SQLGlot is a no-dependency SQL parser, transpiler, optimizer, and engine. It can be used to format SQL or translate between 23 different dialects like DuckDB, Presto / Trino, Spark / Databricks, Snowflake, and BigQuery. It aims to read a wide variety of SQL inputs and output syntactically and semantically correct SQL in the targeted dialects.

It is a very comprehensive generic SQL parser with a robust test suite. It is also quite performant, while being written purely in Python.

You can easily customize the parser, analyze queries, traverse expression trees, and programmatically build SQL.

Syntax errors are highlighted and dialect incompatibilities can warn or raise depending on configurations. However, SQLGlot does not aim to be a SQL validator, so it may fail to detect certain syntax errors.

Learn more about SQLGlot in the API documentation and the expression tree primer.

Contributions are very welcome in SQLGlot; read the contribution guide and the onboarding document to get started!

• Install
• Versioning
• Get in Touch
• FAQ
• Examples
  • Formatting and Transpiling
  • Metadata
  • Parser Errors
  • Unsupported Errors
  • Build and Modify SQL
  • SQL Optimizer
  • AST Introspection
  • AST Diff
  • Custom Dialects
  • SQL Execution
• Used By
• Documentation
• Run Tests and Lint
• Benchmarks
• Optional Dependencies

From PyPI:

pip3 install "sqlglot[rs]"

# Without Rust tokenizer (slower):
# pip3 install sqlglot

Or with a local checkout:

make install

Requirements for development (optional):

make install-dev

Given a version number MAJOR.MINOR.PATCH, SQLGlot uses the following versioning strategy:

• The PATCH version is incremented when there are backwards-compatible fixes or feature additions.
• The MINOR version is incremented when there are backwards-incompatible fixes or feature additions.
• The MAJOR version is incremented when there are significant backwards-incompatible fixes or feature additions.

We'd love to hear from you. Join our community Slack channel!

I tried to parse SQL that should be valid but it failed, why did that happen?

• Most of the time, issues like this occur because the "source" dialect is omitted during parsing. For example, this is how to correctly parse a SQL query written in Spark SQL: parse_one(sql, dialect="spark") (alternatively: read="spark"). If no dialect is specified, parse_one will attempt to parse the query according to the "SQLGlot dialect", which is designed to be a superset of all supported dialects. If you tried specifying the dialect and it still doesn't work, please file an issue.

I tried to output SQL but it's not in the correct dialect!

• Like parsing, generating SQL also requires the target dialect to be specified, otherwise the SQLGlot dialect will be used by default. For example, to transpile a query from Spark SQL to DuckDB, do parse_one(sql, dialect="spark").sql(dialect="duckdb") (alternatively: transpile(sql, read="spark", write="duckdb")).

I tried to parse invalid SQL and it worked, even though it should raise an error! Why didn't it validate my SQL?

• SQLGlot does not aim to be a SQL validator - it is designed to be very forgiving. This makes the codebase more comprehensive and also gives more flexibility to its users, e.g. by allowing them to include trailing commas in their projection lists.

What happened to sqlglot.dataframe?

• The PySpark dataframe api was moved to a standalone library called SQLFrame in v24. It now allows you to run queries as opposed to just generate SQL.

Easily translate from one dialect to another. For example, date/time functions vary between dialects and can be hard to deal with:

import sqlglot
sqlglot.transpile("SELECT EPOCH_MS(1618088028295)", read="duckdb", write="hive")[0]

'SELECT FROM_UNIXTIME(1618088028295 / POW(10, 3))'

SQLGlot can even translate custom time formats:

import sqlglot
sqlglot.transpile("SELECT STRFTIME(x, '%y-%-m-%S')", read="duckdb", write="hive")[0]

"SELECT DATE_FORMAT(x, 'yy-M-ss')"

Identifier delimiters and data types can be translated as well:

import sqlglot

# Spark SQL requires backticks (`) for delimited identifiers and uses `FLOAT` over `REAL`
sql = """WITH baz AS (SELECT a, c FROM foo WHERE a = 1) SELECT f.a, b.b, baz.c, CAST("b"."a" AS REAL) d FROM foo f JOIN bar b ON f.a = b.a LEFT JOIN baz ON f.a = baz.a"""

# Translates the query into Spark SQL, formats it, and delimits all of its identifiers
print(sqlglot.transpile(sql, write="spark", identify=True, pretty=True)[0])

WITH `baz` AS (
  SELECT
    `a`,
    `c`
  FROM `foo`
  WHERE
    `a` = 1
)
SELECT
  `f`.`a`,
  `b`.`b`,
  `baz`.`c`,
  CAST(`b`.`a` AS FLOAT) AS `d`
FROM `foo` AS `f`
JOIN `bar` AS `b`
  ON `f`.`a` = `b`.`a`
LEFT JOIN `baz`
  ON `f`.`a` = `baz`.`a`

Comments are also preserved on a best-effort basis:

sql = """
/* multi
   line
   comment
*/
SELECT
  tbl.cola /* comment 1 */ + tbl.colb /* comment 2 */,
  CAST(x AS SIGNED), # comment 3
  y               -- comment 4
FROM
  bar /* comment 5 */,
  tbl #          comment 6
"""

# Note: MySQL-specific comments (`#`) are converted into standard syntax
print(sqlglot.transpile(sql, read='mysql', pretty=True)[0])

/* multi
   line
   comment
*/
SELECT
  tbl.cola /* comment 1 */ + tbl.colb /* comment 2 */,
  CAST(x AS INT), /* comment 3 */
  y /* comment 4 */
FROM bar /* comment 5 */, tbl /*          comment 6 */

You can explore SQL with expression helpers to do things like find columns and tables in a query:

from sqlglot import parse_one, exp

# print all column references (a and b)
for column in parse_one("SELECT a, b + 1 AS c FROM d").find_all(exp.Column):
    print(column.alias_or_name)

# find all projections in select statements (a and c)
for select in parse_one("SELECT a, b + 1 AS c FROM d").find_all(exp.Select):
    for projection in select.expressions:
        print(projection.alias_or_name)

# find all tables (x, y, z)
for table in parse_one("SELECT * FROM x JOIN y JOIN z").find_all(exp.Table):
    print(table.name)

Read the ast primer to learn more about SQLGlot's internals.

When the parser detects an error in the syntax, it raises a ParseError:

import sqlglot
sqlglot.transpile("SELECT foo FROM (SELECT baz FROM t")

sqlglot.errors.ParseError: Expecting ). Line 1, Col: 34.
  SELECT foo FROM (SELECT baz FROM t
                                   ~

Structured syntax errors are accessible for programmatic use:

import sqlglot
try:
    sqlglot.transpile("SELECT foo FROM (SELECT baz FROM t")
except sqlglot.errors.ParseError as e:
    print(e.errors)

[{
  'description': 'Expecting )',
  'line': 1,
  'col': 34,
  'start_context': 'SELECT foo FROM (SELECT baz FROM ',
  'highlight': 't',
  'end_context': '',
  'into_expression': None
}]

It may not be possible to translate some queries between certain dialects. For these cases, SQLGlot may emit a warning and will proceed to do a best-effort translation by default:

import sqlglot
sqlglot.transpile("SELECT APPROX_DISTINCT(a, 0.1) FROM foo", read="presto", write="hive")

APPROX_COUNT_DISTINCT does not support accuracy
'SELECT APPROX_COUNT_DISTINCT(a) FROM foo'

This behavior can be changed by setting the unsupported_level attribute. For example, we can set it to either RAISE or IMMEDIATE to ensure an exception is raised instead:

import sqlglot
sqlglot.transpile("SELECT APPROX_DISTINCT(a, 0.1) FROM foo", read="presto", write="hive", unsupported_level=sqlglot.ErrorLevel.RAISE)

sqlglot.errors.UnsupportedError: APPROX_COUNT_DISTINCT does not support accuracy

There are queries that require additional information to be accurately transpiled, such as the schemas of the tables referenced in them. This is because certain transformations are type-sensitive, meaning that type inference is needed in order to understand their semantics. Even though the qualify and annotate_types optimizer rules can help with this, they are not used by default because they add significant overhead and complexity.

Transpilation is generally a hard problem, so SQLGlot employs an "incremental" approach to solving it. This means that there may be dialect pairs that currently lack support for some inputs, but this is expected to improve over time. We highly appreciate well-documented and tested issues or PRs, so feel free to reach out if you need guidance!

SQLGlot supports incrementally building SQL expressions:

from sqlglot import select, condition

where = condition("x=1").and_("y=1")
select("*").from_("y").where(where).sql()

'SELECT * FROM y WHERE x = 1 AND y = 1'

It's possible to modify a parsed tree:

from sqlglot import parse_one
parse_one("SELECT x FROM y").from_("z").sql()

'SELECT x FROM z'

Parsed expressions can also be transformed recursively by applying a mapping function to each node in the tree:

from sqlglot import exp, parse_one

expression_tree = parse_one("SELECT a FROM x")

def transformer(node):
    if isinstance(node, exp.Column) and node.name == "a":
        return parse_one("FUN(a)")
    return node

transformed_tree = expression_tree.transform(transformer)
transformed_tree.sql()

'SELECT FUN(a) FROM x'

SQLGlot can rewrite queries into an "optimized" form. It performs a variety of techniques to create a new canonical AST. This AST can be used to standardize queries or provide the foundations for implementing an actual engine. For example:

import sqlglot
from sqlglot.optimizer import optimize

print(
    optimize(
        sqlglot.parse_one("""
            SELECT A OR (B OR (C AND D))
            FROM x
            WHERE Z = date '2021-01-01' + INTERVAL '1' month OR 1 = 0
        """),
        schema={"x": {"A": "INT", "B": "INT", "C": "INT", "D": "INT", "Z": "STRING"}}
    ).sql(pretty=True)
)

SELECT
  (
    "x"."a" <> 0 OR "x"."b" <> 0 OR "x"."c" <> 0
  )
  AND (
    "x"."a" <> 0 OR "x"."b" <> 0 OR "x"."d" <> 0
  ) AS "_col_0"
FROM "x" AS "x"
WHERE
  CAST("x"."z" AS DATE) = CAST('2021-02-01' AS DATE)

You can see the AST version of the parsed SQL by calling repr:

from sqlglot import parse_one
print(repr(parse_one("SELECT a + 1 AS z")))

Select(
  expressions=[
    Alias(
      this=Add(
        this=Column(
          this=Identifier(this=a, quoted=False)),
        expression=Literal(this=1, is_string=False)),
      alias=Identifier(this=z, quoted=False))])

SQLGlot can calculate the semantic difference between two expressions and output changes in a form of a sequence of actions needed to transform a source expression into a target one:

from sqlglot import diff, parse_one
diff(parse_one("SELECT a + b, c, d"), parse_one("SELECT c, a - b, d"))

[
  Remove(expression=Add(
    this=Column(
      this=Identifier(this=a, quoted=False)),
    expression=Column(
      this=Identifier(this=b, quoted=False)))),
  Insert(expression=Sub(
    this=Column(
      this=Identifier(this=a, quoted=False)),
    expression=Column(
      this=Identifier(this=b, quoted=False)))),
  Keep(
    source=Column(this=Identifier(this=a, quoted=False)),
    target=Column(this=Identifier(this=a, quoted=False))),
  ...
]

See also: Semantic Diff for SQL.

Dialects can be added by subclassing Dialect:

from sqlglot import exp
from sqlglot.dialects.dialect import Dialect
from sqlglot.generator import Generator
from sqlglot.tokens import Tokenizer, TokenType


class Custom(Dialect):
    class Tokenizer(Tokenizer):
        QUOTES = ["'", '"']
        IDENTIFIERS = ["`"]

        KEYWORDS = {
            **Tokenizer.KEYWORDS,
            "INT64": TokenType.BIGINT,
            "FLOAT64": TokenType.DOUBLE,
        }

    class Generator(Generator):
        TRANSFORMS = {exp.Array: lambda self, e: f"[{self.expressions(e)}]"}

        TYPE_MAPPING = {
            exp.DataType.Type.TINYINT: "INT64",
            exp.DataType.Type.SMALLINT: "INT64",
            exp.DataType.Type.INT: "INT64",
            exp.DataType.Type.BIGINT: "INT64",
            exp.DataType.Type.DECIMAL: "NUMERIC",
            exp.DataType.Type.FLOAT: "FLOAT64",
            exp.DataType.Type.DOUBLE: "FLOAT64",
            exp.DataType.Type.BOOLEAN: "BOOL",
            exp.DataType.Type.TEXT: "STRING",
        }

print(Dialect["custom"])

<class '__main__.Custom'>

SQLGlot is able to interpret SQL queries, where the tables are represented as Python dictionaries. The engine is not supposed to be fast, but it can be useful for unit testing and running SQL natively across Python objects. Additionally, the foundation can be easily integrated with fast compute kernels, such as Arrow and Pandas.

The example below showcases the execution of a query that involves aggregations and joins:

from sqlglot.executor import execute

tables = {
    "sushi": [
        {"id": 1, "price": 1.0},
        {"id": 2, "price": 2.0},
        {"id": 3, "price": 3.0},
    ],
    "order_items": [
        {"sushi_id": 1, "order_id": 1},
        {"sushi_id": 1, "order_id": 1},
        {"sushi_id": 2, "order_id": 1},
        {"sushi_id": 3, "order_id": 2},
    ],
    "orders": [
        {"id": 1, "user_id": 1},
        {"id": 2, "user_id": 2},
    ],
}

execute(
    """
    SELECT
      o.user_id,
      SUM(s.price) AS price
    FROM orders o
    JOIN order_items i
      ON o.id = i.order_id
    JOIN sushi s
      ON i.sushi_id = s.id
    GROUP BY o.user_id
    """,
    tables=tables
)

user_id price
      1   4.0
      2   3.0

See also: Writing a Python SQL engine from scratch.

• SQLMesh
• Apache Superset
• Dagster
• Fugue
• ibis
• mysql-mimic
• Querybook
• Quokka
• Splink
• SQLFrame

SQLGlot uses pdoc to serve its API documentation.

A hosted version is on the SQLGlot website, or you can build locally with:

make docs-serve

make style  # Only linter checks
make unit   # Only unit tests (or unit-rs, to use the Rust tokenizer)
make test   # Unit and integration tests (or test-rs, to use the Rust tokenizer)
make check  # Full test suite & linter checks

Benchmarks run on Python 3.10.12 in seconds.

SQLGlot uses dateutil to simplify literal timedelta expressions. The optimizer will not simplify expressions like the following if the module cannot be found:

x + interval '1' month