Key takeaways:
- FME reads and writes Snowflake, supporting spatial types, semi-structured VARIANT, and bulk loading for large datasets.
- SQL pushdown lets filters and spatial queries run inside Snowflake rather than pulling full tables into FME, cutting processing time and cost.
- FME Flow automates workflows with event-driven triggers (webhooks, file drops, schedules) and can run FME engines inside Snowpark Container Services for in-warehouse compute.
- FME bridges Snowflake’s spatial limits, handling geometry validation, reprojection, raster-to-vector conversion, 3D/BIM processing, and LiDAR point clouds.
- FME Flow can host an MCP server that turns any FME workspace into a tool AI agents can call, enabling natural-language access to Snowflake data and spatial analysis without exposing raw data to external systems.
When Austin’s transit authority needed to migrate complex transit data and real-time feeds into Snowflake, FME automated the entire ingestion and transformation process: over 35,000 jobs a day across more than 100 workflows, processing a million records in minutes. That’s the kind of gap FME is built to close.
Snowflake is well-established as a cloud data platform for storage, querying, and collaboration. When it comes to connecting to hundreds of heterogeneous source formats, handling complex spatial data, running transformations beyond SQL, and automating the pipelines that keep data current, FME bridges that gap.
- For a deeper dive into this topic and live demos, watch our webinar, Connect, Process, Enrich: Extending Snowflake’s Reach with FME.
FME sits between source data and Snowflake services, converting and loading data from over 500 formats: GIS files, CAD drawings, BIM models, databases, cloud APIs, raster imagery, point clouds, and more. The core integration works through three components.
The Snowflake Reader and Writer. FME connects to Snowflake and supports a wide range of native data types: TIMESTAMP, GEOGRAPHY and GEOMETRY for spatial data, VARIANT for semi-structured content, and standard scalar types. For large datasets, bulk loading provides a significant throughput advantage over row-by-row inserts. The writer also gives you control over schema, so you can define column data types and set table handling directly in the workspace.
SQL integration and pushdown. FME supports running Snowflake SQL through transformers like SQLCreator and SQLExecutor. A WHERE clause on the reader is the simplest form of pushdown: only matching rows come into FME. The SQLExecutor goes further. You can flatten nested VARIANT fields into columns at read time, push aggregations like GROUP BY into Snowflake so only the summary returns, or run custom load statements the standard writer doesn’t expose. Spatial filters work too: ST_ functions run inside Snowflake before any data leaves the warehouse.
Authentication options. FME supports basic username/password, browser-based SSO and MFA, RSA key-pair authentication, and OAuth, so however your organization manages access, there’s a secure option that fits.
Spatial data: where FME extends what Snowflake can do natively
Snowflake supports GEOGRAPHY and GEOMETRY types and a set of ST_ functions. To go beyond standard spatial queries, such as geometry repair, coordinate system reprojection, topology analysis, raster processing, 3D models, or LiDAR, you need FME.
Vector data. FME supports thousands of coordinate systems and can reproject data before loading into Snowflake or after retrieving it. It handles format-native geometry types (arcs, meshes, multi-part geometries), validates and repairs invalid geometries, performs advanced spatial overlays, builds and analyzes network topology, and runs Esri ArcPy for organizations on the Esri stack. Converting between points, lines, and polygons is a single transformer operation rather than custom SQL.
Raster data. FME can clip, tile, mosaic, and resample raster imagery, overlay vector features on rasters to extract pixel values, and convert raster cells to vector polygons that load directly into a Snowflake GEOMETRY column. A practical example: a land-cover GeoTIFF where each pixel represents a classification (urban, forest, agricultural land) can be clipped to a study area, reclassified into fewer categories, converted to polygons, dissolved, and written to Snowflake with area statistics, all without writing or maintaining custom raster code. The resulting table (land-cover class, area in km², geometry) is queryable, joinable, and dashboard-ready.
3D and BIM data. Building Information Models contain thousands of properties per object, right down to manufacturer and warranty data for individual components. FME extracts that structured data and writes it to Snowflake, where it can be joined with business data and queried by AI tools. For example, an agent with access to the right Snowflake tables could answer questions like “which buildings have HVAC units whose warranty expires in the next 90 days?”, provided the BIM data has been loaded and normalized. FME also handles LiDAR point clouds, creating point tables from raw x/y/z coordinates stored in Snowflake.
Common workflow patterns
A few patterns come up repeatedly when integrating FME with Snowflake.
Scheduled ingestion: source data is read, transformed, staged, and merged into production Snowflake tables on a schedule. This is the straightforward ETL pattern, but with FME handling the format conversion and geometry normalization that would otherwise require preprocessing scripts.
Event-driven pipelines: a trigger fires when a file is dropped, a webhook is received, or a message arrives on a queue. FME Flow runs a workspace that validates, transforms, and loads the data, then sends a notification, useful for near-real-time ingestion like a field survey form that writes to Snowflake within seconds of submission.
In-database processing: use SQL pushdown and spatial filtering to do aggregation and analysis inside Snowflake, pulling only the result set into FME for final transformation or output.
Reverse publishing: read from Snowflake, apply transformations or format conversions, and push results to a REST API, a GIS web service, an ArcGIS feature service, or a BI dashboard refresh.
Automated delta sync: detect changes, compare new and existing records, perform upserts, and write an audit log.
Automating pipelines with FME Flow
FME Flow is the orchestration layer that runs FME workspaces on a schedule, in response to events, or via API call. A few capabilities matter most for Snowflake workflows.
Automations connect triggers to actions: a file lands in an S3 bucket, a webhook fires from a field data app, or a schedule hits, and Flow runs the appropriate workspace and loads the data to Snowflake without manual intervention. Jobs can run in sequence or in parallel, retry on failure, and process large datasets across multiple engines.
Workspace Apps turn any FME workflow into a no-code web form. Parameters in the workspace become form fields (file uploads, map selections, date pickers). A user submits the form, Flow executes the workspace, and the results stream back without the user needing to know anything about the underlying pipeline. A good example is CAD file validation before ingestion: a user drops in a CAD file, FME checks the geometry, and either writes valid features to Snowflake or returns an HTML report listing the bad ones.
Streaming lets FME Flow connect to high-velocity message brokers such as Kafka, RabbitMQ, and IBM MQ, processing and loading real-time streams directly into Snowflake.
Custom APIs expose FME’s data processing as REST endpoints, so other systems can trigger spatial queries, format conversions, or Snowflake loads on demand.
Running FME engines inside Snowflake (Remote Engines)
FME Flow Remote Engines, available as a Native App through Snowpark Container Services, lets FME processing jobs run directly inside Snowflake’s compute environment. An existing FME Flow instance handles orchestration and scheduling; the engine runs inside the container and can deliver up to a 30% performance improvement by eliminating data movement overhead. The tradeoff is higher Snowflake compute consumption, but for jobs where data gravity matters, keeping sensitive data inside the warehouse or processing datasets that are expensive to egress, it’s the right architecture.
Learn more: Integrate FME into Snowflake Marketplace
Connecting AI agents to Snowflake through an FME Flow MCP server
FME Flow lets you host a Model Context Protocol (MCP) server, turning any FME workspace into a callable tool for AI agents. An AI client (Claude, Copilot, Gemini, or any MCP-compatible application) connects to that server, sees the available tools and their descriptions, and can invoke them.
This matters for Snowflake workflows because it gives AI agents a governed, auditable way to query and act on Snowflake data. The agent never has direct database access: it calls a named tool, FME runs the workspace (which holds the connection credentials, query logic, and any validation), and returns a structured result. Every call is logged as a Flow job, and the data the agent can see is exactly what the tool exposes and nothing more.
Snowflake Cortex agents, as well as GIS tools like Geocortex, can connect to the MCP server to access FME’s full range of capabilities, whether that’s spatial validation beyond what Snowflake natively supports, ingesting additional data formats, or any other workflow the team has built. There is also an MCPCaller transformer that works in the other direction: FME workspaces can call external MCP tools, combining FME’s data processing with AI inference steps in a single pipeline.
Deployment flexibility
FME Flow deploys on-premises, in any cloud (AWS, Azure, Google Cloud, and sovereign clouds), in Docker/Kubernetes containers, or as FME Flow Hosted, Safe Software’s fully managed SaaS version on AWS. Remote Engines extend this to run inside Snowpark. For organizations with data sovereignty or air-gap requirements, FME can run on-premises with a local AI model, keeping all data, and the AI, inside the firewall.
Get Started
Snowflake gives you a powerful platform to store, scale, and query all your data. FME complements it by connecting hundreds of source formats, transforming and validating spatial data beyond native SQL, and automating the pipelines that keep everything current. Together, they let your team move from raw, scattered source data to governed, analysis-ready Snowflake tables, without writing and maintaining custom code.
Learn more:
