1)%2 == 0 { +// // Use previously calculated value for increased precision. +// bound = sparseBounds[i-1][j/2+1] +// } else { +// bound = bounds[j] * factor +// } +// bounds = append(bounds, bound) +// } +// numBuckets *= 2 +// sparseBounds[i] = bounds +// } +// } + // A Histogram counts individual observations from an event or sample stream in // configurable buckets. Similar to a summary, it also provides a sum of // observations and an observation count. @@ -68,7 +238,10 @@ var DefBuckets = []float64{.005, .01, .025, .05, .1, .25, .5, 1, 2.5, 5, 10} // DefSparseBucketsZeroThreshold is the default value for // SparseBucketsZeroThreshold in the HistogramOpts. -var DefSparseBucketsZeroThreshold = 1e-128 +const DefSparseBucketsZeroThreshold = 2.938735877055719e-39 + +// This is 2^-128 (or 0.5*2^-127 in the actual IEEE 754 representation), which +// is a bucket boundary at all possible resolutions. var errBucketLabelNotAllowed = fmt.Errorf( "%q is not allowed as label name in histograms", bucketLabel, @@ -162,24 +335,41 @@ type HistogramOpts struct { // buckets here explicitly.) Buckets []float64 - // If SparseBucketsResolution is not zero, sparse buckets are used (in - // addition to the regular buckets, if defined above). Every power of - // ten is divided into the given number of exponential buckets. For - // example, if set to 3, the bucket boundaries are approximately […, - // 0.1, 0.215, 0.464, 1, 2.15, 4,64, 10, 21.5, 46.4, 100, …] Histograms - // can only be properly aggregated if they use the same - // resolution. Therefore, it is recommended to use 20 as a resolution, - // which is generally expected to be a good tradeoff between resource - // usage and accuracy (resulting in a maximum error of quantile values - // of about 6%). - SparseBucketsResolution uint8 + // If SparseBucketsFactor is greater than one, sparse buckets are used + // (in addition to the regular buckets, if defined above). Sparse + // buckets are exponential buckets covering the whole float64 range + // (with the exception of the “zero” bucket, see + // SparseBucketsZeroThreshold below). From any one bucket to the next, + // the width of the bucket grows by a constant factor. + // SparseBucketsFactor provides an upper bound for this factor + // (exception see below). The smaller SparseBucketsFactor, the more + // buckets will be used and thus the more costly the histogram will + // become. A generally good trade-off between cost and accuracy is a + // value of 1.1 (each bucket is at most 10% wider than the previous + // one), which will result in each power of two divided into 8 buckets + // (e.g. there will be 8 buckets between 1 and 2, same as between 2 and + // 4, and 4 and 8, etc.). + // + // Details about the actually used factor: The factor is calculated as + // 2^(2^n), where n is an integer number between (and including) -8 and + // 4. n is chosen so that the resulting factor is the largest that is + // still smaller or equal to SparseBucketsFactor. Note that the smallest + // possible factor is therefore approx. 1.00271 (i.e. 2^(2^-8) ). If + // SparseBucketsFactor is greater than 1 but smaller than 2^(2^-8), then + // the actually used factor is still 2^(2^-8) even though it is larger + // than the provided SparseBucketsFactor. + SparseBucketsFactor float64 // All observations with an absolute value of less or equal // SparseBucketsZeroThreshold are accumulated into a “zero” bucket. For // best results, this should be close to a bucket boundary. This is - // most easily accomplished by picking a power of ten. If + // usually the case if picking a power of two. If // SparseBucketsZeroThreshold is left at zero (or set to a negative // value), DefSparseBucketsZeroThreshold is used as the threshold. SparseBucketsZeroThreshold float64 + // TODO(beorn7): Need a setting to limit total bucket count and to + // configure a strategy to enforce the limit, e.g. if minimum duration + // after last reset, reset. If not, half the resolution and/or expand + // the zero bucket. } // NewHistogram creates a new Histogram based on the provided HistogramOpts. It @@ -217,20 +407,24 @@ func newHistogram(desc *Desc, opts HistogramOpts, labelValues ...string) Histogr } h := &histogram{ - desc: desc, - upperBounds: opts.Buckets, - sparseResolution: uint32(opts.SparseBucketsResolution), - sparseThreshold: opts.SparseBucketsZeroThreshold, - labelPairs: MakeLabelPairs(desc, labelValues), - counts: [2]*histogramCounts{{}, {}}, - now: time.Now, - } - if len(h.upperBounds) == 0 && opts.SparseBucketsResolution == 0 { + desc: desc, + upperBounds: opts.Buckets, + sparseThreshold: opts.SparseBucketsZeroThreshold, + labelPairs: MakeLabelPairs(desc, labelValues), + counts: [2]*histogramCounts{{}, {}}, + now: time.Now, + } + if len(h.upperBounds) == 0 && opts.SparseBucketsFactor <= 1 { h.upperBounds = DefBuckets } if h.sparseThreshold <= 0 { h.sparseThreshold = DefSparseBucketsZeroThreshold } + if opts.SparseBucketsFactor <= 1 { + h.sparseThreshold = 0 // To mark that there are no sparse buckets. + } else { + h.sparseSchema = pickSparseSchema(opts.SparseBucketsFactor) + } for i, upperBound := range h.upperBounds { if i < len(h.upperBounds)-1 { if upperBound >= h.upperBounds[i+1] { @@ -264,14 +458,14 @@ type histogramCounts struct { sumBits uint64 count uint64 buckets []uint64 - // sparse buckets are implemented with a sync.Map for this PoC. A - // dedicated data structure will likely be more efficient. - // There are separate maps for negative and positive observations. - // The map's value is a *uint64, counting observations in that bucket. - // The map's key is the logarithmic index of the bucket. Index 0 is for an - // upper bound of 1. Each increment/decrement by SparseBucketsResolution - // multiplies/divides the upper bound by 10. Indices in between are - // spaced exponentially as defined in spareBounds. + // sparse buckets are implemented with a sync.Map for now. A dedicated + // data structure will likely be more efficient. There are separate maps + // for negative and positive observations. The map's value is an *int64, + // counting observations in that bucket. (Note that we don't use uint64 + // as an int64 won't overflow in practice, and working with signed + // numbers from the beginning simplifies the handling of deltas.) The + // map's key is the index of the bucket according to the used + // sparseSchema. Index 0 is for an upper bound of 1. sparseBucketsPositive, sparseBucketsNegative sync.Map // sparseZeroBucket counts all (positive and negative) observations in // the zero bucket (with an absolute value less or equal @@ -312,10 +506,10 @@ func (hc *histogramCounts) observe(v float64, bucket int, doSparse bool, whichSp atomic.AddUint64(&hc.count, 1) } -func addToSparseBucket(buckets *sync.Map, key int, increment uint64) { +func addToSparseBucket(buckets *sync.Map, key int, increment int64) { if existingBucket, ok := buckets.Load(key); ok { // Fast path without allocation. - atomic.AddUint64(existingBucket.(*uint64), increment) + atomic.AddInt64(existingBucket.(*int64), increment) return } // Bucket doesn't exist yet. Slow path allocating new counter. @@ -323,7 +517,7 @@ func addToSparseBucket(buckets *sync.Map, key int, increment uint64) { if actualBucket, loaded := buckets.LoadOrStore(key, &newBucket); loaded { // The bucket was created concurrently in another goroutine. // Have to increment after all. - atomic.AddUint64(actualBucket.(*uint64), increment) + atomic.AddInt64(actualBucket.(*int64), increment) } } @@ -339,7 +533,7 @@ type histogram struct { // perspective of the histogram) swap the hot–cold under the writeMtx // lock. A cooldown is awaited (while locked) by comparing the number of // observations with the initiation count. Once they match, then the - // last observation on the now cool one has completed. All cool fields must + // last observation on the now cool one has completed. All cold fields must // be merged into the new hot before releasing writeMtx. // // Fields with atomic access first! See alignment constraint: @@ -356,11 +550,11 @@ type histogram struct { // http://golang.org/pkg/sync/atomic/#pkg-note-BUG. counts [2]*histogramCounts - upperBounds []float64 - labelPairs []*dto.LabelPair - exemplars []atomic.Value // One more than buckets (to include +Inf), each a *dto.Exemplar. - sparseResolution uint32 // Instead of uint8 to be ready for protobuf encoding. - sparseThreshold float64 + upperBounds []float64 + labelPairs []*dto.LabelPair + exemplars []atomic.Value // One more than buckets (to include +Inf), each a *dto.Exemplar. + sparseSchema int32 + sparseThreshold float64 // This is zero iff no sparse buckets are used. now func() time.Time // To mock out time.Now() for testing. } @@ -407,7 +601,7 @@ func (h *histogram) Write(out *dto.Metric) error { Bucket: make([]*dto.Bucket, len(h.upperBounds)), SampleCount: proto.Uint64(count), SampleSum: proto.Float64(math.Float64frombits(atomic.LoadUint64(&coldCounts.sumBits))), - SbResolution: &h.sparseResolution, + SbSchema: &h.sparseSchema, SbZeroThreshold: &h.sparseThreshold, } out.Histogram = his @@ -448,7 +642,7 @@ func (h *histogram) Write(out *dto.Metric) error { atomic.AddUint64(&hotCounts.buckets[i], atomic.LoadUint64(&coldCounts.buckets[i])) atomic.StoreUint64(&coldCounts.buckets[i], 0) } - if h.sparseResolution != 0 { + if h.sparseThreshold != 0 { zeroBucket := atomic.LoadUint64(&coldCounts.sparseZeroBucket) defer func() { @@ -478,21 +672,41 @@ func makeSparseBuckets(buckets *sync.Map) *dto.SparseBuckets { } sbs := dto.SparseBuckets{} - var prevCount uint64 + var prevCount int64 var nextI int + + appendDelta := func(count int64) { + *sbs.Span[len(sbs.Span)-1].Length++ + sbs.Delta = append(sbs.Delta, count-prevCount) + prevCount = count + } + for n, i := range ii { v, _ := buckets.Load(i) - count := atomic.LoadUint64(v.(*uint64)) - if n == 0 || i-nextI != 0 { + count := atomic.LoadInt64(v.(*int64)) + // Multiple spans with only small gaps in between are probably + // encoded more efficiently as one larger span with a few empty + // buckets. Needs some research to find the sweet spot. For now, + // we assume that gaps of one ore two buckets should not create + // a new span. + iDelta := int32(i - nextI) + if n == 0 || iDelta > 2 { + // We have to create a new span, either because we are + // at the very beginning, or because we have found a gap + // of more than two buckets. sbs.Span = append(sbs.Span, &dto.SparseBuckets_Span{ - Offset: proto.Int32(int32(i - nextI)), - Length: proto.Uint32(1), + Offset: proto.Int32(iDelta), + Length: proto.Uint32(0), }) } else { - *sbs.Span[len(sbs.Span)-1].Length++ + // We have found a small gap (or no gap at all). + // Insert empty buckets as needed. + for j := int32(0); j < iDelta; j++ { + appendDelta(0) + } } - sbs.Delta = append(sbs.Delta, int64(count)-int64(prevCount)) // TODO(beorn7): Do proper overflow handling. - nextI, prevCount = i+1, count + appendDelta(count) + nextI = i + 1 } return &sbs } @@ -504,9 +718,9 @@ func makeSparseBuckets(buckets *sync.Map) *dto.SparseBuckets { // recreated on the next scrape). func addAndReset(hotBuckets *sync.Map) func(k, v interface{}) bool { return func(k, v interface{}) bool { - bucket := v.(*uint64) - addToSparseBucket(hotBuckets, k.(int), atomic.LoadUint64(bucket)) - atomic.StoreUint64(bucket, 0) + bucket := v.(*int64) + addToSparseBucket(hotBuckets, k.(int), atomic.LoadInt64(bucket)) + atomic.StoreInt64(bucket, 0) return true } } @@ -528,7 +742,8 @@ func (h *histogram) findBucket(v float64) int { // observe is the implementation for Observe without the findBucket part. func (h *histogram) observe(v float64, bucket int) { - doSparse := h.sparseResolution != 0 + // Do not add to sparse buckets for NaN observations. + doSparse := h.sparseThreshold != 0 && !math.IsNaN(v) var whichSparse, sparseKey int if doSparse { switch { @@ -537,13 +752,20 @@ func (h *histogram) observe(v float64, bucket int) { case v < -h.sparseThreshold: whichSparse = -1 } - // TODO(beorn7): This sometimes gives inaccurate results for - // floats that are actual powers of 10, e.g. math.Log10(0.1) is - // calculated as -0.9999999999999999 rather than -1 and thus - // yields a key unexpectedly one off. Maybe special-case precise - // powers of 10. - // TODO(beorn7): This needs special-casing for ±Inf and NaN. - sparseKey = int(math.Ceil(math.Log10(math.Abs(v)) * float64(h.sparseResolution))) + frac, exp := math.Frexp(math.Abs(v)) + switch { + case math.IsInf(v, 0): + sparseKey = math.MaxInt32 // Largest possible sparseKey. + case h.sparseSchema > 0: + bounds := sparseBounds[h.sparseSchema] + sparseKey = sort.SearchFloat64s(bounds, frac) + (exp-1)*len(bounds) + default: + sparseKey = exp + if frac == 0.5 { + sparseKey-- + } + sparseKey /= 1 << -h.sparseSchema + } } // We increment h.countAndHotIdx so that the counter in the lower // 63 bits gets incremented. At the same time, we get the new value @@ -797,3 +1019,24 @@ func (s buckSort) Swap(i, j int) { func (s buckSort) Less(i, j int) bool { return s[i].GetUpperBound() < s[j].GetUpperBound() } + +// pickSparseschema returns the largest number n between -4 and 8 such that +// 2^(2^-n) is less or equal the provided bucketFactor. +// +// Special cases: +// - bucketFactor <= 1: panics. +// - bucketFactor < 2^(2^-8) (but > 1): still returns 8. +func pickSparseSchema(bucketFactor float64) int32 { + if bucketFactor <= 1 { + panic(fmt.Errorf("bucketFactor %f is <=1", bucketFactor)) + } + floor := math.Floor(math.Log2(math.Log2(bucketFactor))) + switch { + case floor <= -8: + return 8 + case floor >= 4: + return -4 + default: + return -int32(floor) + } +}