feat(snowflake)!: transpile NTH_VALUE from Snowflake to DuckDB

[fivetran-felixhuang]

https://docs.snowflake.com/en/sql-reference/functions/nth_value

To transpile NTH_VALUE, we need to handle the following

1. NTH_VALUE can have an optional FROM LAST or FROM FIRST. However, DuckDB doesn't have a way to support FROM LAST (counting from the last item). In this case we just ignore the FROM LAST part
2. Snowflake's default frame for ranking window function is ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING (https://docs.snowflake.com/en/sql-reference/functions-window-syntax#usage-notes-for-window-frames), while in DuckDB the default frame is RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW when an ORDER BY clause is present. So we want to set the default Snowflake window frame during transpilation. Will handle this in a follow up PR
3. NULLS FIRST/LAST is not properly handled during tranpilation of Window functions.

   sqlglot/sqlglot/generator.py

   Line 2741 in ad2e9d3

   if isinstance(window, exp.Window) and window.args.get("spec"):

4. The syntax for NTH_VALUE in DuckDB is nth_value(expr, nth[ ORDER BY ordering][ IGNORE NULLS]). However, the [ ORDER BY ordering] part is not parsed correctly. Will need to address this in a different PR

[github-actions]

SQLGlot Integration Test Results

Comparing:

• this branch (sqlglot:transpile_nth_value_from_snowflake_to_duckdb, sqlglot version: transpile_nth_value_from_snowflake_to_duckdb)
• baseline (main, sqlglot version: 28.9.1.dev5)

⚠️ Limited to dialects: bigquery, duckdb, snowflake

By Dialect

dialect	main	sqlglot:transpile_nth_value_from_snowflake_to_duckdb	difference	links
bigquery -> bigquery	2592/2624 passed (98.8%)	2592/2624 passed (98.8%)	No change	full result / delta
bigquery -> duckdb	1847/2623 passed (70.4%)	1847/2623 passed (70.4%)	No change	full result / delta
duckdb -> duckdb	4003/4003 passed (100.0%)	4003/4003 passed (100.0%)	No change	full result / delta
snowflake -> duckdb	1315/2623 passed (50.1%)	1284/2623 passed (49.0%)	⬇ regressed by 1.2%	full result / delta
snowflake -> snowflake	2606/2623 passed (99.4%)	2553/2623 passed (97.3%)	⬇ regressed by 2.0%	full result / delta

Overall

main: 14496 total, 12363 passed (pass rate: 85.3%), sqlglot version: 28.9.1.dev5

sqlglot:transpile_nth_value_from_snowflake_to_duckdb: 14496 total, 12279 passed (pass rate: 84.7%), sqlglot version: transpile_nth_value_from_snowflake_to_duckdb

Difference: ⬇ regressed by 0.6%

[georgesittas]

@treysp @fivetran-felixhuang can we double-check these reported regressions in the test suite? Are they noise or?

[fivetran-felixhuang]

currently NULLS LAST/FIRST transpilation is not properly handled in window function, will need to investigate

sqlglot/sqlglot/generator.py

Line 2741 in ad2e9d3

if isinstance(window, exp.Window) and window.args.get("spec"):

[georgesittas]

I know this is still in draft mode, but wanted to raise awareness on the integration test suite regressions again and block the merge until we investigate them.