Dead salmon brain

The dead salmon brain is an open source project for industrial A/B statistical analysis and reporting. The core library extends Apache Spark functionality, making it easy to integrate into your data processing pipelines and providing scalable implementations of common online experimentation techniques and mathematical methods.

API

Java/Scala/Python

Getting Started

Build from source

1) Clone git repository
2) Check spark,scala,hadoop and set correct in 3 step
3) bash setup.sh spark=3.1.2 scala=2.12.11 hadoop=3.2 is_build=true is_install_python_lib=false

Setup from released packages

1) Clone git repository
2) Check version,spark,scala,hadoop and set correct in 3 step
3) bash setup.sh version=0.0.4 spark=3.1.2 scala=2.12.11 hadoop=3.2 is_build=false is_install_python_lib=false

Add as library

Maven Central

PIP

Data model

case class ExpData(
    timestamp: Long, //metric timestamp
    variantId: String, //(treatment and control)
    entityUid: String, //unique entity id
    experimentUid: String, //unique experiment id
    metricValue: Double, //numeric metric value
    metricName: String, //unique metric name in metricSource space
    categoryName: String, //entity category name (i.e gender)
    categoryValue: String, // entity category value (i.e male, female, other)
    metricSource: String, //metric source 
    isAdditive: Boolean   // is additive metric or not
)

Usage

import org.apache.commons.math3.distribution.NormalDistribution
import org.apache.spark.ml.Pipeline
import ai.salmonbrain.computing.{ExpData, CumulativeMetricTransformer, OutlierRemoveTransformer, AutoStatisticsTransformer}


val control = new NormalDistribution(1, 1)
  .sample(1000)
  .zipWithIndex.map { case (value, idx) =>
  ExpData(System.currentTimeMillis(), "control", idx.toString, "exp", value, "timeSpent")
}

val treatment = new NormalDistribution(2, 4)
  .sample(1000)
  .zipWithIndex.map { case (value, idx) =>
  ExpData(System.currentTimeMillis(), "treatment", idx.toString, "exp", value, "timeSpent")
}


val model = new Pipeline().setStages(
  Array(
    new CumulativeMetricTransformer(), // aggregate all metrics
    new OutlierRemoveTransformer() // remove outliers by percentile
           .setLowerPercentile(0.01)
           .setUpperPercentile(0.99),
    new AutoStatisticsTransformer() // auto choose and compute Welch or MannWhitney test
            .setAlpha(0.05)
            .setBeta(0.2)
  )
)

val report = model.fit(data).transform(data)

report.select(
  "metricName",
  "experimentUid",
  "statisticsData.srm",
  "statisticsData.testType",
  "statisticsData.statResult.pValue",
  "statisticsData.statResult.requiredSampleSizeByVariant",
  "statisticsData.statResult.percentageLeft",
  "statisticsData.statResult.percentageRight"
).show
+----------+-------------+-----+--------+-------+---------------------------+-----------------+-----------------+
|metricName|experimentUid|  srm|testType|pValue |requiredSampleSizeByVariant|   percentageLeft|  percentageRight|
+----------+-------------+-----+--------+-------+---------------------------+-----------------+-----------------+
| timeSpent|          exp|false|   WELCH|1.3e-21|                         82|          94.6294|         148.8822|
+----------+-------------+-----+--------+-------+---------------------------+-----------------+-----------------+

What they are and how to interpret them?

srm if true than your test is invalid

pValue if less than 0.05 or 0.01 then great

requiredSampleSizeByVariant estimated required sample size by each variant for observed test and data parameters

percentageLeft lower percent confidence interval

percentageRight upper percent confidence interval

Extra features

import ai.salmon.computing.RatioMetricData

val cum = new CumulativeMetricTransformer()
          .setNumBuckets(256) // you can split your data by buckets and use buckets like new entity
          .setNumeratorNames(Array("clicks"))  // you can set numerator
          .setDenominatorNames(Array("views")) // and denominator
          .setRatioNames(Array("ctr"))  // and create new ratio metric

Contributing

Pull requests are welcome.

License

Apache License 2.0