Cloud-native geospatial — without the copy.

We build tools that query Zarr and NetCDF archives with SQL, right where they sit in object storage. You don't copy anything into a database first.

stratoscale.io · Blog · hi@stratoscale.io

Climate and Earth-observation archives already sit in object storage, chunked and compressed. The usual advice is to copy them into a warehouse before you can query them. That costs egress, duplicates petabytes, and the copy is stale the day it lands.

We'd rather send the query to the data. That's what we build.

Everything uses open formats: Zarr, Apache Arrow, Apache DataFusion.

A SQL engine for Zarr-native array data, built on Apache DataFusion.

It reads Zarr v2 and v3, and infers the Arrow schema from the store's own metadata. Filters and projections push down, so a query skips the chunks it doesn't need. MIN, MAX and COUNT come from statistics without reading any data. Scans run in parallel and read straight from GCS and S3. VirtualiZarr reference stores point the same SQL at NetCDF and GRIB files in place.

The CLI has an extended DESCRIBE and per-query I/O stats. Larger scans can spread across several nodes.

CREATE EXTERNAL TABLE era5 STORED AS ZARR
LOCATION 'gs://gcp-public-data-arco-era5/ar/model-level-1h-0p25deg.zarr-v1';

SELECT latitude, longitude, AVG(2m_temperature)
FROM era5
WHERE time > '2020-01-01'
GROUP BY latitude, longitude;

An OGC API - Environmental Data Retrieval server built on that engine.

position, radius, area and cube queries turn into SQL and push down to chunk reads. Results come back as CoverageJSON, GeoJSON or HTML. With no configuration it serves the public ARCO-ERA5 store. Give it a TOML file and it serves as many stores as you list, reading the axes, extents and parameters from the stores themselves at startup.

Every resource has an HTML page at the same URL as its JSON. A result page's address is the request that produced it, one f= away from CoverageJSON.

Each recipe is checked against a reference.

• ONI from ERA5 computes the Oceanic Niño Index for every overlapping three-month season since 1950. It's one SQL query over public ERA5 on GCS. In the satellite era it tracks NOAA's official table within 0.14 °C on average, at r = 0.96.
• NDVI is (b08 - b04) / (b08 + b04) over two Sentinel-2 bands, written as a single SQL projection. It matches xarray to four decimals.

We write these up on the blog.

We're a small senior team specialising in cloud-native array data. We take hands-on production work. Write to hi@stratoscale.io.