Databricks / Lakehouse architecture preset
Icons: drawio-ai search "databricks …". Databricks has no group/container stencils — draw containers with frame/stage/band; the icons carry identity. Accent red #FF3621 on borders only, never as a fill (clean house style).
Pick the view first — the two Databricks diagrams have different containment:
A. Data Intelligence Platform / lakehouse reference (default — “lakehouse / data platform / pipeline”)
The canonical Databricks reference. Copy drawio-ai scaffold build_data_intelligence_platform.mjs and adapt.
Signature look — do these, they’re what makes a diagram read as Databricks:
- Coral platform header band (fill
#FF3621, WHITE bold text) titling the platform, plus a navyOrchestrationband (fill#1B3139, WHITE text) just inside it.box()hardcodes dark text, so draw these bands with a raw style:rounded=0;whiteSpace=wrap;html=1;fillColor=<coral|navy>;strokeColor=none;fontColor=#FFFFFF;fontStyle=1;verticalAlign=middle;align=center;. Give the bands an explicit width so the column hugs them. - Side zones (Data Sources · Ingestion · Consumers) = WHITE frames with a navy border — never a gray or tinted fill. Fill them with the concept line-icons.
- Medallion = the database icon recolored per Delta layer, in a row: Landing (green) → Bronze → Silver → Gold (
dbx_landing,dbx_bronze,dbx_silver,dbx_gold). - Foundation = a coral band “Unified, Open, Scalable Lakehouse Architecture” over two EQUAL-width cards — Governance (
unity_catalog) and Open Storage (delta,parquet,iceberg) — each with its own navy sub-header band. Give both cards the same width so they align (they sit side by side). - Unity Catalog is cross-cutting governance — dash-link (
governs) to storage AND serving/ML; never a node inside the flow. - The only tint used is the light-coral foundation (
#FDECEA); everything else is white + colored bands. No gray anywhere.
Icon vocabulary (all merged into the databricks pack — drawio-ai search "databricks …"):
- Concept line-icons (navy):
dbx_data_warehouse·dbx_external(on-prem) ·dbx_apps_line·dbx_logs·dbx_events·dbx_cloud_database·dbx_ingestion·dbx_streaming·dbx_pipeline·dbx_data_engineering·dbx_ai_ml·dbx_query·dbx_dashboards_line·dbx_data_sharing·dbx_business_users·dbx_orchestration·dbx_catalog_line·dbx_notebook_line·dbx_cluster·dbx_table·dbx_lineage·dbx_security. - Medallion layers (recolored DB icons, in the Big Data pack):
medallion_landing(green) ·medallion_bronze·medallion_silver·medallion_gold. - Products/logos:
unity_catalog·databricks_sql·mosaic_ai·bi_genie·lakeflow_connect/lakeflow_declarative_pipelines/lakeflow_jobs·delta/parquet/iceberg(Big Data pack) ·tableau·power_bi.
Alternative flavour — some references use compute lanes (Source → Ingest → Transform → Serve → Analysis) over a horizontal Storage band holding the medallion, with a Unity Catalog band under everything. Same principles; pick whichever the request implies.
B. Platform deployment topology (“control plane / data plane / serverless / VPC / on a cloud”)
The plane split is an account-ownership boundary, NOT a network tier:
- Control plane = frame “Databricks account (control plane)” (red border) — workspace/web app (
databricks),notebooks,lakeflow_jobs(jobs UI),databricks_sql(query UI),unity_catalog(metastore + governance). Databricks-managed. - Classic compute plane = a dashed customer-cloud VPC frame (
clusterBox) in the customer cloud account, holding the Spark/photonclusters. Serverless compute plane = a separate frame on the Databricks-account side. - Storage lives in the customer cloud account —
s3/ADLS/GCS bucket(s) for workspace storage + the data lake (medallion). Object storage is not in the VPC — draw beside the classic compute VPC. - Edges: control plane ↔ compute = secure cluster connectivity (solid); compute → storage = read/write (flow); UC → storage & compute =
governs(dashed). Composing with a cloud followsdiagram-types.md§Composing — the cloud is a sibling frame; only the classic compute VPC genuinely nests in the customer account.
C. MLOps across workspaces (“MLOps / CI/CD / dev-staging-prod / promote model”)
The “Big Book of MLOps” layout — copy drawio-ai scaffold build_mlops.mjs.
- Git provider band on top (dev → main → release repos,
github; CI/CD boxes) → three workspace zones side by side: Development · Staging · Production → Unity Catalog band with per-env catalogs (each = Tables + Models) → Lakehouse band at the bottom. - Each workspace holds an MLflow Tracking Server + its stage’s work (dev: EDA + train/validate/deploy/monitor; staging: integration tests; prod: model train-deploy Workflow + Batch Inference + Monitoring) and a Model Serving Endpoint (coral).
- Refined tones (match the reference): each zone is a muted header band flush to the top edge (white text, left-aligned, with the Databricks logo at the left) over a WHITE body — muted steel-blue (Dev), dusty maroon (Staging), sage green (Prod), gray (Git), coral (Unity Catalog), navy (Lakehouse). Only the header is colored; bodies stay white. Catalog cards are white with a navy sub-header. The logo is a rasterized white mark in a
shape=label;image=…;imageAlign=leftband (embed a PNG data-URI — draw.io’s PNG export does not rasterize embedded SVG). - Edges carry the flow:
Pull request to main,Merge to release,CI trigger,Continuous Deployment,Logging,Register · promote model.
Governance hierarchy (when the diagram is about Unity Catalog)
Account → Metastore (one per cloud region per account) → Catalog → Schema → Table | View | Volume | Function | Model. Storage credentials, external locations, connections and shares sit under the metastore. Tables/volumes are managed (UC owns storage) or external (UC governs only). Draw as a nested frame tree (3-level namespace catalog.schema.object), not a flow.
Edges
Solid = data/control flow; dashed = governance/lineage/sharing. Medallion Bronze→Silver→Gold = flow. Connect to a band’s border, not each child. flow:true on the main ingest→medallion→serve spine.
Source: skills/drawio/references/databricks-architecture.md on GitHub ↗