inbq

A library for parsing BigQuery queries and extracting schema-aware, column-level lineage.

Features

• Parse BigQuery queries into well-structured ASTs with easy-to-navigate nodes.
• Extract schema-aware, column-level lineage.
• Trace data flow through nested structs and arrays.
• Capture referenced columns and the specific query components (e.g., select, where, join) they appear in.
• Process both single and multi-statement queries with procedural language constructs.
• Built for speed and efficiency, with lightweight Python bindings that add minimal overhead.

Python

Install

pip install inbq

Example (Pipeline API)

import inbq

catalog = {"schema_objects": []}

def add_table(name: str, columns: list[tuple[str, str]]) -> None:
    catalog["schema_objects"].append({
        "name": name,
        "kind": {
            "table": {
                "columns": [{"name": name, "dtype": dtype} for name, dtype in columns]
            }
        }
    })

add_table("project.dataset.out", [("id", "int64"), ("val", "float64")])
add_table("project.dataset.t1", [("id", "int64"), ("x", "float64")])
add_table("project.dataset.t2", [("id", "int64"), ("s", "struct<source string, x float64>")])

query = """
declare default_val float64  default (select min(val) from project.dataset.out);

insert into `project.dataset.out`
select
    id,
    if(x is null or s.x is null, default_val, x + s.x)
from `project.dataset.t1` inner join `project.dataset.t2` using (id)
where s.source = "baz";
"""

pipeline = (
    inbq.Pipeline()
    .config(
        # If the `pipeline` is configured with `raise_exception_on_error=False`,
        # any error that occurs during parsing or lineage extraction is
        # captured and returned as a `inbq.PipelineError`
        raise_exception_on_error=False,
        # No effect with only one query (may provide a speedup with multiple queries)
        parallel=True,
    )
    .parse()
    .extract_lineage(catalog=catalog, include_raw=False)
)
sqls = [query]
pipeline_output = inbq.run_pipeline(sqls, pipeline=pipeline)

# This loop will iterate just once as we have only one query
for i, (ast, output_lineage) in enumerate(
    zip(pipeline_output.asts, pipeline_output.lineages)
):
    assert isinstance(ast, inbq.ast_nodes.Ast), (
        f"Could not parse query `{sqls[i][:20]}...` due to: {ast.error}"
    )

    print(f"{ast=}")

    assert isinstance(output_lineage, inbq.lineage.Lineage), (
        f"Could not extract lineage from query `{sqls[i][:20]}...` due to: {output_lineage.error}"
    )

    print("\nLineage:")
    for lin_obj in output_lineage.lineage.objects:
        print("Inputs:")
        for lin_node in lin_obj.nodes:
            print(
                f"{lin_obj.name}->{lin_node.name} <- {[f'{input_node.obj_name}->{input_node.node_name}' for input_node in lin_node.inputs]}"
            )

        print("\nSide inputs:")
        for lin_node in lin_obj.nodes:
            print(
                f"""{lin_obj.name}->{lin_node.name} <- {[f"{input_node.obj_name}->{input_node.node_name} @ {','.join(input_node.sides)}" for input_node in lin_node.side_inputs]}"""
            )

    print("\nReferenced columns:")
    for ref_obj in output_lineage.referenced_columns.objects:
        for ref_node in ref_obj.nodes:
            print(
                f"{ref_obj.name}->{ref_node.name} referenced in {ref_node.referenced_in}"
            )

# Prints:
# ast=Ast(...)

# Lineage:
# Inputs:
# project.dataset.out->id <- ['project.dataset.t2->id', 'project.dataset.t1->id']
# project.dataset.out->val <- ['project.dataset.t2->s.x', 'project.dataset.t1->x', 'project.dataset.out->val']
#
# Side inputs:
# project.dataset.out->id <- ['project.dataset.t2->s.source @ where', 'project.dataset.t2->id @ join', 'project.dataset.t1->id @ join']
# project.dataset.out->val <- ['project.dataset.t2->s.source @ where', 'project.dataset.t2->id @ join', 'project.dataset.t1->id @ join']
#
# Referenced columns:
# project.dataset.out->val referenced in ['default_var', 'select']
# project.dataset.t1->id referenced in ['join', 'select']
# project.dataset.t1->x referenced in ['select']
# project.dataset.t2->id referenced in ['join', 'select']
# project.dataset.t2->s.x referenced in ['select']
# project.dataset.t2->s.source referenced in ['where']

Note: What happens if you remove the insert and just keep the select in the query? inbq is designed to handle this gracefully. It will return the lineage for the last SELECT statement, but since the destination is no longer explicit, the output object (an anonymous query) will be assigned an anonymous identifier (e.g., !anon_4). Try it yourself and see how the output changes!

To learn more about the output elements (Lineage, Side Inputs, and Referenced Columns), please see the Concepts section.

Example (Individual Functions)

If you don't like the Pipeline API, you can use these functions instead:

parse_sql and parse_sql_to_dict

Parse a single SQL query:

ast = inbq.parse_sql(query)

# You can also get a dictionary representation of the AST
ast_dict = inbq.parse_sql_to_dict(query)

parse_sqls

Parse multiple SQL queries in parallel:

sqls = [query]
asts = inbq.parse_sqls(sqls, parallel=True)

parse_sqls_and_extract_lineage

Parse SQLs and extract lineage in one go:

asts, lineages = inbq.parse_sqls_and_extract_lineage(
    sqls=[query],
    catalog=catalog,
    parallel=True
)

AST Navigation

import inbq
import inbq.ast_nodes as ast_nodes

sql = """
UPDATE proj.dataset.t1
SET quantity = quantity - 10,
    supply_constrained = DEFAULT
WHERE product like '%washer%';

UPDATE proj.dataset.t2
SET quantity = quantity - 10,
WHERE product like '%console%';
"""

ast = inbq.parse_sql(sql)

# Example: find updated tables and columns
for node in ast.find_all(
    ast_nodes.UpdateStatement,
):
    match node:
        case ast_nodes.UpdateStatement(
            table=table,
            alias=_,
            update_items=update_items,
            from_=_,
            where=_,
        ):
            print(f"Found updated table: {table.name}. Updated columns:")
            for update_item in update_items:
                for node in update_item.column.find_all(
                    ast_nodes.Identifier,
                    ast_nodes.QuotedIdentifier
                ):
                    match node:
                        case ast_nodes.Identifier(name=name) | ast_nodes.QuotedIdentifier(name=name):
                            print(f"- {name}")

# Example: find `like` filters
for node in ast.find_all(
    ast_nodes.BinaryExpr,
):
    match node:
        case ast_nodes.BinaryExpr(
            left=left,
            operator=ast_nodes.BinaryOperator_Like(),
            right=right,
        ):
            print(left, "like", right)

Variants and Variant Types in Python

The AST nodes in Python are auto-generated dataclasses from their Rust definitions. For instance, a Rust enum Expr might be defined as:

pub enum Expr {
    // ... more variants here ...
    Binary(BinaryExpr),
    Identifier(Identifier),
    // ... more variants here ...
}

In Python, this translates to corresponding classes like Expr_Binary(vty=BinaryExpr), Expr_Identifier(vty=Identifier), etc. The vty attribute stands for "variant type" (unit variants do not have a vty attribute). You can search for any type of object using .find_all(), whether it's the variant (e.g., Expr_Identifier) or the concrete variant type (e.g., Identifier).

Rust

Install

cargo add inbq

Example

use inbq::{
    lineage::{
        catalog::{Catalog, Column, SchemaObject, SchemaObjectKind},
        extract_lineage,
    },
    parser::Parser,
    scanner::Scanner,
};

fn column(name: &str, dtype: &str) -> Column {
    Column {
        name: name.to_owned(),
        dtype: dtype.to_owned(),
    }
}

fn main() -> anyhow::Result<()> {
    env_logger::init();

    let sql = r#"
        declare default_val float64 default (select min(val) from project.dataset.out);

        insert into `project.dataset.out`
        select
            id,
            if(x is null or s.x is null, default_val, x + s.x)
        from `project.dataset.t1` inner join `project.dataset.t2` using (id)
        where s.source = "baz";
    "#;
    let mut scanner = Scanner::new(sql);
    scanner.scan()?;
    let mut parser = Parser::new(scanner.tokens());
    let ast = parser.parse()?;
    println!("Syntax Tree: {:?}", ast);

    let data_catalog = Catalog {
        schema_objects: vec![
            SchemaObject {
                name: "project.dataset.out".to_owned(),
                kind: SchemaObjectKind::Table {
                    columns: vec![column("id", "int64"), column("val", "int64")],
                },
            },
            SchemaObject {
                name: "project.dataset.t1".to_owned(),
                kind: SchemaObjectKind::Table {
                    columns: vec![column("id", "int64"), column("x", "float64")],
                },
            },
            SchemaObject {
                name: "project.dataset.t2".to_owned(),
                kind: SchemaObjectKind::Table {
                    columns: vec![
                        column("id", "int64"),
                        column("s", "struct<source string, x float64>"),
                    ],
                },
            },
        ],
    };

    let lineage = extract_lineage(&[&ast], &data_catalog, false, true)
        .pop()
        .unwrap()?;

    println!("\nLineage: {:?}", lineage.lineage);
    println!("\nReferenced columns: {:?}", lineage.referenced_columns);
    Ok(())
}

Command Line Interface

Install binary

cargo install inbq

Extract Lineage

1. Prepare your data catalog: create a JSON file (e.g., catalog.json) that defines the schema for all tables and views referenced in your SQL queries.

2. Run inbq: pass the catalog file and your SQL file or directory of multiple SQL files to the inbq lineage command.

inbq extract-lineage \
    --pretty \
    --catalog ./examples/lineage/catalog.json  \
    ./examples/lineage/query.sql

The output is written to stdout.

Concepts

Lineage

Column-level lineage tracks how data flows from a destination column back to its original source columns. A destination column's value is derived from its direct input columns, and this process is applied recursively to trace the lineage back to the foundational source columns. For example, in with tmp as (select a+b as tmp_c from t) select tmp_c as c from t, the lineage for column c traces back to a and b as its source columns (the source table is t).

Lineage - Side Inputs

Side inputs are columns that indirectly contribute to the final set of output values. As the name implies, they aren't part of the direct SELECT list, but are found in the surrounding clauses that shape the result, such as WHERE, JOIN, WINDOW, etc. Side inputs influence is traced recursively. For example, in the query:

with cte as (select id, c1 from table1 where f1>10)
select c2 as z
from table2 inner join cte using (id)

table1.f1 is a side input to z with sides join and where (cte.id, later used in the join condition, is filtered by table1.f1). The other two side inputs are table1.id with side join and table2.id with side join.

Referenced Columns

Referenced columns provide a detailed map of where each input column is mentioned within a query. This is the entry point for a column into the query's logic. From this initial reference, the column can then influence other parts of the query indirectly through subsequent operations.

Limitations

While this library can parse and extract lineage for most BigQuery syntax, there are some current limitations. For example, the pipe (|) syntax and the recently introduced MATCH_RECOGNIZE clause are not yet supported. Requests and contributions for unsupported features are welcome.

Contributing

Here's a brief overview of the project's key modules:

• crates/inbq/src/parser.rs: contains the hand-written top-down parser.
• crates/inbq/src/ast.rs: defines the Abstract Syntax Tree (AST) nodes.
  • Note: If you add or modify AST nodes here, you must regenerate the corresponding Python nodes. You can do this by running cargo run --bin inbq_genpy, which will update crates/py_inbq/python/inbq/ast_nodes.py.
• crates/inbq/src/lineage.rs: contains the core logic for extracting column-level lineage from the AST.
• crates/py_inbq/: this crate exposes the Rust backend as a Python module via PyO3.
• crates/inbq/tests/: this directory contains the tests. You can add new test cases for parsing and lineage extraction by editing the .toml files:
  • parsing_tests.toml
  • lineage_tests.toml