Extract parquet statistics for `StructArray` #11289

Lordworms · 2024-07-05T17:39:28Z

Which issue does this PR close?

Closes #10609

Rationale for this change

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

Lordworms · 2024-07-05T17:43:11Z

@alamb I'll try to add support for struct in DataPage in next PR.

efredine · 2024-07-12T14:31:50Z

datafusion/core/src/datasource/physical_plan/parquet/statistics.rs

-        // For example a ListArray could correspond to anything from 1 to 3 levels
-        // in the parquet schema
-        return None;
+        match field.data_type() {


Perhaps we should extract the repeated code into a helper function? Could also reduce nesting with an early return for the case where it's not nested. So something like:

Suggested change

match field.data_type() {

pub(crate) fn parquet_column<'a>(

parquet_schema: &SchemaDescriptor,

arrow_schema: &'a Schema,

name: &str,

) -> Option<(usize, &'a FieldRef)> {

let (root_idx, field) = arrow_schema.fields.find(name)?;

if !field.data_type().is_nested() {

let parquet_idx = find_parquet_idx(parquet_schema, root_idx)?;

return Some((parquet_idx, field));

}

// Nested field

match field.data_type() {

DataType::Struct(_) => {

let parquet_idx = find_parquet_idx(parquet_schema, root_idx)?;

Some((parquet_idx, field))

}

_ => {

if field.data_type().is_nested() {

// Nested fields are not supported and require non-trivial logic

// to correctly walk the parquet schema accounting for the

// logical type rules - <https://github.com/apache/parquet-format/blob/master/LogicalTypes.md>

//

// For example a ListArray could correspond to anything from 1 to 3 levels

// in the parquet schema

None

} else {

let parquet_idx = find_parquet_idx(parquet_schema, root_idx)?;

Some((parquet_idx, field))

}

}

}

}

// Helper function to find parquet index - TBD: this could be more efficient

fn find_parquet_idx(parquet_schema: &SchemaDescriptor, root_idx: usize) -> Option<usize> {

(0..parquet_schema.columns().len())

.find(|x| parquet_schema.get_column_root_idx(*x) == root_idx)

}

Sure, I can do it

efredine · 2024-07-12T14:39:56Z

datafusion/core/tests/parquet/arrow_statistics.rs

+            Arc::new(BooleanArray::from(vec![Some(true)])) as ArrayRef,
+        ),
+    ]);
+


I wondered if we needed a struct array with top-level nulls. But for the stats, it shouldn't matter whether the null is coming from the top level or the nested arrays. That would be testing the how the stats are calculated. (This assumes I'm interpreting the way nulls are handled correctly.)

But it might be worth having some null counts to see that they get calculated correctly?

That's what I concern about. Since it is hard and seems meaningless to have a top-level nullcount. The reason I leave it like this is just to follow the test pattern.

This blog post provides good examples on when you'd represent nulls at different levels: https://arrow.apache.org/blog/2022/10/08/arrow-parquet-encoding-part-2/

I was going to read the valid mask but it actually contains in the data page, not sure whether we need to read the actual data at this time since it is a Metadata level analysis.

I don't think you need to do anything differently. I was trying to find some tests in the arrow crate for when the statistics are written but I haven't been able to find any for writing nested structs.

Here's some tests in the arrow crate that are writing a 2 level struct: https://github.com/apache/arrow-rs/blob/master/parquet/src/arrow/arrow_writer/mod.rs#L1516-L1618 Not sure if that's helpful or not.

Sure, I'll take a look at it.

efredine · 2024-07-13T22:18:55Z

datafusion/core/src/datasource/physical_plan/parquet/statistics.rs

+        metadata: &[&RowGroupMetaData],
+        index: usize,
+        data_type: &DataType,
+    ) -> Vec<u64> {


If you preferred, this could also be expressed as a fold:

let num_row_groups = metadata.len(); fields.iter().fold(vec![0; num_row_groups], |mut acc, field| { let field_null_counts = Self::get_null_counts_recursive( metadata, index + 1, field.data_type(), ); acc.iter_mut().zip(field_null_counts.iter()).for_each(|(a, b)| *a += b); acc })

Sure, sorry for the late response

alamb · 2024-07-18T19:48:41Z

Since we have merged apache/arrow-rs#6046 now upstream in arrow-rs, @Lordworms would you be willing to port this PR to the other repo?

Lordworms · 2024-07-19T01:22:11Z

Since we have merged apache/arrow-rs#6046 now upstream in arrow-rs, @Lordworms would you be willing to port this PR to the other repo?

Sure, I'll open a PR later today.

Lordworms · 2024-07-19T01:53:11Z

filed apache/arrow-rs#6090

alamb · 2024-07-19T19:29:42Z

Thank you @Lordworms - sorry for the delay / runaround. I just haven't had a chance to focus on this PR. I was hoping someone with more structured type experience would be able to help review it. 🎣

alamb · 2024-07-24T18:40:25Z

superceded by apache/arrow-rs#6090

(which I know is waiting for a review -- I just need to find time to study the nested types or find someone else who does)

github-actions bot added the core Core DataFusion crate label Jul 5, 2024

alamb changed the title ~~extract statistics read for struct array in parquet~~ Extract parquet statistics for StructArray Jul 7, 2024

This was referenced Jul 7, 2024

DataFusion weekly project plan (Andrew Lamb) - July 1, 2024 #11190

Closed

DataFusion weekly project plan (Andrew Lamb) - July 8, 2024 #11334

Closed

efredine reviewed Jul 12, 2024

View reviewed changes

extract statistics read for struct array in parquet

15cd7c5

Lordworms force-pushed the issue_10609_main branch from c1a3e62 to 15cd7c5 Compare July 12, 2024 22:39

alamb mentioned this pull request Jul 13, 2024

Add parquet StatisticsConverter for arrow reader apache/arrow-rs#6046

Merged

efredine reviewed Jul 13, 2024

View reviewed changes

alamb mentioned this pull request Jul 15, 2024

DataFusion weekly project plan (Andrew Lamb) - July 15, 2024 #11474

Closed

7 tasks

refactor code

f2e0a79

alamb mentioned this pull request Jul 22, 2024

DataFusion weekly project plan (Andrew Lamb) - July 22, 2024 #11601

Closed

5 tasks

alamb closed this Jul 24, 2024

This pull request was closed.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Extract parquet statistics for `StructArray` #11289

Extract parquet statistics for `StructArray` #11289

Lordworms commented Jul 5, 2024

Lordworms commented Jul 5, 2024

efredine Jul 12, 2024

Lordworms Jul 12, 2024

efredine Jul 12, 2024

Lordworms Jul 12, 2024

efredine Jul 12, 2024

Lordworms Jul 12, 2024 •

edited

Loading

efredine Jul 12, 2024

efredine Jul 12, 2024

Lordworms Jul 13, 2024

efredine Jul 13, 2024

Lordworms Jul 15, 2024

alamb commented Jul 18, 2024

Lordworms commented Jul 19, 2024

Lordworms commented Jul 19, 2024

alamb commented Jul 19, 2024

alamb commented Jul 24, 2024

-        match field.data_type() {
+pub(crate) fn parquet_column<'a>(
+    parquet_schema: &SchemaDescriptor,
+    arrow_schema: &'a Schema,
+    name: &str,
+) -> Option<(usize, &'a FieldRef)> {
+    let (root_idx, field) = arrow_schema.fields.find(name)?;
+    if !field.data_type().is_nested() {
+        let parquet_idx = find_parquet_idx(parquet_schema, root_idx)?;
+        return Some((parquet_idx, field));
+    }
+    // Nested field
+    match field.data_type() {
+        DataType::Struct(_) => {
+            let parquet_idx = find_parquet_idx(parquet_schema, root_idx)?;
+            Some((parquet_idx, field))
+        }
+        _ => {
+            if field.data_type().is_nested() {
+                // Nested fields are not supported and require non-trivial logic
+                // to correctly walk the parquet schema accounting for the
+                // logical type rules - <https://github.com/apache/parquet-format/blob/master/LogicalTypes.md>
+                //
+                // For example a ListArray could correspond to anything from 1 to 3 levels
+                // in the parquet schema
+                None
+            } else {
+                let parquet_idx = find_parquet_idx(parquet_schema, root_idx)?;
+                Some((parquet_idx, field))
+            }
+        }
+    }
+}
+// Helper function to find parquet index - TBD: this could be more efficient
+fn find_parquet_idx(parquet_schema: &SchemaDescriptor, root_idx: usize) -> Option<usize> {
+    (0..parquet_schema.columns().len())
+        .find(|x| parquet_schema.get_column_root_idx(*x) == root_idx)
+}

Extract parquet statistics for StructArray #11289

Extract parquet statistics for StructArray #11289

Conversation

Lordworms commented Jul 5, 2024

Which issue does this PR close?

Rationale for this change

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

Lordworms commented Jul 5, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Lordworms Jul 12, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

alamb commented Jul 18, 2024

Lordworms commented Jul 19, 2024

Lordworms commented Jul 19, 2024

alamb commented Jul 19, 2024

alamb commented Jul 24, 2024

Extract parquet statistics for `StructArray` #11289

Extract parquet statistics for `StructArray` #11289

Lordworms Jul 12, 2024 •

edited

Loading