Query Xarray with SQL
pip install xarray-sqlThis is an experiment to provide a SQL interface for raster data.
import xarray as xr
import xarray_sql as qr
ds = xr.tutorial.open_dataset('air_temperature')
# The same as a dask-sql Context; i.e. an Apache DataFusion Context.
c = qr.Context(ds)
c.create_table('air', ds, chunks=dict(time=24))
df = c.sql('''
SELECT
"lat", "lon", AVG("air") as air_total
FROM
"air"
GROUP BY
"lat", "lon"
''')
# A table of the average temperature for each location across time.
df.compute()
# Alternatively, you can just create the DataFrame from the Dataset:
df = qr.read_xarray(ds)
df.head()Succinctly, we "pivot" Xarray Datasets to treat them like tables so we can run SQL queries against them.
A few reasons:
- Even though SQL is the lingua franca of data, scientific datasets are often inaccessible to non-scientists (SQL users).
- Joining tabular data with raster data is common yet difficult. It could be easy.
- There are many cloud-native, Xarray-openable datasets, from Google Earth Engine to Pangeo Forge. Wouldn’t it be great if these were also SQL-accessible? How can the bridge be built with minimal effort?
This is a light-weight way to prove the value of the interface.
The larger goal is to explore the hypothesis that Pangeo is a scientific database. Here, xarray-sql can be thought of as a missing DB front end.
All chunks in an Xarray Dataset are transformed into a Dask DataFrame via
from_map() and to_dataframe(). For SQL support, we just use dask-sql.
That's it!
Underneath Xarray, Dask, and Pandas, there are NumPy arrays. These are paged in
chuncks and represented contiguously in memory. It is only a matter of metadata
that breaks them up into ndarrays. to_dataframe()
just changes this metadata (via a ravel()/reshape()), back into a column
amenable to a DataFrame.
There is added overhead from duplicating dimensions as columns, which we see as worth the convenience of DataFrames.
Dask doesn't support
MultiIndexs (dask/dask#1493). If
it did, I suspect performance for many types of queries would greatly improve.
Further, while this does play well with dask-geopandas (for geospatial query
support), certain types of operations don't quite match standard geopandas.
Spatial joins come to mind as a killer feature, but only inner joins are
supported (geopandas/dask-geopandas#72)
.
I have a few ideas so far. One approach involves applying operations directly on
Xarray Datasets. This approach is being pursued
here, as xql.
Deeper still: I was thinking we could make a virtual filesystem for parquet that would internally map to Zarr. Raster-backed virtual parquet would open up integrations to numeroustools like dask, pyarrow, duckdb, and BigQuery. More thoughts on this in #4.
I want to give a special thanks to the following folks and institutions:
- Pramod Gupta and the Anthromet Team at Google Research for the problem formation and design inspiration.
- Jake Wall and AI2/Ecoscope for compute resources and key use cases.
- Charles Stern, Stephan Hoyer, Alexander Kmoch, Wei Ji, and Qiusheng Wu for the early review and discussion of this project.
Copyright 2024 Alexander Merose
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
https://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
Some sources are re-distributed from Google LLC via https://github.com/google/Xee (also Apache-2.0 License) with and without modification (specifically, Github Actions workflows). These files are subject to the original copyright; they include the original license header comment as well as a note to indicate modifications (when appropriate).