feat(snowflake)!: update transpilation of DATE_TRUNC to duckdb

[toriwei]

Improve Snowflake --> DuckDB transpilation of DATE_TRUNC.

Return types / casting

• Snowflake's return type matches the type of the date/time value (second argument) -- docs
• DuckDB's return type matches the date/time granularity (first argument) -- docs

Snowflake: SELECT DATE_TRUNC(YEAR, TIMESTAMP '2026-01-01 00:00:00')
2026-01-01 00:00:00
Transpiled DuckDB: SELECT DATE_TRUNC('YEAR', CAST('2026-01-01 00:00:00' AS TIMESTAMP))
2026-01-01


Snowflake: SELECT DATE_TRUNC('HOUR', DATE '2026-01-01')
2026-01-01
Transpiled DuckDB: DATE_TRUNC('HOUR', CAST('2026-01-01' AS DATE))
2026-01-01 00:00:00

Fix:
In the Snowflake parser, modify the AST when the input of DATE_TRUNC(granularity_part, value) is either not both dates or both not times. Then, handle casting to the granularity_part type on DuckDB side.

Snowflake: DATE_TRUNC(YEAR, TIMESTAMP '2026-01-01 00:00:00')
2026-01-01 00:00:00
Transpiled DuckDB: CAST(DATE_TRUNC('YEAR', CAST('2026-01-01 00:00:00' AS TIMESTAMP)) AS TIMESTAMP)
2026-01-01 00:00:00

Snowflake: DATE_TRUNC('HOUR', DATE '2026-01-01')
2026-01-01
Transpiled DuckDB: CAST(DATE_TRUNC('HOUR', CAST('2026-01-01' AS DATE)) AS DATE)
2026-01-01

Support Snowflake's DATE_TRUNC(time_part, time)
Snowflake can truncate time input (HH:MM:SS), but DuckDB cannot.

Snowflake: SELECT DATE_TRUNC(HOUR, TIME '14:23:45.123456') AS time_hour
14:00:00
DuckDB: SELECT DATE_TRUNC('HOUR', CAST('14:23:45.123456' AS TIME)) AS time_hour
Binder Error:
No function matches the given name and argument types 'date_trunc(STRING_LITERAL, TIME)'. You might need to add explicit type casts.
        Candidate functions:
        date_trunc(VARCHAR, DATE) -> TIMESTAMP
        date_trunc(VARCHAR, INTERVAL) -> INTERVAL
        date_trunc(VARCHAR, TIMESTAMP) -> TIMESTAMP

Fix: Add a dummy date to the time expression to create a full date and time. After truncating, cast it to TIME to isolate the time.

DuckDB: SELECT CAST(DATE_TRUNC('HOUR', CAST('1970-01-01' AS DATE) + CAST('14:23:45.123456' AS TIME)) AS TIME);
14:00:00

[github-actions]

SQLGlot Integration Test Results

Comparing:

• this branch (sqlglot:tori/date-trunc-casting, sqlglot version: tori/date-trunc-casting)
• baseline (main, sqlglot version: 28.5.1.dev72)

⚠️ Limited to dialects: snowflake, duckdb

By Dialect

dialect	main	sqlglot:tori/date-trunc-casting	difference	links
duckdb -> duckdb	4003/4003 passed (100.0%)	4003/4003 passed (100.0%)	No change	full result / delta
snowflake -> duckdb	578/847 passed (68.2%)	579/847 passed (68.4%)	⬆ improved by 0.1%	full result / delta
snowflake -> snowflake	847/847 passed (100.0%)	847/847 passed (100.0%)	No change	full result / delta

Overall

main: 5697 total, 5428 passed (pass rate: 95.3%), sqlglot version: 28.5.1.dev72

sqlglot:tori/date-trunc-casting: 5697 total, 5429 passed (pass rate: 95.3%), sqlglot version: tori/date-trunc-casting

Difference: No change

[georgesittas]

Don't we want to set a flag in Snowflake that means "the output type matches the value's type" instead of specifying an "action" like "cast", etc? This feels backwards to me.

Perhaps input_type_preserved more accurately represents this? Ditto for TimestampTrunc.

[toriwei]

renamed to input_type_preserved. i agree, this naming is a better representation since we set the flag in Snowflake (snowflake.py#L267)

[georgesittas]

Let's only annotate if the boolean flag is set, otherwise the type attribute is not needed.

[georgesittas]

Couple of final comments, feel free to merge when addressed.