THINK Gas

Target design · TD-001 · proposed

Deployable technical architecture

A lakehouse plus warehouse serving certified measures, operational APIs and permission-aware retrieval. All paths are environment-separated, identity-controlled, observable and replayable.

Flow

00

Sources

01

Connectors

02

Bronze Delta

03

Silver products

04

Semantic models

05

AI / BI serving

00 — Sources

16 registered applications; extraction contracts pending validation

DMAHEATiPortalGIS PortalSCADA ApplicationNFA PortalHSE iReporting PortalLoyalty PortalQAM PortalSupplier PortalDBMS / LMS PortalHRMSCustomer Self-Service PortalSAP (ERP)TIPMSpot Billing

01 — Ingestion

Incremental extraction and contract enforcement

  • CDC (Debezium) on MariaDB / MySQL / PostgreSQL sources
  • SAP SLT + OData extractors for ERP objects
  • MQTT / OPC-UA bridge for SCADA telemetry
  • REST + webhook connectors for .NET and Laravel portals
  • Schema registry with contract validation at the edge

02 — Data Lake (Bronze)

Immutable Delta landing and replay

  • Object storage with open table format (Iceberg / Delta)
  • Partitioned by source system and event date
  • Exactly-once landing with checkpoint offsets
  • Raw payload retained with ingest metadata columns

03 — Conformed (Silver)

Conformed entities, history and quality controls

  • Deduplication, type casting, late-arriving record handling
  • Master data resolution: customer, premise, asset, employee, supplier
  • GIS geometry standardisation (single CRS) for spatial joins
  • Data quality tests (freshness, uniqueness, referential integrity)

04 — Warehouse (Gold)

Certified semantic measures and governed SQL

  • Dimensional models: dim_customer, dim_asset, dim_geo, dim_employee
  • Facts: fct_connection_funnel, fct_consumption, fct_incident, fct_billing
  • Semantic/metric layer so every tool shares one definition
  • Incremental builds orchestrated with dbt + Airflow

05 — AI & Consumption

Policy-aware retrieval and monitored model serving

  • BI layer on the semantic models
  • Feature store for ML training and serving
  • Vector store over documents (HSE, policies, SOPs) for RAG
  • Text-to-SQL assistant scoped to governed gold models

Non-functional control plane

Controls applied to every environment

Access & privacy

Entra groups, least privilege, managed identities, secrets vault, row/column security, masking and immutable audit.

Catalogue & lineage

Owner, classification, schema, quality SLO and column lineage required before certification.

Cost & reliability

Freshness, completeness, duplicates, drift, job failures, model quality, token and capacity consumption.

Release & recovery

Dev/test/prod isolation, IaC, pull-request gates, deployment identity, rollback, backup and timed restore exercise.

Decision record · benchmark required

One primary platform; earn any second platform

No final selection is valid before identical SAP/GIS workloads are benchmarked for elapsed time, source impact, concurrency, operability, security controls and total capacity. Pricing is excluded until measured volumes and commercial terms are available.

Best fit

Microsoft Fabric
Fast governed analytics, Power BI and lower operating overhead
Databricks
Advanced ML, high-volume telemetry and engineering flexibility

Storage & BI

Microsoft Fabric
OneLake, Delta and native DirectLake semantic models
Databricks
Open Delta lakehouse with broad BI connectivity

Data science

Microsoft Fabric
Integrated and improving; enough for early models
Databricks
Mature MLflow, feature engineering and model serving

Governance

Microsoft Fabric
Purview plus Microsoft identity and policy ecosystem
Databricks
Unity Catalog with strong fine-grained lineage and controls

Selection gate

Microsoft Fabric
Choose if benchmark meets SLOs and Microsoft support model is viable
Databricks
Choose if telemetry/ML/geospatial benchmark materially outperforms and skills exist

Company Brain query path

  1. 01 · Authenticate user and resolve role, geography and purpose.
  2. 02 · Route metric questions to certified semantic models; documents to hybrid retrieval.
  3. 03 · Apply row/column policy before retrieval or SQL execution.
  4. 04 · Return citations, freshness and metric definition; abstain below confidence threshold.
  5. 05 · Log prompt, sources, policy result, response and user feedback.

Recovery objectives · proposed

Critical serving
RTO 4h / RPO 1h — validate
Analytical products
RTO 8h / RPO 24h — validate
Pipeline failure
Quarantine; alert owner; replay from durable checkpoint
AI failure
Disable endpoint; retain governed report/manual workflow