Filter expressions not being applied

[jpereiranexar]

Environment

Delta-rs version: 0.22 (Tested multiple versions up to latest)
PyArrow version: 18.0.0

• OS: Mac 14.4 (23E214)
• Delta table on S3

---

Bug

What happened:

When applying a filter expression (lighting == "day") using pyarrow.dataset, no results are returned. However, if I do not apply the filter at this stage and instead filter the resulting pandas DataFrame (results[results["lighting"] == "day"]), I find that rows are filtered out, confirming that data matching the condition exists in the dataset.

What you expected to happen:

The filter method should correctly return rows where lighting == "day" when applied directly on the pyarrow.dataset.

How to reproduce it:

Given a delta table as such

CREATE TABLE hive_metastore.dwh.table_name (
  key STRING,
  ...
  lighting STRING,
  ...
  h3_id_res9 BIGINT)
USING delta
PARTITIONED BY (h3_id_res9)
LOCATION 'dbfs:s3_path'
TBLPROPERTIES (
  'delta.minReaderVersion' = '1',
  'delta.minWriterVersion' = '2')

# Python code

delta_table = get_delta_table(table_path, dynamo_table_name)
partitions = [("h3_id_res9", "in", str(608716487191953407))]
condition = pc.equal(ds.field("lighting"), "day")

# Apply filter directly on pyarrow dataset
results = (
    delta_table.to_pyarrow_dataset(partitions=partitions)
    .filter(expression=condition)
    .to_table()
    .to_pandas()
)

# Results are empty
assert results.empty, "Expected non-empty results, but got none."

# Remove filter and filter using pandas
results = (
    dt.to_pyarrow_dataset(partitions=partitions)
    .to_table()
    .to_pandas()
)
results_filtered = results[results["lighting"] == "day"].reset_index(drop=True)

# Results are non-empty and rows were filtered as expected
assert not results_filtered.empty, "Expected non-empty results, but got none after pandas filtering."

More Details:

• In other tables, I am able to filter the data, so I don't think it's tied to data type

[ion-elgreco]

Can you provide an MRE that I can run as-is with deltalake only?

[jpereiranexar]

I haven't been able to replicate the error using mock data. However, I discovered that applying the filter with match_substring does return the expected values.

For example:
This condition fails:

condition = (ds.field("lighting") == np.str_("day"))

This returns the expected values:

condition = pc.match_substring(ds.field("lighting"), "day")

I've already verified that the data does not contain leading or trailing spaces or unexpected characters in the lighting column.
Any ideas on what else I can check?

[ion-elgreco]

I haven't been able to replicate the error using mock data. However, I discovered that applying the filter with match_substring does return the expected values.

For example:
This condition fails:

condition = (ds.field("lighting") == np.str_("day"))

This returns the expected values:

condition = pc.match_substring(ds.field("lighting"), "day")

I've already verified that the data does not contain leading or trailing spaces or unexpected characters in the lighting column.
Any ideas on what else I can check?

Can you check the statistics for that column for all your add actions? Maybe there is something happening there.

Without an MRE it's difficult to help here