Operating on Dask Dataframes with SQL

Dask-SQL is an open source project and Python package leveraging Apache Calcite to provide a SQL frontend for Dask dataframe operations, allowing SQL users to take advantage of Dask's distributed capabilities without requiring an extensive knowledge of the dataframe API.

In [ ]:
! pip install dask-sql

Set up a Dask cluster

Setting up a Dask Cluster is optional, but can dramatically expand our options for distributed computation by giving us access to Dask workers on GPUs, remote machines, common cloud providers, and more). Additionally, connecting our cluster to a Dask Client will give us access to a dashboard, which can be used to monitor the progress of active computations and diagnose issues.

For this notebook, we will create a local cluster and connect it to a client. Once the client has been created, a link will appear to its associated dashboard, which can be viewed throughout the following computations.

In [ ]:
from dask.distributed import Client

client = Client(n_workers=2, threads_per_worker=2, memory_limit='1GB')

Create a context

A dask_sql.Context is the Python equivalent to a SQL database, serving as an interface to register all tables and functions used in SQL queries, as well as execute the queries themselves. In typical usage, a single Context is created and used for the duration of a Python script or notebook.

In [ ]:
from dask_sql import Context
c = Context()

Load and register data

Once a Context has been created, there are a variety of ways to register tables in it. The simplest way to do this is through the create_table method, which accepts a variety of input types which Dask-SQL then uses to infer the table creation method. Supported input types include:

Input type can also be specified explicitly by providing a format. When being registered, tables can optionally be persisted into memory by passing persist=True, which can greatly speed up repeated queries on the same table at the cost of loading the entire table into memory. For more information, see Data Loading and Input.

In [ ]:
import pandas as pd
from dask.datasets import timeseries

# register and persist a dask table
ddf = timeseries()
c.create_table("dask", ddf, persist=True)

# register a pandas table (implicitly converted to a dask table)
df = pd.DataFrame({"a": [1, 2, 3]})
c.create_table("pandas", df)

# register a table from local storage; kwargs are passed on to the underlying table creation method

Tables can also be registered through SQL CREATE TABLE WITH or CREATE TABLE AS statements, using the sql method.

In [ ]:
# replace our table from local storage
    WITH (
        location = 'surveys/data/2021-user-survey-results.csv.gz',
        format = 'csv',
        parse_dates = ARRAY [ 'Timestamp' ]

# create a new table from a SQL query
    CREATE TABLE filtered AS (
        SELECT id, name FROM dask WHERE name = 'Zelda'

All registered tables can be listed with a SHOW TABLES statement.

In [ ]:
c.sql("SHOW TABLES FROM root").compute()

Dask-SQL currently offers experimental GPU support, powered by the RAPIDS suite of open source GPU data science libraries. Input support is currently limited to Dask / Pandas-like dataframes and data in local/remote storage, and though most queries run without issue, users should expect some bugs or undefined behavior. To register a table and mark it for use on GPUs, gpu=True can be passed to a standard create_table call, or its equivalent CREATE TABLE WITH query (note that this requires cuDF and Dask-cuDF).

# register a dask table for use on GPUs (not possible in this binder)
c.create_table("gpu_dask", ddf, gpu=True)

# load in a table from disk using GPU-accelerated IO operations
    WITH (
        location = 'surveys/data/2021-user-survey-results.csv.gz',
        format = 'csv',
        parse_dates = ARRAY [ 'Timestamp' ],
        gpu = True

Query the data

When the sql method is called, Dask-SQL hands the query off to Apache Calcite to convert into a relational algebra - essentially a list of SQL tasks that must be executed in order to get a result. The relational algebra of any query can be viewed directly using the explain method.

In [ ]:
print(c.explain("SELECT AVG(x) FROM dask"))

From here, this relational algebra is then converted into a Dask computation graph, which ultimately returns (or in the case of CREATE TABLE statements, implicitly assigns) a Dask dataframe.

In [ ]:
c.sql("SELECT AVG(x) FROM dask")

Dask dataframes are lazy, meaning that at the time of their creation, none of their dependent tasks have been executed yet. To actually execute these tasks and get a result, we must call compute.

In [ ]:
c.sql("SELECT AVG(x) FROM dask").compute()

Looking at the dashboard, we can see that executing this query has triggered some Dask computations.

Because the return value of a query is a Dask dataframe, it is also possible to do follow-up operations on it using Dask's dataframe API. This can be useful if we want to perform some complex operations on a dataframe that are not possible through Dask, then follow up with some simpler operations that can easily be expressed through the dataframe API.

In [ ]:
# perform a multi-column sort that isn't possible in Dask
res = c.sql("""
    SELECT * FROM dask ORDER BY name ASC, id DESC, x ASC

# now do some follow groupby aggregations
res.groupby("name").agg({"x": "sum", "y": "mean"}).compute()

Custom functions and aggregations

When standard SQL functionality is insufficient, it is possible to register custom functions for use in queries. These functions can be classified as one of the following:

  • Column-wise functions
  • Row-wise functions
  • Aggregations

Column-wise functions

Column-wise functions can take columns or literal values as input and return a column of an identical length. Column-wise functions can be registered in a Context using the register_function method.

In [ ]:
import numpy as np

def f(x):
    return x ** 2

c.register_function(f, "f", [("x", np.float64)], np.float64)

Function registration requires the following inputs:

  • A callable function
  • A name for the function to be referred to in queries
  • A list of tuples, representing the input variables and their respective types, which can be either Pandas or NumPy types
  • A type for the output column

Once a function has been registered, it can be called like any other standard SQL function.

In [ ]:
c.sql("SELECT F(x) FROM dask").compute()

Row-wise functions

In some cases, it may be easier to write a custom function that processes a dict-like row object - otherwise known as a row-wise function. These functions can also be registered using register_function by passing row_udf=True, and used in the same manner as a column-wise function.

In [ ]:
def g(row):
    if row["x"] > row["y"]:
        return row["x"] - row["y"]
    return row["y"] - row["x"]

c.register_function(g, "g", [("x", np.float64), ("y", np.float64)], np.float64, row_udf=True)

c.sql("SELECT G(x, y) FROM dask").compute()

Note that unlike column-wise functions, which are called directly using specified columns and literals as input, row-wise functions are called using apply, which can have unpredictable performance depending on the underlying dataframe library (e.g. Pandas, cuDF) and the function itself.


Aggregations take a single column as input and return a single value - thus, they can only be used to reduce the results of a GROUP BY query. Aggregations can be registered using the register_aggregation method, which is functionally similar to register_function but takes a Dask Aggregation as input instead of a callable function.

In [ ]:
import dask.dataframe as dd

my_sum = dd.Aggregation("my_sum", lambda x: x.sum(), lambda x: x.sum())

c.register_aggregation(my_sum, "my_sum", [("x", np.float64)], np.float64)

c.sql("SELECT MY_SUM(x) FROM dask").compute()

Machine learning in SQL

Dask-SQL has support for both model training and prediction, enabling machine learning workflows with a flexible combination of both Python and SQL. A model can be registered in a Context either through the register_model method or a CREATE MODEL statement.

In [ ]:
from dask_ml.linear_model import LinearRegression
from sklearn.ensemble import GradientBoostingClassifier

# create a dask-ml model and train it
model = GradientBoostingClassifier()
data = c.sql("SELECT x, y, x * y > 0 AS target FROM dask LIMIT 50")
model.fit(data[["x", "y"]], data["target"])

# register this model in the context
c.register_model("python_model", model, training_columns=["x", "y"])

# create and train a model directly from SQL
    CREATE MODEL sql_model WITH (
        model_class = 'sklearn.ensemble.GradientBoostingClassifier',
        wrap_predict = True,
        target_column = 'target'
    ) AS (
        SELECT x, y, x * y > 0 AS target
        FROM dask
        LIMIT 50

Registered models must follow the scikit-learn interface by implementing a predict method. As with tables, all registered models can be listed with a SHOW MODEL statement.

In [ ]:
c.sql("SHOW MODELS").compute()

From here, the models can be used to make predictions using the PREDICT keyword as part of a SELECT query.

In [ ]:
        MODEL sql_model,
        SELECT x, y, x * y > 0 AS actual FROM dask
        OFFSET 50
In [ ]: