feat(snowflake)!: Transpilation support for TRY_CAST

[fivetran-ashashankar]

Transpilation support for try_cast snowflake function

[github-actions]

Benchmark Results

Legend: 🟢🟢 = 5%+ faster | 🟢 = 3-5% faster | 🟩 = 1-3% faster | ⚪ = unchanged | 🟧 = 1-3% slower | 🔴 = 3-5% slower | 🔴🔴 = 5%+ slower

sqlglot

Query	main	PR	diff
tpch	2.6ms	2.5ms	0.9% faster	⚪
short	202us	200us	1.0% faster	⚪
deep_arithmetic	7.4ms	7.5ms	1.3% slower	🟧
large_in	387.9ms	389.9ms	0.5% slower	⚪
values	431.0ms	429.8ms	0.3% faster	⚪
many_joins	10.8ms	10.8ms	0.0%	⚪
many_unions	41.2ms	41.0ms	0.6% faster	⚪
nested_subqueries	1.3ms	1.2ms	5.5% faster	🟢🟢
many_columns	11.8ms	11.6ms	1.3% faster	🟩
large_case	34.2ms	33.6ms	1.7% faster	🟩
complex_where	28.1ms	27.3ms	2.9% faster	🟩
many_ctes	17.1ms	17.0ms	1.0% faster	⚪
many_windows	20.8ms	20.5ms	1.5% faster	🟩
nested_functions	696us	687us	1.3% faster	🟩
large_strings	5.3ms	5.4ms	0.6% slower	⚪
many_numbers	99.7ms	98.8ms	0.8% faster	⚪

sqlglot[c]

Query	main	PR	diff
tpch	722us	696us	3.6% faster	🟢
short	61us	61us	0.6% faster	⚪
deep_arithmetic	2.8ms	2.8ms	0.2% faster	⚪
large_in	128.2ms	127.0ms	0.9% faster	⚪
values	131.1ms	132.7ms	1.2% slower	🟧
many_joins	2.8ms	2.6ms	7.0% faster	🟢🟢
many_unions	9.4ms	9.0ms	4.1% faster	🟢
nested_subqueries	253us	250us	1.5% faster	🟩
many_columns	3.4ms	3.2ms	5.1% faster	🟢🟢
large_case	10.4ms	10.1ms	3.4% faster	🟢
complex_where	7.7ms	7.4ms	4.6% faster	🟢
many_ctes	3.8ms	3.7ms	1.9% faster	🟩
many_windows	5.3ms	5.2ms	3.0% faster	🟢
nested_functions	171us	166us	3.1% faster	🟢
large_strings	1.5ms	1.5ms	0.7% slower	⚪
many_numbers	28.0ms	28.5ms	1.7% slower	🟧

---

Comment /benchmark to re-run.

[github-actions]

SQLGlot Integration Test Results

Comparing:

• this branch (sqlglot:RD-1069322-transpile_try_cast, sqlglot version: RD-1069322-transpile_try_cast)
• baseline (main, sqlglot version: 0.0.1.dev1)

⚠️ Limited to dialects: duckdb

By Dialect

dialect	main	sqlglot:RD-1069322-transpile_try_cast	transitions	links
duckdb -> duckdb	4003/4004 passed (100.0%)	4003/4004 passed (100.0%)	No change	full result / delta

Overall

main: 4004 total, 4003 passed (pass rate: 100.0%), sqlglot version: 0.0.1.dev1

sqlglot:RD-1069322-transpile_try_cast: 4004 total, 4003 passed (pass rate: 100.0%), sqlglot version: RD-1069322-transpile_try_cast

Transitions:
No change

[georgesittas]

Why are we doing this? I don't think this is an established pattern. Can you please elaborate on the difficulties you encountered while trying to transpile this function?

[fivetran-ashashankar]

on duckdb - SELECT TRY_CAST('05-Mar-2016' AS DATE)
on snowflake - TRY_CAST('05-Mar-2016' AS DATE) → returns DATE type (2016-03-05)

For DATE target
SELECT CAST(TRY_STRPTIME('05-Mar-2016', '%d-%b-%Y') AS DATE)

[fivetran-ashashankar]

so my thought was - DuckDB's TRY_STRPTIME needs to know the format string for parsing:

-- Different formats need different format strings
TRY_STRPTIME('05-Mar-2016', '%d-%b-%Y') -- DD-Mon-YYYY
TRY_STRPTIME('Mar-05-2016', '%b-%d-%Y') -- Mon-DD-YYYY
TRY_STRPTIME('03/05/2016', '%m/%d/%Y') -- MM/DD/YYYY
TRY_STRPTIME('March 5, 2016', '%B %d, %Y') -- Month DD, YYYY

_DATE_FORMAT_PATTERNS maps each format to its format string.

The pattern matching detects which format the literal string is using, so you know which format string to pass to TRY_STRPTIME.

[georgesittas]

There's an existing pattern in the codebase for this, though: calling build_formatted_time. Can't it be applied in this case as well? Doing manual regex matching doesn't seem like the right approach.

[geooo109]

I will close this for now, it's complicated and the solution isn't trivial.