GIS in the Cloud: Beyond ArcGIS Online

GIS is one of those domains where the dominant vendor — Esri — has earned its position through decades of genuine product work, but where the economics of cloud-native alternatives have become hard to ignore. If you are building geospatial analytics on top of a modern data platform, you have options that did not exist five years ago. Most of them are open source, most of them are cheaper than ArcGIS Online, and the maturity is finally good enough for production use.

Here is what a modern GIS stack looks like when you build it on cloud primitives, and where ArcGIS still wins.

PostGIS Is Still the Foundation

The quiet truth of GIS on the cloud is that PostGIS — the spatial extension for PostgreSQL — remains the most important piece of open-source geospatial software ever written. Every analysis pipeline, every web map, every data product we build starts with data in PostGIS or easily loaded into it.

Managed PostgreSQL services on every cloud support PostGIS. AWS RDS, Azure Database for PostgreSQL, GCP Cloud SQL, Supabase, Neon — all of them. Turn it on with a single extension command and you have a production-grade spatial database that will handle multi-million-record datasets without breaking a sweat.

Where PostGIS wins

Complex spatial joins, analytical queries, CRUD on large vector datasets, serving data to web maps via pg_tileserv or Martin, feeding dbt pipelines that join spatial and non-spatial tables. Anything that is fundamentally a database operation with a spatial predicate.

Where PostGIS struggles

Raster processing at scale. PostGIS has raster support but it is not the right tool for processing satellite imagery or LiDAR. For that, use purpose-built tools.

Cloud-Optimized GeoTIFF and STAC

The biggest shift in geospatial data over the last five years has been the Cloud-Optimized GeoTIFF (COG) format and the SpatioTemporal Asset Catalog (STAC) standard. Together they let you store large raster datasets in object storage and query them without downloading the entire file.

The workflow is simple: put your raster data in S3 or Azure Blob as COGs, register it in a STAC catalog (pgstac for self-hosted, or a managed service like Element 84's Earth Search), and then any STAC-aware client can query by space, time, and metadata. Tools like rio-tiler, titiler, and GDAL all know how to read COGs directly from the cloud without downloading them first.

This is how you process terabytes of satellite imagery without moving terabytes of data. The compute reads only the bytes it needs.

Use case fit

This pattern is the right default for any project involving satellite imagery, aerial surveys, or any raster data at multi-gigabyte scale. It is the standard adopted by Microsoft Planetary Computer, AWS Earth on AWS, and most of the modern earth observation ecosystem.

Vector Tiles and Modern Web Maps

Esri's legacy web map stack (ArcGIS Server, ArcGIS Online, the Web AppBuilder) is being quietly replaced in most new projects by the open-source vector tile ecosystem: MapLibre GL (a community fork of Mapbox GL JS), PMTiles, and servers like tileserver-gl or Martin.

PMTiles in particular is a brilliant piece of work. It is a single-file format for vector and raster tiles that can be served from a static object store with no backend. Generate a PMTiles file from your data with Tippecanoe, upload it to S3, point MapLibre at it, and you have a fast web map for less than $5 a month in hosting cost.

This is not a toy architecture. We use it in production for web maps with millions of features. It is dramatically cheaper than any ArcGIS Online solution and the performance is often better because the tiles are served from a CDN close to the user.

Raster Processing on Cloud Compute

For heavy raster processing — mosaicking, reprojection, classification, time-series analysis — the honest answer in 2023 is "spin up cloud compute and run GDAL." The GDAL library is decades of work at this point and the feature set is hard to beat.

For workflows at scale, three tools stand out:

• Dask with xarray for interactive raster analysis on cloud clusters. The Pangeo stack has made this the standard in climate and earth science.
• Sentinel Hub and Google Earth Engine for managed global datasets where you do not want to deal with the data pipeline at all. You write a query and get an answer. Earth Engine has a free tier for research.
• Apache Sedona (formerly GeoSpark) for spatial joins on Spark clusters when you are processing billions of features.

Each of these has a learning curve, but the learning curve for ArcGIS at equivalent scale is not smaller.

Where Esri Still Wins

Let us be honest about where the open-source stack is not yet competitive.

Cartographic output

ArcGIS Pro is still the best tool in the world for making a nice-looking printed map. QGIS is good and improving, but for serious cartography — a thesis, a published atlas, a legal document — ArcGIS still produces better output with less fuss.

Domain-specific extensions

Esri's extensions for utilities, transportation, forestry, and public safety represent years of specialized development. Replacing them with open-source equivalents means rewriting a lot of logic. If your workflow depends on one of these, budget the migration honestly or stay put.

Enterprise support

A lot of organizations want a vendor phone number to call when something breaks. Esri has one. The open-source stack has consulting companies instead, which is a different thing.

The Hybrid Pattern That Usually Wins

For a mid-sized GIS team, the architecture that holds up is:

• PostGIS on managed PostgreSQL for the authoritative spatial database.
• COGs in object storage for raster data, catalogued with STAC.
• Vector tiles (PMTiles or tileserver-gl) for web map delivery.
• MapLibre GL for web clients.
• Python with GeoPandas, Rasterio, xarray for analysis.
• QGIS for desktop work and ArcGIS Pro kept on a few seats for cartographic output and any legacy workflows.

This is cheaper than an all-Esri stack by an order of magnitude for most use cases, and the capabilities are converging fast.

Three Takeaways

1. PostGIS plus COGs plus PMTiles covers 80 percent of GIS use cases at 10 percent of the cost. Start there and add Esri only where it earns its seat.
2. Cloud-optimized formats change the data architecture. Process data in place in object storage instead of moving it around.
3. Do not underestimate the migration cost off Esri's extensions. If your workflow is deep in ArcGIS-specific functionality, the savings have to clear the rewrite.