Skip to content

Distribution Analytics

This tutorial adds descriptive numeric analytics and model-score analytics on top of Getting started. Complete that tutorial first so examples/demo has ingested aggregates.

The demo workspace configures two distribution processors:

  • ih_response_time (numeric_distribution) — t-digest of decision-to-outcome response times.
  • ih_propensity_scores (score_distribution) — t-digests of propensity, final propensity, priority, and rank, split by positive/negative outcome for model-quality curves.

Query Numeric Distributions

95th-percentile response time by channel:

uv run valuestream query examples/demo VS_ResponseTime_P95 --by Channel --grain Day

Query Score Distributions

Median and P90 final propensity:

uv run valuestream query examples/demo VS_FinalPropensity_Median --by Channel --grain Day
uv run valuestream query examples/demo VS_FinalPropensity_P90 --by Channel --grain Day

Median priority:

uv run valuestream query examples/demo VS_Priority_Median --by Channel --grain Day

Query Model-Quality Metrics

ROC AUC reconstructed from the positive/negative propensity t-digests:

uv run valuestream query examples/demo ih_propensity_scores_roc_auc --by Channel --grain Day

How to Read These Numbers

Quantile metrics are computed from mergeable t-digest sketches, not raw rows, so they are approximate with bounded error; ROC AUC, average precision, and calibration are reconstructed from the score digests. Reports mark these values with approximation badges.

  • Algorithms — sketch formulas, error behavior, and curve reconstruction.
  • Processorsnumeric_distribution and score_distribution state layouts.