dask-sql is a distributed SQL query engine in Python. It allows you to query and transform your data using a mixture of common SQL operations and Python code and also scale up the calculation easily if you need it.
- Combine the power of Python and SQL: load your data with Python, transform it with SQL, enhance it with Python and query it with SQL - or the other way round. With
dask-sqlyou can mix the well known Python dataframe API ofpandasandDaskwith common SQL operations, to process your data in exactly the way that is easiest for you. - Infinite Scaling: using the power of the great
Daskecosystem, your computations can scale as you need it - from your laptop to your super cluster - without changing any line of SQL code. From k8s to cloud deployments, from batch systems to YARN - ifDasksupports it, so willdask-sql. - Your data - your queries: Use Python user-defined functions (UDFs) in SQL without any performance drawback and extend your SQL queries with the large number of Python libraries, e.g. machine learning, different complicated input formats, complex statistics.
- Easy to install and maintain:
dask-sqlis just a pip/conda install away (or a docker run if you prefer). No need for complicated cluster setups -dask-sqlwill run out of the box on your machine and can be easily connected to your computing cluster. - Use SQL from wherever you like:
dask-sqlintegrates with your jupyter notebook, your normal Python module or can be used as a standalone SQL server from any BI tool. It even integrates natively with Apache Hue.
Read more in the documentation.
Example
For this example, we use some data loaded from disk and query them with a SQL command from our python code. Any pandas or dask dataframe can be used as input and dask-sql understands a large amount of formats (csv, parquet, json,...) and locations (s3, hdfs, gcs,...).
import dask.dataframe as dd
from dask_sql import Context
# Create a context to hold the registered tables
c = Context()
# Load the data and register it in the context
# This will give the table a name, that we can use in queries
df = dd.read_csv("...")
c.create_table("my_data", df)
# Now execute a SQL query. The result is again dask dataframe.
result = c.sql("""
SELECT
my_data.name,
SUM(my_data.x)
FROM
my_data
GROUP BY
my_data.name
""", return_futures=False)
# Show the result
print(result)
Quickstart
Have a look into the documentation or start the example notebook on binder.
dask-sqlis currently under development and does so far not understand all SQL commands (but a large fraction). We are actively looking for feedback, improvements and contributors!
If you would like to utilize GPUs for your SQL queries, have a look into the blazingSQL project.
Installation
dask-sql can be installed via conda (preferred) or pip - or in a development environment.
With conda
Create a new conda environment or use your already present environment:
conda create -n dask-sql
conda activate dask-sql
Install the package from the conda-forge channel:
conda install dask-sql -c conda-forge
With pip
dask-sql needs Java for the parsing of the SQL queries. Make sure you have a running java installation with version >= 8.
To test if you have Java properly installed and set up, run
$ java -version
openjdk version "1.8.0_152-release"
OpenJDK Runtime Environment (build 1.8.0_152-release-1056-b12)
OpenJDK 64-Bit Server VM (build 25.152-b12, mixed mode)
After installing Java, you can install the package with
pip install dask-sql
For development
If you want to have the newest (unreleased) dask-sql version or if you plan to do development on dask-sql, you can also install the package from sources.
git clone https://github.com/nils-braun/dask-sql.git
Create a new conda environment and install the development environment:
conda create -n dask-sql --file conda.txt -c conda-forge
It is not recommended to use pip instead of conda for the environment setup. If you however need to, make sure to have Java (jdk >= 8) and maven installed and correctly setup before continuing. Have a look into conda.txt for the rest of the development environment.
After that, you can install the package in development mode
pip install -e ".[dev]"
To compile the Java classes (at the beginning or after changes), run
python setup.py java
This repository uses pre-commit hooks. To install them, call
pre-commit install
Testing
You can run the tests (after installation) with
pytest tests
SQL Server
dask-sql comes with a small test implementation for a SQL server. Instead of rebuilding a full ODBC driver, we re-use the presto wire protocol. It is - so far - only a start of the development and missing important concepts, such as authentication.
You can test the sql presto server by running (after installation)
dask-sql-server
or by using the created docker image
docker run --rm -it -p 8080:8080 nbraun/dask-sql
in one terminal. This will spin up a server on port 8080 (by default) that looks similar to a normal presto database to any presto client.
You can test this for example with the default presto client:
presto --server localhost:8080
Now you can fire simple SQL queries (as no data is loaded by default):
=> SELECT 1 + 1;
EXPR$0
--------
2
(1 row)
You can find more information in the documentation.
CLI
You can also run the CLI dask-sql for testing out SQL commands quickly:
dask-sql --load-test-data --startup
(dask-sql) > SELECT * FROM timeseries LIMIT 10;
How does it work?
At the core, dask-sql does two things:
- translate the SQL query using Apache Calcite into a relational algebra, which is specified as a tree of java objects - similar to many other SQL engines (Hive, Flink, ...)
- convert this description of the query from java objects into dask API calls (and execute them) - returning a dask dataframe.
For the first step, Apache Calcite needs to know about the columns and types of the dask dataframes, therefore some java classes to store this information for dask dataframes are defined in planner. After the translation to a relational algebra is done (using RelationalAlgebraGenerator.getRelationalAlgebra), the python methods defined in dask_sql.physical turn this into a physical dask execution plan by converting each piece of the relational algebra one-by-one.
![[BUG] CVEs in conda release](https://avatars.githubusercontent.com/u/4249447?v=4)
![[ENH] substr() not supported in dask-sql](https://avatars.githubusercontent.com/u/29827014?v=4)
![[BUG] Segfaults on](https://avatars.githubusercontent.com/u/1692914?v=4)
![[BUG] Fix `test_random` on Dask cluster](https://avatars.githubusercontent.com/u/53962159?v=4)
![[BUG] Schema: <schema> not found in DaskSQLContext](https://avatars.githubusercontent.com/u/85463385?v=4)
![[ENH] Add ~isna() support for predicate pushdown](https://avatars.githubusercontent.com/u/19949207?v=4)
6 Oct 31, 2022
35 Dec 23, 2022
5 Jun 24, 2021
1.3k Jan 03, 2023
11k Dec 29, 2022
4 Oct 21, 2022
6 May 30, 2022
629 Nov 21, 2022
853 Dec 30, 2022
1.4k Dec 31, 2022
666 Dec 30, 2022
1.9k Dec 25, 2022
554 Dec 04, 2022
12 Jan 01, 2023
3.3k Dec 31, 2022
66 Dec 02, 2022
17 Oct 10, 2021
1 Feb 08, 2022
4 Dec 01, 2021