fixed size list type is not retained when writing to parquet

[matko]

When I create a parquet file from an arrow table with a fixed size array as one of the columns, then read back the resulting parquet, the column is no longer a fixed size array, but instead a dynamically sized array.

Example:

import datafusion as df
import pyarrow as pa

FILENAME = "/tmp/fixed_array_example.parquet"
ctx = df.SessionContext()

array = pa.array([[1.0, 2.0], [3.0, 4.0]], type=pa.list_(pa.float32(), 2))
table = pa.Table.from_pydict({"array": array})
df_table = ctx.from_arrow(table)
print("original schema:")
print(df_table.schema())

df_table.write_parquet(FILENAME)
print("roundtrip schema:")
print(ctx.read_parquet(FILENAME).schema())

Output:

original schema:
array: fixed_size_list<item: float>[2]
  child 0, item: float
roundtrip schema:
array: list<item: float>
  child 0, item: float

As the output demonstrates, the datafusion dataframe that is written out has the proper schema. Nevertheless, the file that is read back does not.

If instead of datafusion, I use pyarrow to write the parquet file, I do get the expected schema when I read it back using datafusion.

import datafusion as df
import pyarrow as pa
import pyarrow.parquet as pq

FILENAME = "/tmp/fixed_array_example_pyarrow.parquet"
ctx = df.SessionContext()

array = pa.array([[1.0, 2.0], [3.0, 4.0]], type=pa.list_(pa.float32(), 2))
table = pa.Table.from_pydict({"array": array})

print("original schema:")
print(table.schema)

pq.write_table(table, FILENAME)
print("roundtrip schema:")
print(ctx.read_parquet(FILENAME).schema())

output:

original schema:
array: fixed_size_list<item: float>[2]
  child 0, item: float
roundtrip schema:
array: fixed_size_list<element: float>[2]
  child 0, element: float

[kosiew]

Research note:
currently FixedSizeList is stored as List in Parquet

[kylebarron]

I think you want to set skip_metadata=False

datafusion-python/python/datafusion/context.py

Line 717 in 79c22d6

skip_metadata: bool = True,

In my opinion that should be the default, otherwise no Arrow-only types will be retained I think.

[kosiew]

hi @kylebarron ,

I amended skip_metadata: bool = False and ran

import datafusion as df
import pyarrow as pa

FILENAME = "/tmp/fixed_array_example.parquet"
ctx = df.SessionContext()

array = pa.array([[1.0, 2.0], [3.0, 4.0]], type=pa.list_(pa.float32(), 2))
table = pa.Table.from_pydict({"array": array})
df_table = ctx.from_arrow(table)
print("original schema:")
print(df_table.schema())

df_table.write_parquet(FILENAME)

ctx.register_parquet("test_fixed_list", FILENAME)

print("register parquet schema:")
print(ctx.table("test_fixed_list").schema())

but still the fixed size list type was still retained:

original schema:
array: fixed_size_list<item: float>[2]
  child 0, item: float
==> register_parquet with skip_metadata:  False
register parquet schema:
array: list<item: float>
  child 0, item: float

[kylebarron]

In this case it's an issue with writing the Parquet file, not reading, which you can see if you try to read the file back with pyarrow:

In [23]: import pyarrow.parquet as pq

In [24]: pq.read_schema(FILENAME)
Out[24]:
array: list<item: float>
  child 0, item: float

In this case it's actually because the writing side doesn't correctly propagate the Arrow metadata either.

Here's how pyarrow.parquet correctly propagates the Arrow schema within the Parquet metadata:

In [32]: pq.write_table(table, "test.parquet")

In [33]: meta2 = pq.read_metadata('test.parquet')

In [34]: meta2.metadata
Out[34]: {b'ARROW:schema': b'/////6gAAAAQAAAAAAAKAAwABgAFAAgACgAAAAABBAAMAAAACAAIAAAABAAIAAAABAAAAAEAAAAEAAAAzP///wAAARAUAAAAIAAAAAQAAAABAAAALAAAAAUAAABhcnJheQAGAAgABAAGAAAAAgAAABAAFAAIAAYABwAMAAAAEAAQAAAAAAABAxAAAAAcAAAABAAAAAAAAAAEAAAAaXRlbQAABgAIAAYABgAAAAAAAQA='}

However there's no embedded Arrow schema in the Parquet file written by DataFusion:

In [35]: df_table.write_parquet(FILENAME)

In [36]: meta = pq.read_metadata(FILENAME)

In [37]: meta.metadata # None

[kylebarron]

IMO not writing the Arrow schema to Parquet is a big bug.

Trying to track this down...

datafusion-python/src/dataframe.rs

Lines 510 to 520 in 79c22d6

	let mut options = TableParquetOptions::default();
	options.global.compression = Some(compression_string);
	wait_for_future(
	py,
	self.df.as_ref().clone().write_parquet(
	path,
	DataFrameWriteOptions::new(),
	Option::from(options),
	),
	)?;

This just calls datafusion::dataframe::DataFrame::write_parquet with the default options.

I'm not sure where on the datafusion side this fails.

[kylebarron]

Looks like it's this bug: apache/datafusion#11770