Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Provide documentation of expose APIs to enable handling of type coercion at UNION plan construction. #12142

Merged
merged 3 commits into from
Aug 26, 2024
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 11 additions & 1 deletion datafusion/expr/src/logical_plan/builder.rs
Original file line number Diff line number Diff line change
Expand Up @@ -1331,7 +1331,17 @@ pub fn validate_unique_names<'a>(
})
}

/// Union two logical plans.
/// Union two [`LogicalPlan`]s.
///
/// Constructs the UNION plan, but does not perform type-coercion. Therefore the
/// subtree expressions will not be properly typed until the optimizer pass.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

hm, does that mean if there is no optimizer phase the query will crash?

Copy link
Contributor Author

@wiedld wiedld Aug 26, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think maybe this depends on how users construct their own logical plans? And if they are already ensuring type correctness in their own code?

Our logical plan construction relied upon the previous behavior of the union() api, which was immediate type coercion. IMO - this was not an explicit api contract, rather an implementation detail. Regardless, we use the typed-coerced union schema & expressions when adding other logical nodes (e.g. limit, sort, group by). Once the behavior of the api changed, we started building incorrect plans -- which produced incorrect results and in one case errored.

I filed this ticket as a doc enhancement, not a bug, since I didn't think(?) type coercion was part of the api contract. My hope was to (a) make it clear when/how unions can be type coerced, and (b) make public the APIs to do so. In other words, make it easier for users to ensure type correctness before (or without) the optimizer pass.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should I have phrased the docs differently? Or is there another approach we should take?

Copy link
Contributor Author

@wiedld wiedld Aug 26, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also, to clarify. The "incorrect plans" is because the new union() behavior doesn't type-coerce the expressions and it takes the left node's schema. When we built logical plans with union + gap filling (adding casted scalar nulls), we started having missing fields etc when inspecting the union before constructing the next node.

Maybe the latter should be a bug?

///
/// If a properly typed UNION plan is needed, refer to [`TypeCoercionRewriter::coerce_union`]
/// or alternatively, merge the union input schema using [`coerce_union_schema`] and
/// apply the expression rewrite with [`coerce_plan_expr_for_schema`].
///
/// [`TypeCoercionRewriter::coerce_union`]: https://docs.rs/datafusion-optimizer/latest/datafusion_optimizer/analyzer/type_coercion/struct.TypeCoercionRewriter.html#method.coerce_union
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

👍

/// [`coerce_union_schema`]: https://docs.rs/datafusion-optimizer/latest/datafusion_optimizer/analyzer/type_coercion/fn.coerce_union_schema.html
pub fn union(left_plan: LogicalPlan, right_plan: LogicalPlan) -> Result<LogicalPlan> {
// Temporarily use the schema from the left input and later rely on the analyzer to
// coerce the two schemas into a common one.
Expand Down
22 changes: 17 additions & 5 deletions datafusion/optimizer/src/analyzer/type_coercion.rs
Original file line number Diff line number Diff line change
Expand Up @@ -56,6 +56,8 @@ use datafusion_expr::{
Projection, ScalarUDF, Union, WindowFrame, WindowFrameBound, WindowFrameUnits,
};

/// Performs type coercion by determining the schema
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

❤️

/// and performing the expression rewrites.
#[derive(Default)]
pub struct TypeCoercion {}

Expand Down Expand Up @@ -128,16 +130,23 @@ fn analyze_internal(
.map_data(|plan| plan.recompute_schema())
}

pub(crate) struct TypeCoercionRewriter<'a> {
/// Rewrite expressions to apply type coercion.
pub struct TypeCoercionRewriter<'a> {
pub(crate) schema: &'a DFSchema,
}

impl<'a> TypeCoercionRewriter<'a> {
/// Create a new [`TypeCoercionRewriter`] with a provided schema
/// representing both the inputs and output of the [`LogicalPlan`] node.
fn new(schema: &'a DFSchema) -> Self {
Self { schema }
}

fn coerce_plan(&mut self, plan: LogicalPlan) -> Result<LogicalPlan> {
/// Coerce the [`LogicalPlan`].
///
/// Refer to [`TypeCoercionRewriter::coerce_join`] and [`TypeCoercionRewriter::coerce_union`]
/// for type-coercion approach.
pub fn coerce_plan(&mut self, plan: LogicalPlan) -> Result<LogicalPlan> {
match plan {
LogicalPlan::Join(join) => self.coerce_join(join),
LogicalPlan::Union(union) => Self::coerce_union(union),
Expand All @@ -153,7 +162,7 @@ impl<'a> TypeCoercionRewriter<'a> {
///
/// For example, on_exprs like `t1.a = t2.b AND t1.x = t2.y` will be stored
/// as a list of `(t1.a, t2.b), (t1.x, t2.y)`
fn coerce_join(&mut self, mut join: Join) -> Result<LogicalPlan> {
pub fn coerce_join(&mut self, mut join: Join) -> Result<LogicalPlan> {
join.on = join
.on
.into_iter()
Expand All @@ -176,7 +185,7 @@ impl<'a> TypeCoercionRewriter<'a> {

/// Coerce the union’s inputs to a common schema compatible with all inputs.
/// This occurs after wildcard expansion and the coercion of the input expressions.
fn coerce_union(union_plan: Union) -> Result<LogicalPlan> {
pub fn coerce_union(union_plan: Union) -> Result<LogicalPlan> {
let union_schema = Arc::new(coerce_union_schema(&union_plan.inputs)?);
let new_inputs = union_plan
.inputs
Expand Down Expand Up @@ -809,7 +818,10 @@ fn coerce_case_expression(case: Case, schema: &DFSchema) -> Result<Case> {
}

/// Get a common schema that is compatible with all inputs of UNION.
fn coerce_union_schema(inputs: &[Arc<LogicalPlan>]) -> Result<DFSchema> {
///
/// This method presumes that the wildcard expansion is unneeded, or has already
/// been applied.
pub fn coerce_union_schema(inputs: &[Arc<LogicalPlan>]) -> Result<DFSchema> {
let base_schema = inputs[0].schema();
let mut union_datatypes = base_schema
.fields()
Expand Down