Repository navigation
Add test coverage for grouping on dictionary encoded columns #8791
Description
Is your feature request related to a problem or challenge?
At InfluxData we use Dictionary(Int32, Utf8) heavily for certain columns of our queries.
This particular type of data is not well represented in the DataFusion codebase, which leads to issues such as #8738
In addition to not being well covered by functional tests, we also don't have particularly good coverage performance wise, especially for some of our important cases like grouping performance. This leads to dueling speculations on better/worse ways to run such queries, such as #7647 but now good way to evaluate them
Describe the solution you'd like
I would like to ensure we have adequate test coverage
- Functional tests so we don't hit regressions accidentally again
- Performance tests (e.g. so we could evaluate the changes contemplated by Materialize Dictionaries in Group Keys #7647
Describe alternatives you've considered
I added some functional sqllogictests in
However, those are not reading from Parquet files (as I could not figure out how to make a parquet file have the same schema I wanted)
Additional context
No response
Activity
Rich-T-kid commentedon Apr 2, 2026
take
Rich-T-kid commentedon Apr 7, 2026
@alamb I’m a bit confused by what you meant by:
"However, those are not reading from Parquet files (as I could not figure out how to make a parquet file have the same schema I wanted)"
In dictionary.slt I see that this statement exists:
which writes out the table as a partitioned Parquet dataset.
I’ve also included my own test in a draft PR where I create a dictionary-encoded column, write it out to a Parquet file, then read it back in again and do a schema check (confirm it remains Dictionary(...)) as well as a basic filter.
Is this what you meant by “reading from a Parquet file”? Or were you referring to a more specific Parquet scenario (e.g., a particular schema/layout/encoding that you weren’t able to reproduce) for the coverage you’re looking for?
writing table to parquet followed by read and schema check (#21444)
writing table to parquet followed by read and schema check (apache#21444