StatsGenerator
Related Doc:
package row
MetricTransform represents a single statistic built on top of an input column.
InputTransform acts as a signal of how to process the metric.
InputTransform acts as a signal of how to process the metric.
IsNull: Check if the input is null.
Raw: Operate in the input column.
One: lit(true) in spark. Used for row counts leveraged to obtain null rate values.
PSI is a measure of the difference between two probability distributions.
PSI is a measure of the difference between two probability distributions. However, it's not defined for cases where a bin can have zero elements in either distribution (meant for continuous measures). In order to support PSI for discrete measures we add a small eps value to perturb the distribution in bins.
Existing rules of thumb are: PSI < 0.10 means "little shift", .10<PSI<.25 means "moderate shift", and PSI>0.25 means "significant shift, action required" https://scholarworks.wmich.edu/dissertations/3208
Post processing for finalized values or IRs when generating a time series of stats.
Post processing for finalized values or IRs when generating a time series of stats. In the case of percentiles for examples we reduce to 5 values in order to generate candlesticks.
Stats applied to any column
Build RowAggregator to use for computing stats on a dataframe based on metrics
For the schema of the data define metrics to be aggregated
Stats applied to numeric columns
Given a PMF add and substract small values to keep a valid probability distribution without zeros
Input schema is the data required to update partial aggregations / stats.
Input schema is the data required to update partial aggregations / stats.
Given a valueSchema and a metric transform list, defines the schema expected by the Stats aggregator (online and offline)
A valueSchema (for join) and Metric list define uniquely the IRSchema to be used for the statistics.
A valueSchema (for join) and Metric list define uniquely the IRSchema to be used for the statistics. In order to support custom storage for statistic percentiles this method would need to be modified. IR Schemas are used to decode streaming partial aggregations as well as KvStore partial stats.
Module managing FeatureStats Schema, Aggregations to be used by type and aggregator construction.
Stats Aggregation has an offline/ batch component and an online component. The metrics defined for stats depend on the schema of the join. The dataTypes and column names. For the online side, we obtain this information from the JoinCodec/valueSchema For the offline side, we obtain this information directly from the outputTable. To keep the schemas consistent we sort the metrics in the schema by name. (one column can have multiple metrics).