Edit on GitHub

PySpark DataFrame SQL Generator

This is a drop-in replacement for the PySpark DataFrame API that will generate SQL instead of executing DataFrame operations directly. This, when combined with the transpiling support in SQLGlot, allows one to write PySpark DataFrame code and execute it on other engines like DuckDB, Presto, Spark, Snowflake, and BigQuery.

Currently many of the common operations are covered and more functionality will be added over time. Please open an issue or PR with your feedback or contribution to help influence what should be prioritized next and make sure your use case is properly supported.

How to use

Instructions

  • Install SQLGlot and that is all that is required to just generate SQL. The examples show generating SQL and then executing that SQL on a specific engine and that will require that engine's client library.
  • Find/replace all from pyspark.sql with from sqlglot.dataframe.
  • Prior to any spark.read.table or spark.table run sqlglot.schema.add_table('<table_name>', <column_structure>, dialect="spark").
    • The column structure can be defined the following ways:
      • Dictionary where the keys are column names and values are string of the Spark SQL type name.
        • Ex: {'cola': 'string', 'colb': 'int'}
      • PySpark DataFrame StructType similar to when using createDataFrame.
        • Ex: StructType([StructField('cola', StringType()), StructField('colb', IntegerType())])
      • A string of names and types similar to what is supported in createDataFrame.
        • Ex: cola: STRING, colb: INT
      • [Not Recommended] A list of string column names without type.
        • Ex: ['cola', 'colb']
        • The lack of types may limit functionality in future releases.
    • See Registering Custom Schema for information on how to skip this step if the information is stored externally.
  • If your output SQL dialect is not Spark, then configure the SparkSession to use that dialect
    • Ex: SparkSession().builder.config("sqlframe.dialect", "bigquery").getOrCreate()
    • See dialects for a full list of dialects.
  • Add sqlglot.dataframe.sql(pretty=True) to your final DataFrame command to return a list of sql statements to run that command.
    • In most cases a single SQL statement is returned. Currently the only exception is when caching DataFrames which isn't supported in other dialects.
    • Ex: sqlglot.dataframe.sql(pretty=True)

Examples

import sqlglot
from sqlglot.dataframe.sql.session import SparkSession
from sqlglot.dataframe.sql import functions as F

dialect = "spark"

sqlglot.schema.add_table(
  'employee',
  {
    'employee_id': 'INT',
    'fname': 'STRING',
    'lname': 'STRING',
    'age': 'INT',
  },
  dialect=dialect,
)  # Register the table structure prior to reading from the table

spark = SparkSession.builder.config("sqlframe.dialect", dialect).getOrCreate()

df = (
    spark
    .table('employee')
    .groupBy(F.col("age"))
    .agg(F.countDistinct(F.col("employee_id")).alias("num_employees")) 
)

print(df.sql(pretty=True))

SELECT
  `employee`.`age` AS `age`,
  COUNT(DISTINCT `employee`.`employee_id`) AS `num_employees`
FROM `employee` AS `employee`
GROUP BY
  `employee`.`age`

Registering Custom Schema Class

The step of adding sqlglot.schema.add_table can be skipped if you have the column structure stored externally like in a file or from an external metadata table. This can be done by writing a class that implements the sqlglot.schema.Schema abstract class and then assigning that class to sqlglot.schema. 

import sqlglot
from sqlglot.dataframe.sql.session import SparkSession
from sqlglot.dataframe.sql import functions as F
from sqlglot.schema import Schema


class ExternalSchema(Schema):
  ...

sqlglot.schema = ExternalSchema()

spark = SparkSession()  # Spark will be used by default is not specific in SparkSession config

df = (
    spark
    .table('employee')
    .groupBy(F.col("age"))
    .agg(F.countDistinct(F.col("employee_id")).alias("num_employees")) 
)

print(df.sql(pretty=True))

Example Implementations

Bigquery

from google.cloud import bigquery
from sqlglot.dataframe.sql.session import SparkSession
from sqlglot.dataframe.sql import types
from sqlglot.dataframe.sql import functions as F

client = bigquery.Client()

data = [
    (1, "Jack", "Shephard", 34),
    (2, "John", "Locke", 48),
    (3, "Kate", "Austen", 34),
    (4, "Claire", "Littleton", 22),
    (5, "Hugo", "Reyes", 26),
]
schema = types.StructType([
    types.StructField('employee_id', types.IntegerType(), False),
    types.StructField('fname', types.StringType(), False),
    types.StructField('lname', types.StringType(), False),
    types.StructField('age', types.IntegerType(), False),
])

sql_statements = (
    SparkSession
    .builder
    .config("sqlframe.dialect", "bigquery")
    .getOrCreate()
    .createDataFrame(data, schema)
    .groupBy(F.col("age"))
    .agg(F.countDistinct(F.col("employee_id")).alias("num_employees"))
    .sql()
)

result = None
for sql in sql_statements:
  result = client.query(sql)

assert result is not None
for row in client.query(result):
    print(f"Age: {row['age']}, Num Employees: {row['num_employees']}")

Snowflake

import os

import snowflake.connector
from sqlglot.dataframe.session import SparkSession
from sqlglot.dataframe import types
from sqlglot.dataframe import functions as F

ctx = snowflake.connector.connect(
    user=os.environ["SNOWFLAKE_USER"],
    password=os.environ["SNOWFLAKE_PASS"],
    account=os.environ["SNOWFLAKE_ACCOUNT"]
)
cs = ctx.cursor()

data = [
    (1, "Jack", "Shephard", 34),
    (2, "John", "Locke", 48),
    (3, "Kate", "Austen", 34),
    (4, "Claire", "Littleton", 22),
    (5, "Hugo", "Reyes", 26),
]
schema = types.StructType([
    types.StructField('employee_id', types.IntegerType(), False),
    types.StructField('fname', types.StringType(), False),
    types.StructField('lname', types.StringType(), False),
    types.StructField('age', types.IntegerType(), False),
])

sql_statements = (
    SparkSession
    .builder
    .config("sqlframe.dialect", "snowflake")
    .getOrCreate()
    .createDataFrame(data, schema)
    .groupBy(F.col("age"))
    .agg(F.countDistinct(F.col("lname")).alias("num_employees"))
    .sql()
)

try:
    for sql in sql_statements:
        cs.execute(sql)
    results = cs.fetchall()
    for row in results:
        print(f"Age: {row[0]}, Num Employees: {row[1]}")
finally:
    cs.close()
ctx.close()

Spark

from pyspark.sql.session import SparkSession as PySparkSession
from sqlglot.dataframe.sql.session import SparkSession
from sqlglot.dataframe.sql import types
from sqlglot.dataframe.sql import functions as F

data = [
    (1, "Jack", "Shephard", 34),
    (2, "John", "Locke", 48),
    (3, "Kate", "Austen", 34),
    (4, "Claire", "Littleton", 22),
    (5, "Hugo", "Reyes", 26),
]
schema = types.StructType([
    types.StructField('employee_id', types.IntegerType(), False),
    types.StructField('fname', types.StringType(), False),
    types.StructField('lname', types.StringType(), False),
    types.StructField('age', types.IntegerType(), False),
])

sql_statements = (
    SparkSession()
    .createDataFrame(data, schema)
    .groupBy(F.col("age"))
    .agg(F.countDistinct(F.col("employee_id")).alias("num_employees"))
    .sql()
)

pyspark = PySparkSession.builder.master("local[*]").getOrCreate()

df = None
for sql in sql_statements:
    df = pyspark.sql(sql)

assert df is not None
df.show()

Unsupportable Operations

Any operation that lacks a way to represent it in SQL cannot be supported by this tool. An example of this would be rdd operations. Since the DataFrame API though is mostly modeled around SQL concepts most operations can be supported.

1"""
2.. include:: ./README.md
3"""