Databricks Genie now exposes a demand‑forecast model (Chronos‑2 via Many Model Forecasting) directly to Amazon Quick, enabling a closed‑loop process that detects demand surges, matches them to live supplier capacity stored in S3 Tables, and automatically creates purchase orders or escalation tickets. Practitioners can replace manual spreadsheet‑driven replenishment with an end‑to‑end automated workflow that runs on a schedule and requires no per‑SKU tuning.
Architecture Overview
The solution consists of four linked stages. Forecast runs in a Databricks workspace where the MMF package serves Chronos‑2 to generate a seven‑day demand projection for every SKU. Detect uses a Databricks Genie Agent to surface SKUs whose next‑seven‑day average exceeds 1.5 × the prior‑fourteen‑day average, filtering out low‑volume noise (average ≥ 1). Decide occurs in Amazon Quick, which reads the forecast via the Genie MCP connector and queries live supplier availability from Amazon S3 Tables using a Direct Query dataset. Act is performed by an Amazon Quick Flow that invokes a Supplier Order API for the cheapest supplier that can satisfy the surge, or raises a human‑review ticket when no single supplier can cover the demand.
Implementation Steps
All components are provisioned from a local terminal. Required tools include the Databricks CLI ≥ 0.299.0, AWS CLI ≥ 2.36.2, jq 1.7, and either uv or Python 3.11 for the supplier‑feed loader. The repository provides three key artifacts: genie/genie_space.json (Genie Agent definition), flow/flow_definition.json (Quick Flow definition), and cleanup/cleanup.sh (teardown script). After cloning the repo, the workflow is:
- Authenticate to the Databricks workspace with
databricks auth loginand note the<WORKSPACE_USER>fromdatabricks auth describe. - Create a Quick space, data source, and flow using the
aws quicksight create‑space,create‑data‑source, andcreate‑flowsubcommands. - Deploy the Genie Agent via the Databricks CLI, ensuring the workspace identity has
CREATE CATALOGpermission. - Run the supplier‑feed loader to populate the S3 Tables dataset.
- Validate that the Quick user holds an Author or Author Pro role; otherwise the MCP and OpenAPI connectors cannot be created.
Each command reads account‑specific values from environment variables, so replace placeholders with your own identifiers before execution.
Operational and Security Implications
Because the forecast and supplier feed remain in separate managed services, the loop avoids materializing a combined data lake, reducing storage cost and latency. However, the integration points introduce cross‑service access requirements: the Genie Agent must be reachable from Quick via the Model Context Protocol, and Quick must have read permissions on the S3 Tables bucket and invoke the external Order API through an OpenAPI connector. Practitioners should audit the IAM policies attached to the Quick user and the Databricks workspace identity to ensure least‑privilege access, especially for the CREATE CATALOG privilege and any role that can trigger purchase orders.
Operationally, the scheduled Quick flow provides deterministic execution, but failures in any stage (e.g., missing supplier data or API errors) will surface as human‑review tickets. Monitoring should therefore include health checks on the Genie Agent response, S3 Tables query latency, and the success rate of Order API calls. Logging from both Databricks and Quick can be aggregated into a central observability platform to trace end‑to‑end order creation.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Engineers can now replace manual replenishment pipelines with a model‑driven, automated loop that leverages existing Databricks forecasting and Amazon Quick orchestration capabilities. The primary actions are to provision the Genie Agent, configure Quick to read the forecast and supplier feed, and grant the minimal IAM roles needed for order creation. Ongoing responsibilities include monitoring cross‑service connectivity, validating that the surge‑detection thresholds remain appropriate for the catalog, and ensuring that any escalation path (human‑review tickets) is integrated with existing incident‑response processes.
