Query Pandas-like Dataframes Using SQL

QPD let you run the same SQL (SELECT for now) statements on different computing frameworks with pandas-like interfaces. Currently, it support Pandas, Dask and Ray (via Modin on Ray).

QPD directly translates SQL into pandas-like operations to run on the backend computing frameworks, so it can be significantly faster than some other approaches, for example, to dump pandas dataframes into SQLite, run SQL and convert the result back into a pandas dataframe. However, the top priorities of QPD are correctness and consistency. It ensures the results of implemented SQL features following SQL convention, and it ensures consistent behavior regardless of backend computing frameworks. A typical case is groupby().agg(). In pandas or pandas like frameworks, if any of the group keys is null, the default behavior is to drop that group, however, in SQL they are not dropped. QPD follows the SQL way.

QPD syntax is a subset of the intersection of Spark SQL and SQLite. The correctness and consistency are extensively tested against SQLite. Practically, Spark SQL and SQLite are highly consistent on both syntax and behavior. So, in other words, QPD enables you to run common SQLs and get the same result on Pandas, SQLite, Spark, Dask, Ray and other backends that QPD will support in the future.

SQL is one of the most important data processing languages. It is very scale agnostic, and one of the major goals of the Fugue project is to enrich SQL and make SQL platform agnostic. QPD, as a subproject of Fugue, focuses on running SQL on pandas-like frameworks, it is an essential component to achieve the ultimate goal.

Installation

QPD can be installed from PyPI:

pip install qpd # install qpd + pandas

If you want to use Ray or Dask as the backend, you will need to install QPD with one of the targets:

pip install qpd[dask] # install qpd + dask[dataframe]
pip install qpd[ray] # install qpd + ray
pip install qpd[all] # install all dependencies above

Using QPD

On Pandas

from qpd_pandas import run_sql_on_pandas
import pandas as pd

df = pd.DataFrame([[0,1],[2,3],[0,5]], columns=["a","b"])
res = run_sql_on_pandas("SELECT a, SUM(b) AS b, COUNT(*) AS c FROM df GROUP BY a", {"df": df})
print(res)

On Dask

Please read this to learn the best practice for initializing dask.

from qpd_dask import run_sql_on_dask
import dask.dataframe as pd
import pandas

df = pd.from_pandas(pandas.DataFrame([[0,1],[2,3],[0,5]], columns=["a","b"]))
res = run_sql_on_dask("SELECT a, SUM(b) AS b, COUNT(*) AS c FROM df GROUP BY a", {"df": df})
print(res.compute())  # dask dataframe is lazy, you need to call compute

On Ray

Please read this to learn the best practice for initializing ray. And read this for initializing modin + ray.

Please don't use dask as modin backend if you want to use QPD, it's not tested

import ray
ray.init()

from qpd_ray import run_sql_on_ray
import modin.pandas as pd

df = pd.DataFrame([[0,1],[2,3],[0,5]], columns=["a","b"])
res = run_sql_on_ray("SELECT a, SUM(b) AS b, COUNT(*) AS c FROM df GROUP BY a", {"df": df})
print(res)

Ignoring Case in SQL

By default, QPD requires users to use upper cased keywords, otherwise syntax errors will be raised. However if you really don't like this behavior, you can turn it off, the parameter is ignore_case, here is an example:

from qpd_pandas import run_sql_on_pandas
import pandas as pd

df = pd.DataFrame([[0,1],[2,3],[0,5]], columns=["a","b"])
res = run_sql_on_pandas(
    "select a, sum(b) as b, count(*) as c from df group by a",
    {"df": df}, ignore_case=True)
print(res)

Things to clarify

QPD on Spark (Koalas)?

No, that will not happen. QPD is using Spark SQL syntax file. Spark SQL is highly optimized. If we create a Koalas backend, correctness and consistency can be guaranteed, but there will be no performance advantage. So for Spark, please use Spark SQL. If you use Fugue SQL on Spark backend, it will also directly use Spark to run the SQL statements. We don't see the value to make QPD run on Spark.

Update History

• 0.4.4: Remove triad version constraint
• 0.4.3: Fix packing issue
• 0.4.2: Refactor to use the latest triad, fix packaging issues
• 0.4.1: Make Pandas 2 compatible
• 0.4.0: Support arbitrary column name
• 0.2.6: Update pandas indexer import
• 0.2.5: Update antlr to 4.9
• 0.2.4: Fix a bug: set operations will alter the input dataframe to add columns
• 0.2.3: Refactor and extract out PandasLikeUtils class
• 0.2.2: Accept constant select without FROM, SELECT 1 AS a, 'b' AS b
• <= 0.2.1: Pandas, Dask, Ray SQL support