Atlas Go-Live Feedback¶
Post-go-live observations from running the full data-eng-lab platform in a production-like environment.
Summary¶
All A1–A9 capabilities verified during go-live. The platform is fully operational. 19 scenarios executed with parity between Scala and PySpark notebooks where applicable.
Key Observations¶
- MinIO stability — MinIO handled the full dataset load (all 5 datasets at medium scale) without issues. Disk usage should be monitored as scenarios are re-run.
- Iceberg REST catalog — The REST catalog responded to concurrent queries from Spark, Trino, and PyIceberg without contention.
- Spark Connect — The shared PySpark session in JupyterHub is stable across notebook executions. Each notebook manages its own session lifecycle.
- Trino performance — Trino performed well on the dataset sizes used. TPC-H at larger scales may need tuning.
- Streaming reliability — Redpanda handled streaming workloads without issues. The
foreachBatchCDC pattern produced correct results. - Airflow orchestration — Airflow successfully scheduled Spark jobs via
SparkSubmitOperator. DAG execution was reliable. - Jenkins CI — The JAR build and publish pipeline works end-to-end.
Recommendations¶
- Add a scheduled cleanup for streaming checkpoint directories.
- Consider implementing dataset versioning for landing data to support reproducible scenario runs.
- Add observability metrics for the Iceberg REST catalog (query counts, latency).
- Consider adding a data quality monitoring dashboard for Bronze/Silver/Gold tables.
Related¶
- Atlas Expectations — Full delivery log
- Go-Live Results — Detailed validation results
- Atlas Feedback A7/A9 — Streaming and federated query feedback