game-monetization-platform

End-to-end F2P monetization analytics on real GA4 data — BigQuery, dbt, Dagster, Soda, Looker Studio.

A production-shaped data platform built around a working free-to-play puzzle game (Google's public GA4 Flood-It! sample) with calibrated synthetic ad-mediation and IAP layers filling the schema gaps that real public datasets don't cover. The goal is to demonstrate the full monetization analyst toolkit — canonical metric definitions, dbt models, data quality at three layers, scheduled orchestration, BI dashboards, product analytics — against a realistic data shape, with every architectural choice captured in a versioned ADR. Every piece of infrastructure is the same one a real game studio would use; the data is real where public, transparently synthetic where it isn't (ADR-0013).

Live dashboard

Flood-It! Monetization Dashboard →

ARPDAU, D0–D30 retention curve with industry benchmark lines, paying conversion rate, cohort revenue, payer-segment breakdown, total revenue. Six tiles, all reading from the dbt marts.

Architecture

flowchart LR
  subgraph SRC["Sources"]
    FI["GA4 Flood-It<br/>(real, public BQ)"]
    GEN_AD["generate_synthetic_ad_events.py"]
    GEN_IAP["generate_synthetic_iap_events.py"]
  end

  subgraph BQ["BigQuery: monetization-warehouse"]
    direction TB
    subgraph RAW["raw"]
      RAW_AD["synthetic_ad_events"]
      RAW_IAP["synthetic_iap_events"]
    end
    subgraph STG["analytics: staging (dbt views)"]
      STG_FI["stg_floodit__events"]
      STG_AD["stg_synthetic_ads__events"]
      STG_IAP["stg_synthetic_iap__events"]
    end
    subgraph MARTS["analytics: marts (dbt tables)"]
      FRD["fct_revenue_daily"]
      DIM["dim_players"]
      FRC["fct_retention_cohorts"]
    end
  end

  subgraph CONS["Consumers"]
    LS["Looker Studio<br/>(dashboard)"]
    AMP["Amplitude<br/>(funnel/retention)"]
  end

  subgraph OPS["Orchestration & quality"]
    DAGSTER["Dagster<br/>(Components, dg-managed)"]
    SODA["Soda Core<br/>(30 checks)"]
    SENTRY["Sentry<br/>(errors)"]
  end

  FI -. read in place .-> STG_FI
  GEN_AD --> RAW_AD
  GEN_IAP --> RAW_IAP
  RAW_AD --> STG_AD
  RAW_IAP --> STG_IAP
  STG_FI --> FRD
  STG_FI --> FRC
  STG_AD --> FRD
  STG_IAP --> FRD
  STG_IAP --> DIM
  FRD --> DIM
  FRD --> LS
  DIM --> LS
  FRC --> LS
  FRD --> AMP
  STG_IAP --> AMP

  DAGSTER -. schedules .-> SODA
  SODA -. checks .-> RAW
  SODA -. checks .-> MARTS
  SENTRY -. monitors .-> DAGSTER

Loading

Real Flood-It events are read in place from firebase-public-project (no copy, no storage cost). Synthetic generators write Parquet to raw.synthetic_*; dbt views stage everything; three mart tables (fct_revenue_daily, dim_players, fct_retention_cohorts) serve the dashboard and Amplitude. Dagster schedules Soda's 30 quality checks daily. Three glossary metrics' source of truth is dim_players; two come from fct_revenue_daily; one from fct_retention_cohorts.

Stack

Layer	Tool
Warehouse	Google BigQuery (free tier)
Transformation	dbt-bigquery (staging / intermediate / marts per ADR-0012)
Orchestration	Dagster (Components-based, dg-managed)
Data quality	dbt tests + Soda Core + Dagster sensors (three-layer split per ADR-0005)
BI	Looker Studio
Product analytics	Amplitude
Observability	Sentry

Analyses

Three writeups demonstrating analytical thinking against the warehouse:

1. Retention curve health and D7 improvement priorities — leaky front door, sticky back door
2. Revenue concentration and live-ops priorities — top 1% of users drive 86% of revenue
3. Starter pack price-test design — sample-size math at 0.40% baseline conversion

Documentation

• Live dbt docs site → — generated lineage graph with column-level documentation, source freshness, and test coverage. Static HTML at docs/dbt-docs/index.html. Regenerate with cd monetization_warehouse && dbt docs generate --static && cp target/static_index.html ../docs/dbt-docs/index.html.
• docs/glossary.md — canonical metric definitions (ARPDAU, ARPPU, LTV, D1/D7/D30 retention, conversion rate, whale concentration), each pointing at its source-of-truth dbt mart
• docs/adr/ — 13 architecture decision records covering every constraining choice
• docs/SETUP.md — the executable build runbook
• docs/analyses/ — three analytical writeups against the warehouse
• docs/CAREER.md — parallel job-search track
• docs/WORKFLOW.md — human/agent collaboration loop

Status

End-to-end pipeline live: BigQuery datasets, three staging models, three marts, 30 Soda checks scheduled in Dagster (cron 0 6 * * *, verified passing end-to-end as a dagster asset materialize run), Looker Studio dashboard published, 500 events ingested into Amplitude for funnel/retention demos. Six canonical metrics encoded with glossary entries.

Data quality: every refresh of the marts is gated by three layers (ADR-0005) — 12 dbt tests on structural invariants, 30 Soda checks on freshness / row-count SLAs / distributional calibration (soda/checks.yml), and a Dagster-scheduled soda_quality_scan asset that fails the run if any check fails (monetization_orchestration/.../defs/soda_quality.py). Live status surfaces in the Dagster UI, not the BI dashboard — quality monitoring belongs at the operations layer, not the consumer-facing one.

See docs/SETUP.md for the per-step state.