Skip to content

Implementation Status

This is the canonical source of truth for implementation readiness. All claims in other pages must be consistent with this page.

Last updated: May 24, 2026

v1.0 completed May 24, 2026. See Requirements — Definition of Done and Roadmap — v1.0 Demo Track for scope details.

Legend

  • ✅ Operational — Implemented, tested, and working in practice
  • 🚧 Partial — Partially implemented or requires manual steps
  • 📋 Planned — Designed but not yet implemented

Implementation Matrix

Component Status Notes
Data Engineering
Airflow ETL ✅ Operational Scraping + PostgreSQL ingestion
MinIO Storage ✅ Operational S3-compatible object storage
Fonbet raw snapshot collection ✅ Operational airflow/dags/etl_odds_fonbet_01_raw.py — Selenoid CDP scrape every 4 h → odds_fonbet/ in MinIO
Fonbet match linking ✅ Operational airflow/dags/etl_odds_fonbet_02_link.py → src/pipelines/link_fonbet_odds.py — fuzzy 3-layer match → match_links/fonbet_links.parquet
Fonbet odds extraction ✅ Operational airflow/dags/etl_odds_fonbet_03_odds.py → src/pipelines/fetch_fonbet_odds.py — 1X2 extraction → match_links/fonbet_odds.parquet
DVC Versioning ✅ Operational Data + model artifacts tracked
PostgreSQL ✅ Operational Canonical data store
ML Pipeline
Feature Engineering ✅ Operational stats_matches.py — time-windowed stats
DVC Pipeline ✅ Operational dvc.yaml orchestration working
MLflow Tracking ✅ Operational Experiment logging functional
Train/Test Splitting ✅ Operational Time-based + CV folds
Model Training ✅ Operational XGBoost + HGB classifiers active; LogReg is in the DAG but disabled (tuning_logreg.enabled: false); production-scale parameters active (classification.frac=0.01, tuning.n_trials=20); three-stage tuning → select_model → final_train
Model Selection ✅ Operational select_model stage compares CV log-loss across XGB / HGB tuners; writes data/models/best_model.json
Model Registry 🚧 Partial Registration automated; promote_model stage gates candidate alias automatically; champion alias requires manual sign-off
Model Promotion 🚧 Partial promote_model stage promotes to candidate alias if new logloss within tolerance; champion promotion is manual-only
Serving
FastAPI App ✅ Operational Routers, middleware, lifespan; CORS allow-list driven by CORS_ALLOWED_ORIGINS (default empty = no cross-origin)
GET /predict/predictions/ ✅ Operational Bulk read from predictions.parquet in-memory cache; display cols only
GET /predict/precomputed/{match_id} ✅ Operational Single-match parquet lookup; no Celery, no MLflow call at request time
GET /predict/cards/ ✅ Operational Predictions merged with Fonbet 1X2 odds; primary Streamlit UI data source
GET /predict/region-roi/ ✅ Operational Flat-stake ROI per region from live-betting simulation
GET /predict/odds/ ✅ Operational Fonbet 1X2 odds from fonbet_odds.parquet
GET /predict/{match_id} ✅ Operational Sync Celery dispatch to ml queue; features from match_features.parquet; 30 s timeout
GET /predict/model/info ✅ Operational MLflow model metadata via sync Celery dispatch
POST /predict/async/ 🚧 Partial Schemas + Celery task + polling endpoint operational; HTTP route not yet registered in src/app/routers/predict.py
Request Validation ✅ Operational Pydantic schemas in src/app/schemas/predict.py
Model Loading ✅ Operational Lazy-loaded once per worker process via PredictionService
Batch Predictions API ❌ Out of scope DVC batch_inference stage exists; HTTP batch endpoint is out of scope by design — see ADR-0006
Streamlit UI ✅ Operational src/ui/app/main.py — match list with Fonbet odds, Value bets signal (>5 pp edge), prediction accuracy per match (Pred), dynamic ROI panel (Accuracy / ROI all picks / ROI value bets), Min region ROI slider, filters: Region / Status / Period. Demo disclaimer on every page.
Monitoring
Prometheus Metrics ✅ Operational GET /metrics, _PrometheusMiddleware, 9 counters/histograms/gauges; prediction_requests_total label is source="sync" for on-demand requests
Service Health ✅ Operational GET /healthcheck/, liveness probes, DB connectivity check
Celery Queue Stats ✅ Operational GET /monitoring/celery/queues, /celery/workers
Task Status Polling ✅ Operational GET /monitoring/task_status/{task_id}
Evidently Drift Detection ✅ Operational GET /monitoring/drift — reads reports/drift/latest.json; refreshes model_feature_drift_score Prometheus gauge. Report written by monitor_drift DVC stage / Airflow DAG soccer_ml_monitor_drift_01.
ML Quality Monitor ✅ Operational soccer_ml_monitor_quality_01 — computes log-loss, ECE, hit-rate on finished matches; writes Prometheus textfile + Evidently HTML report; triggered by soccer_etl_inference_01.
Grafana Dashboards ✅ Operational Two dashboards deployed: "Soccer — ML Quality & Betting" (tags: ml, monitoring, soccer), "SoccerPredictAI" (tags: mlops, prediction, soccer).
Alerting Rules 📋 Planned Runbooks documented; rules not deployed in Alertmanager
Infrastructure
Docker Images ✅ Operational Multi-stage builds for API + workers
K8s Manifests ✅ Operational Deployments + Services + ConfigMaps
Helm Charts ✅ Operational Values + templates parameterized; nginx-ingress rate-limit (rps/burst/connections) configurable via ingress.rateLimit
GitLab CI ✅ Operational Build + test + deploy pipeline
Secrets (SOPS) ✅ Operational age encryption for sensitive data
Quality
pytest Framework ✅ Operational 564 tests collected (unit, property, service, contract, load); see reports/planning/20260522_test.md for coverage gap matrix
Unit Tests ✅ Operational tests/unit/ — splitting, schemas, preprocess
Property Tests ✅ Operational tests/property/ — Hypothesis: features, splitting, metrics
Service Tests ✅ Operational tests/service/ — prediction service, Celery tasks
Contract Tests ✅ Operational tests/contract/test_pipeline_contracts.py
Load Tests ✅ Operational tests/load/locustfile.py — Locust load scenarios
Integration Tests 🚧 Partial API mock tests; no live MLflow/Celery required in CI
Pre-commit Hooks ✅ Operational ruff + basic linting
Data Validation ✅ Operational Great Expectations suites at raw / finished / future / features stages

Known Limitations

  1. Model promotion — candidate alias is automated via promote_model stage; champion alias requires manual sign-off. No alerting on promotion failures.
  2. Batch HTTP endpoint — out of scope by design (ADR-0006). Batch inference runs via the DVC batch_inference stage only.
  3. Alerting — Alerting rules documented in runbooks; not deployed in Alertmanager. No on_failure_callback on critical Airflow DAGs.
  4. Integration tests — All CI tests use mocks. No live Celery/MLflow dependency in CI.
  5. Feature store — Features are file-based Parquet; no dedicated online store.
  6. API authentication — /predict/* endpoints require X-API-Key header. /monitoring/* endpoints are unauthenticated (planned).

See Architecture Limitations for deployment-level constraints.

How to verify

# Reproduce the ML pipeline
dvc pull && dvc repro

# Inspect experiments
mlflow ui --port 5001

# Run the test suite
pytest tests/ -q

# Check API health
curl http://localhost:8000/healthcheck/

# Check Prometheus metrics
curl http://localhost:8000/metrics

See Quickstart for the full reproducibility path.