Unable to query by partition column

[ChewingGlass]

Environment

Max
Delta-rs version: latest(deltalake = { git = "https://github.com/delta-io/delta-rs", branch = "main", features = ["s3", "datafusion"] })

Binding:

Environment: OsX

• Cloud provider: AWS
• OS: Mac
• Other:

---

Bug

What happened:

I create a table with a "date" partition column,

SchemaField::new(
            "date".to_string(),
            SchemaDataType::primitive("date".to_string()),
            false,
            HashMap::new(),
        );

DeltaOps(table)
                .create()
                .with_columns(delta_schema)
                .with_partition_columns(vec!["date"])
                .await?

If I run any query that either selects date or filters by date, I get:

Error: External error: External error: Arrow error: External error: Arrow error: Invalid argument error: column types must match schema types, expected Dictionary(UInt16, Date32) but found Dictionary(UInt16, Null) at column index 1

Caused by:
    0: External error: Arrow error: External error: Arrow error: Invalid argument error: column types must match schema types, expected Dictionary(UInt16, Date32) but found Dictionary(UInt16, Null) at column index 1
    1: Arrow error: External error: Arrow error: Invalid argument error: column types must match schema types, expected Dictionary(UInt16, Date32) but found Dictionary(UInt16, Null) at column index 1
    2: External error: Arrow error: Invalid argument error: column types must match schema types, expected Dictionary(UInt16, Date32) but found Dictionary(UInt16, Null) at column index 1
    3: Arrow error: Invalid argument error: column types must match schema types, expected Dictionary(UInt16, Date32) but found Dictionary(UInt16, Null) at column index 1
    4: Invalid argument error: column types must match schema types, expected Dictionary(UInt16, Date32) but found Dictionary(UInt16, Null) at column index 1

Removing the partition from the table and querying no longer results in an error, but querying on the date field is not working as expected (any query on the field results in no data).

What you expected to happen:
I expect the queries to run and filter by date

How to reproduce it:
See above

[cmackenzie1]

Thanks for the report! I put up a draft PR demonstrating the issue as a test case.

[ChewingGlass]

I also noticed that with no partition, in order to get date filters to work you have to do arrow_cast(date, 'Utf8') >= arrow_cast('{}', 'Utf8')

[Blajda]

I played around with this. I think the issue is that Datafusion / arrow cannot cast a Date32 into a Dictionary(Int32, Date32) and it silently fails. If you explicitly try to do this it will return an error and mention it is not supported.

When I prevent Date32 from being dictionary encoded here my tests don't error out.

   #[tokio::test]
    async fn test_date() -> Result<(), Box<dyn Error>>{
        let ctx = SessionContext::new();
        let schema: Schema = serde_json::from_value(json!({
            "type": "struct",
            "fields": [
                {"name": "id", "type": "string", "nullable": true, "metadata": {}},
                {"name": "my_date", "type": "date", "nullable": true, "metadata": {}},
            ]
        }))
        .unwrap();
        let table = DeltaOps::new_in_memory()
            .create()
            .with_save_mode(SaveMode::ErrorIfExists)
            .with_columns(schema.get_fields().clone())
            .with_partition_columns(["my_date"])
            .await
            .unwrap();
        assert_eq!(table.version(), 0);

        let data = ctx.sql("select 1 as id, now() as my_date").await?.collect().await?;

        let table = DeltaOps(table)
            .write(data)
            .await
            .unwrap();
        
        ctx.register_table("test", Arc::new(table)).unwrap();
        println!("test1");
        let data = ctx.sql("select * from test where my_date <=  arrow_cast(now(), 'Date32')").await?;
        data.show().await?;

        println!("test2");
        let data = ctx.sql("select * from test where my_date > '2023-06-07'").await?;
        data.show().await?;

        println!("test3");
        ctx.table("test").await?
            .filter(col("my_date").gt(lit("2023-06-05")))?
            .show().await?;

        Ok(())
    }

[cmackenzie1]

Looking at a few of my tables, using the type string for partition values avoids this issue as well.

SchemaField::new(
            "date".to_string(),
            SchemaDataType::primitive("string".to_string()),
            false,
            HashMap::new(),
        );

[wjones127]

I think the issue is that Datafusion / arrow cannot cast a Date32 into a Dictionary(Int32, Date32)

That's very odd. Arrow should be able to cast any type into it's dictionary form.
https://github.com/apache/arrow-rs/blob/ab56693985826bb8caea30558b8c25db286a5e37/arrow-cast/src/cast.rs#LL127C16-L127C16
https://github.com/apache/arrow-rs/blob/ab56693985826bb8caea30558b8c25db286a5e37/arrow-cast/src/cast.rs#L216

If it's the only quick fix we can find, we can remove the dictionary encoding on partition columns. But we should add it back when we can, since it's much more performant for those to be dictionary.

[Blajda]

This is the query I used and the resulting error

let data = ctx.sql("select * from test where my_date <=  arrow_cast(now(), 'Dictionary(UInt16, Date32)')").await?;

Error

Error: Context("Optimizer rule 'simplify_expressions' failed", ArrowError(CastError("Unsupported output type for dictionary packing: Date32")))

which traces to here
https://github.com/apache/arrow-rs/blob/ab56693985826bb8caea30558b8c25db286a5e37/arrow-cast/src/cast.rs#L3572

[wjones127]

Reported this upstream: apache/arrow-rs#4390

Sounds like we should remove the dictionary encoding for now then. And we can try bringing it back once we fix that upstream issue.

[tustvold]

remove the dictionary encoding

FWIW Dictionary<Int32, Date32> will be larger, slower to process, and less well supported than Date32. I don't see a reason to ever dictionary encode primitives

[wjones127]

Ah okay. If Arrow won't support that well, then let's not plan on using dictionary columns for small values.