bug: `re_extract` behavior differs between backends

The behavior of re_extract differs between backends, with regards to what the index parameter means.

This method has signature re_extract(pattern, index), where pattern is a regex pattern with optional groups, and index is an index of the group to extract (returning NULL if no match or no group matches that index).

• duckdb, postgres, clickhouse, …: re_extract(pattern, 0) returns the part matching the first group if there’s a match, and NULL otherwise.
• sqlite, pandas, dask, pyspark: re_extract(pattern, 0) returns the whole string if there’s a match, and NULL otherwise. You need to pass in 1 not 0 to extract the first group.

To put it another way, given a column with a value "row_one", column.re_extract("row_([a-z])", 0) returns "one" for backends in the first group, and "row_one" for backends in the second.

Given the docstring, I think the first group has the intended behavior.

[cpcloud]

Is there a reason to support arbitrary values here, or could we restrict pattern and index to literal strings/integers?

A column of strings and a column of groups to extract might be a case where you’d want this.

I think we just handle it in the backend and raise an error during compilation. We do this elsewhere in the codebase.

I’m in favor of not restricting it unless it’s breaking something to keep it that way.

[cpcloud]

+1 on matching the behavior of re.match.