Remove LargeUtf8|Binary, Utf8|BinaryView, and Dictionary from ScalarValue #11978

notfilippo · 2024-08-14T10:00:01Z

This PR removes LargeUtf8|Binary, Utf8|BinaryView, and Dictionary from ScalarValue, following the discussion on #11513

Open questions

Top level ScalarValue cast

This change initially failed the select_arrow_cast test (see review comments for updates). The test fails because of an interaction of expression_simplifier + optimize_projections.

The query is: SELECT arrow_cast(1234, 'Float64') as f64, arrow_cast('foo', 'LargeUtf8') as large

expression_simplifier correctly tries to evaluate_to_scalar the arrow_cast('foo', 'LargeUtf8') cast (here)
this leads to the cast physical expression (here), but this is a no-op since it transforms Utf8('foo') (scalar) → LargeUtf8('foo') (array) → Utf8('foo') (scalar)
- TODO: if this is what we want this should probably not execute at all
This edit doesn't change the underlying schema of the LogicalPlan
optimize_projections rewrites the Projection and updates the schema, seeing the Utf8('foo') it correctly assumes that the LogicalPlan's schema field for this value should have DataType == Utf8

This check is the one raising this error but I guess it should instead check if schema fields are logically equivalent to eachother. I'm not totally convinced this is the correct solution because it removes some guarantees that might be expected by users downstream. Happy to hear everyone's opinion on this.

---- optimizer::select_arrow_cast stdout ----
thread 'optimizer::select_arrow_cast' panicked at datafusion/core/tests/optimizer/mod.rs:118:30:
called `Result::unwrap()` on an `Err` value: Context("Optimizer rule 'optimize_projections' failed", Context("optimize_projections", Internal("Failed due to a difference in schemas, original schema: DFSchema { inner: Schema { fields: [Field { name: \"f64\", data_type: Float64, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }, Field { name: \"large\", data_type: LargeUtf8, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }], metadata: {} }, field_qualifiers: [None, None], functional_dependencies: FunctionalDependencies { deps: [] } }, new schema: DFSchema { inner: Schema { fields: [Field { name: \"f64\", data_type: Float64, nullable: false, dict_id: 0, dict_is_ordered: false, metadata: {} }, Field { name: \"large\", data_type: Utf8, nullable: false, dict_id: 0, dict_is_ordered: false, metadata: {} }], metadata: {} }, field_qualifiers: [None, None], functional_dependencies: FunctionalDependencies { deps: [] } }")))
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

`arrow_typeof`

https://github.com/apache/datafusion/blob/main/datafusion/functions/src/core/arrowtypeof.rs#L59-L72 uses column_value.data_type() to determine the type of the argument but this information is not really accurate. If the ColumnValue is a ScalarValue the data_type() will be "logical". e.g. arrow_typeof(arrow_cast('hello', 'Utf8View')) would yield Utf8.

Type info

Let's take this expr as an example a_function_that_takes_utf8_view(arrow_cast('test', 'Utf8View'))
the cast expression currently evaluates to a ColumnValue::Scalar(Utf8View("test")) and the function is happy to receive that. With this change the cast expression instead evaluates to ColumnValue::Scalar(Utf8("test")) (as ScalarValue::Utf8View doesn't exist it produces a logically equal value) and the cast expression data_type() returns Utf8View.

What would the function do?
How can we communicate to the parent expression in the expression tree that the column value needs to be of that particular type? (it should know because of the schema)

…alue

notfilippo · 2024-08-14T11:13:47Z

datafusion/core/src/datasource/physical_plan/file_scan_config.rs

-    }
-
+fn create_output_array(val: &ScalarValue, len: usize) -> Result<ArrayRef> {
+    // TODO(@notfilippo): should we reintroduce a way to encode as dictionaries?


I'm missing some context for partitions so I'm not really sure if it's the right call to remove the Dictionary output array. IIUC it seems like it's very useful to save space and optimise columns added this way.

notfilippo · 2024-08-14T16:42:30Z

datafusion/sqllogictest/test_files/string_view.slt

-02)--Filter: test.column1_utf8view = Utf8View("Andrew")
+02)--Filter: test.column1_utf8view = CAST(Utf8("Andrew") AS Utf8View)


Casts are not simplified to retain the type information to fix the optimize_projection issue above.

datafusion/common/src/scalar/mod.rs

notfilippo · 2024-08-14T17:04:56Z

datafusion/optimizer/src/simplify_expressions/expr_simplifier.rs

        let col_val = match phys_expr.evaluate(&self.input_batch) {
            Ok(v) => v,
            Err(err) => return ConstSimplifyResult::SimplifyRuntimeError(err, expr),
        };
+
+        // TODO(@notfilippo): a fix for the select_arrow_cast error


This check partially fixes the issue with select_arrow_cast.

datafusion/physical-plan/src/projection.rs

datafusion/common/src/scalar/mod.rs

datafusion/optimizer/src/simplify_expressions/expr_simplifier.rs

datafusion/common/src/logical/eq.rs

datafusion/core/src/datasource/physical_plan/file_scan_config.rs

jayzhan211 · 2024-08-15T06:14:02Z

statement ok
create table test_source as values
    ('Andrew', 'X'),
    ('Xiangpeng', 'Xiangpeng'),
    ('Raphael', 'R'),
    (NULL, 'R')
;

statement ok
create table t as
SELECT
  arrow_cast(column1, 'Utf8') as column1_utf8,
  arrow_cast(column2, 'Utf8') as column2_utf8,
  arrow_cast(column1, 'LargeUtf8') as column1_large_utf8,
  arrow_cast(column2, 'LargeUtf8') as column2_large_utf8,
  arrow_cast(column1, 'Utf8View') as column1_utf8view,
  arrow_cast(column2, 'Utf8View') as column2_utf8view,
  arrow_cast(column1, 'Dictionary(Int32, Utf8)') as column1_dict,
  arrow_cast(column2, 'Dictionary(Int32, Utf8)') as column2_dict
FROM test_source;

query error DataFusion error: Arrow error: Invalid argument error: column types must match schema types, expected Utf8View but found Utf8 at column index 0
select min(column1_utf8view) from t;

I run this test and fail. It success on main branch.

I guess we need schema for converting ScalarValue to ArrayRef

Given that we have ScalarValue::Utf8, and we have StringView in schema. We can then get the corresponding StrginViewArray.

notfilippo · 2024-08-16T13:21:42Z

Given that we have ScalarValue::Utf8, and we have StringView in schema. We can then get the corresponding StrginViewArray.

I feel like this is a little bit more complex because casting a scalar value to a logically equal type (arrow_cast('test', 'Dictionary(Int32, Utf8)')) is a no-op for all cases but one the cast is at the top level, which indicates that you actually want an array of that type. So we need a way to distinguish those two scenarios, which is why I've added this:

datafusion/datafusion/physical-plan/src/projection.rs

Line 312 in 78dc034

.and_then(|v| v.into_array_of_type(batch.num_rows(), field.data_type()))

But I'm not really sure about it...

datafusion/expr-common/src/columnar_value.rs

jayzhan211 · 2024-08-16T14:32:22Z

Since you have no commit yet, so CI is not running.

You could run these command to pass CI
cargo test --test sqllogictests (slt files)
./dev/rust_lint.sh
cargo test --lib --tests --bins --features avro,json (rust test)

notfilippo · 2024-08-16T14:35:54Z

Since you have no commit yet, so CI is not running.

You could run these command to pass CI cargo test --test sqllogictests (slt files) ./dev/rust_lint.sh cargo test --lib --tests --bins --features avro,json (rust test)

Thanks! Don't worry about turning it on as I can still get results from my fork.

notfilippo · 2024-08-19T21:47:06Z

I've pushed a fairly big experiment. I've tried to change ColumnarValue to

#[derive(Clone, Debug)]
pub enum ColumnarValue {
    /// Array of values
    Array(ArrayRef),
    /// A single value
    Scalar(ScalarValue),
    Scalar(Scalar),
}

#[derive(Clone, Debug)]
pub struct Scalar {
    value: ScalarValue,
    data_type: DataType,
}

which follows the approache that was discussed with @jayzhan211 in the comments above. I've opted for this hybrid solution to retain most of the flexibility of the original ColumnarValue and I'm mostly satisfied with how it turned out. Curious to hear your thoughts @jayzhan211 and @alamb

jayzhan211 · 2024-08-20T02:59:43Z

It seems the trick you did is to get the first index of ArrayRef (instead of keeping it as arrow::Scalar) as ScalarValue but we still ends up require DataType to keep the type information.

However, I think we could move on with this approach, we could figure out if there is better approach later on

datafusion/common/src/logical/mod.rs

notfilippo · 2024-08-20T09:25:51Z

I'm happy to report that I've got most sqllogictests to run successfully (albeit there is still the issue pointed out by @alamb, which i plan to address after I've got all tests passing). The only errors I'm seeing are the following:

Aggregates using ScalarValues as state

Aggregates use ScalarValue to represent state and evaluate to a result. Should I look into restricting their return type to the subset of types which can be represented by ScalarValues?

External error: query failed: DataFusion error: Arrow error: Invalid argument error: column types must match schema types, expected LargeUtf8 but found Utf8 at column index 3
[SQL] SELECT
  min(utf8),
  max(utf8),
  count(utf8),
  min(largeutf8),
  max(largeutf8),
  count(largeutf8)
FROM t
at test_files/aggregate.slt:4104

Weird error I don't quite understand

Removing this query doesn't yield any other error in the slt file. I don't have any other clue and I'm not sure where to start 🤷 .

External error: query result mismatch:
[SQL] SELECT
  count,
  LAG(timestamp, 1) OVER (ORDER BY timestamp),
  arrow_typeof(LAG(timestamp, 1) OVER (ORDER BY timestamp))
FROM timestamp_with_tz
LIMIT 10;
[Diff] (-expected|+actual)
    0 NULL Timestamp(Millisecond, Some("UTC"))
    0 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
    0 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
-   4 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
    0 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
-   0 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
+   12 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
+   2 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
    0 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
    14 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
-   0 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
-   0 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
+   5 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
+   1 2014-08-27T14:00:00Z Timestamp(Millisecond, Some("UTC"))
at test_files/parquet.slt:255

jayzhan211 · 2024-08-20T13:13:56Z

Aggregates using ScalarValues as state

The issue is that we have only ScalarValue but no DataType to get the actual ArrayRef back.
One of the solution is to store DataType in MaxAccumulator and call to_array_of_type to return ArrayRef as the output of evaluate, maybe evaluate_as_array to avoid breaking existing function.

Weird error I don't quite understand

I fail to reproduce the error 😕

UIpd: It seems there is only one place that calls Accumulator::evaluate and transform it to ArrayRef, I think we can just change the return value to ArrayRef for Accumulator::evaluate

datafusion/datafusion/physical-plan/src/aggregates/mod.rs

Line 1093 in 37e54ee

.map(|accumulator| accumulator.evaluate().and_then(|v| v.to_array()))

notfilippo · 2024-08-23T09:04:04Z

UIpd: It seems there is only one place that calls Accumulator::evaluate and transform it to ArrayRef, I think we can just change the return value to ArrayRef for Accumulator::evaluate

It seems like this solution is not that easy as the state needs to also be accounted for, but it's definitely a good start!

notfilippo · 2024-08-23T09:37:39Z

Marking as "ready for review" because all tests pass on my end. Some casting issues still remain that I'll look into soon but in the meantime I'm looking forward to some feedback on this huge change ❤️

jayzhan211 · 2024-08-24T05:57:03Z

Will evaluate_as_scalar replaces evaluate and state_as_scalars replaces state? If it is, then it looks good to me

notfilippo · 2024-08-27T13:19:48Z

Will evaluate_as_scalar replaces evaluate and state_as_scalars replaces state? If it is, then it looks good to me

Yes that's the plan!

notfilippo · 2024-09-16T17:35:37Z

I'm back from vacation and I've rebased my PR to the latest upstream.

jayzhan211 · 2024-09-18T06:11:59Z

datafusion/expr-common/src/columnar_value.rs

+pub struct Scalar {
+    value: ScalarValue,
+    data_type: DataType,
+}


I believe this is the main change of this PR

I was thinking that I could open a new PR introducing just introducing this columnar value type, and then little by little removing the scalar value variants (Dictionary, Large* and *View...)

I'm mentioning this because of the discussion in #12488 (comment), as I think it will be very beneficial to split this PR in multiple changes, both for reviewers and datafusion users.

I agree with @findepi 's suggestion to make the smallest PR as possible to begin. I think splitting this PR out into a small part to begin, along with comments explaining the planned end state (or linking back to the ticket that describes this) would be very helpful

alamb · 2024-09-21T11:21:51Z

To make it clear this PR is not waiting on more review (we are discussing on #12536) marking this as draft

notfilippo · 2024-09-30T20:02:54Z

Closing in favour of #12622

Remove LargeUtf8|Binary, Utf8|BinaryView, and Dictionary from ScalarV…

6cec428

…alue

github-actions bot added sql SQL Planner physical-expr Physical Expressions optimizer Optimizer rules core Core DataFusion crate substrait common Related to common crate proto Related to proto crate functions labels Aug 14, 2024

notfilippo mentioned this pull request Aug 14, 2024

[Proposal] Decouple logical from physical types #11513

Open

This comment was marked as resolved.

Sign in to view

notfilippo commented Aug 14, 2024

View reviewed changes

Progress sync for string_view.slt

78dc034

github-actions bot added the sqllogictest SQL Logic Tests (.slt) label Aug 14, 2024

notfilippo commented Aug 14, 2024

View reviewed changes

jayzhan211 reviewed Aug 15, 2024

View reviewed changes

datafusion/common/src/scalar/mod.rs Outdated Show resolved Hide resolved

datafusion/common/src/scalar/mod.rs Outdated Show resolved Hide resolved

jayzhan211 reviewed Aug 15, 2024

View reviewed changes

datafusion/optimizer/src/simplify_expressions/expr_simplifier.rs Outdated Show resolved Hide resolved

jayzhan211 reviewed Aug 15, 2024

View reviewed changes

datafusion/common/src/logical/eq.rs Outdated Show resolved Hide resolved

jayzhan211 reviewed Aug 15, 2024

View reviewed changes

datafusion/core/src/datasource/physical_plan/file_scan_config.rs Show resolved Hide resolved

notfilippo added 4 commits August 16, 2024 14:36

Reintroduce Dictionary in partition values

1b22fa9

Merge remote-tracking branch 'origin/main' into fr/logical-scalar-types

6e50de9

Add peekable

58ce569

Fix impl IntoIterator with Peekable

349a04e

jayzhan211 reviewed Aug 16, 2024

View reviewed changes

datafusion/expr-common/src/columnar_value.rs Outdated Show resolved Hide resolved

Run rust_lint

301f031

github-actions bot added the logical-expr Logical plan and expressions label Aug 19, 2024

jayzhan211 reviewed Aug 20, 2024

View reviewed changes

datafusion/common/src/logical/mod.rs Outdated Show resolved Hide resolved

notfilippo added 2 commits August 20, 2024 10:58

Make expr.slt pass

2578bf9

Apply suggestions and lint

058553e

notfilippo added 2 commits August 20, 2024 11:33

Merge remote-tracking branch 'origin/main' into fr/logical-scalar-types

cb79770

Updates due to merge

4ad4965

Fix aggregates

9308d0b

notfilippo marked this pull request as ready for review August 23, 2024 09:36

This was referenced Sep 9, 2024

DataFusion weekly project plan (Andrew Lamb) - Sep 9, 2024 #12391

Closed

DataFusion weekly project plan (Andrew Lamb) - Sep 2, 2024 #12336

Closed

notfilippo mentioned this pull request Sep 16, 2024

Remove ScalarValue::Dictionary #12488

Closed

notfilippo added 2 commits September 16, 2024 17:59

Merge remote-tracking branch 'origin/main' into fr/logical-scalar-types

9ed1d3d

Resolve conflicts

f158e17

alamb mentioned this pull request Sep 16, 2024

DataFusion weekly project plan (Andrew Lamb) - Sep 16, 2024 #12494

Closed

8 tasks

jayzhan211 reviewed Sep 18, 2024

View reviewed changes

Merge remote-tracking branch 'origin/main' into fr/logical-scalar-types

a4bba8f

notfilippo mentioned this pull request Sep 19, 2024

Introduce Scalar type for ColumnarValue #12536

Merged

1 task

alamb marked this pull request as draft September 21, 2024 11:21

notfilippo closed this Sep 30, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Remove LargeUtf8|Binary, Utf8|BinaryView, and Dictionary from ScalarValue #11978

Remove LargeUtf8|Binary, Utf8|BinaryView, and Dictionary from ScalarValue #11978

notfilippo commented Aug 14, 2024 •

edited

Loading

This comment was marked as resolved.

notfilippo Aug 14, 2024

notfilippo Aug 14, 2024

notfilippo Aug 14, 2024

jayzhan211 commented Aug 15, 2024 •

edited

Loading

notfilippo commented Aug 16, 2024

jayzhan211 commented Aug 16, 2024

notfilippo commented Aug 16, 2024

notfilippo commented Aug 19, 2024

jayzhan211 commented Aug 20, 2024 •

edited

Loading

notfilippo commented Aug 20, 2024 •

edited

Loading

jayzhan211 commented Aug 20, 2024 •

edited

Loading

notfilippo commented Aug 23, 2024

notfilippo commented Aug 23, 2024 •

edited

Loading

jayzhan211 commented Aug 24, 2024

notfilippo commented Aug 27, 2024

notfilippo commented Sep 16, 2024

jayzhan211 Sep 18, 2024 •

edited

Loading

notfilippo Sep 18, 2024 •

edited

Loading

alamb Sep 18, 2024

alamb commented Sep 21, 2024

notfilippo commented Sep 30, 2024

		02)--Filter: test.column1_utf8view = Utf8View("Andrew")
		02)--Filter: test.column1_utf8view = CAST(Utf8("Andrew") AS Utf8View)

Remove LargeUtf8|Binary, Utf8|BinaryView, and Dictionary from ScalarValue #11978

Remove LargeUtf8|Binary, Utf8|BinaryView, and Dictionary from ScalarValue #11978

Conversation

notfilippo commented Aug 14, 2024 • edited Loading

Open questions

Top level ScalarValue cast

arrow_typeof

Type info

This comment was marked as resolved.

notfilippo Aug 14, 2024

Choose a reason for hiding this comment

notfilippo Aug 14, 2024

Choose a reason for hiding this comment

notfilippo Aug 14, 2024

Choose a reason for hiding this comment

jayzhan211 commented Aug 15, 2024 • edited Loading

notfilippo commented Aug 16, 2024

jayzhan211 commented Aug 16, 2024

notfilippo commented Aug 16, 2024

notfilippo commented Aug 19, 2024

jayzhan211 commented Aug 20, 2024 • edited Loading

notfilippo commented Aug 20, 2024 • edited Loading

Aggregates using ScalarValues as state

Weird error I don't quite understand

jayzhan211 commented Aug 20, 2024 • edited Loading

notfilippo commented Aug 23, 2024

notfilippo commented Aug 23, 2024 • edited Loading

jayzhan211 commented Aug 24, 2024

notfilippo commented Aug 27, 2024

notfilippo commented Sep 16, 2024

jayzhan211 Sep 18, 2024 • edited Loading

Choose a reason for hiding this comment

notfilippo Sep 18, 2024 • edited Loading

Choose a reason for hiding this comment

alamb Sep 18, 2024

Choose a reason for hiding this comment

alamb commented Sep 21, 2024

notfilippo commented Sep 30, 2024

notfilippo commented Aug 14, 2024 •

edited

Loading

`arrow_typeof`

jayzhan211 commented Aug 15, 2024 •

edited

Loading

jayzhan211 commented Aug 20, 2024 •

edited

Loading

notfilippo commented Aug 20, 2024 •

edited

Loading

jayzhan211 commented Aug 20, 2024 •

edited

Loading

notfilippo commented Aug 23, 2024 •

edited

Loading

jayzhan211 Sep 18, 2024 •

edited

Loading

notfilippo Sep 18, 2024 •

edited

Loading