feat!: add exp.Trunc for numeric truncation by doripo · Pull Request #6923 · tobymao/sqlglot

doripo · 2026-02-02T15:13:20Z

Add exp.Trunc expression class for numeric truncation, filling a gap alongside existing Round, Floor, and Ceil functions (in AST since 2021).

Implementation:

New exp.Trunc class with _sql_names = ["TRUNC", "TRUNCATE"]
Shared build_trunc in dialect.py for dialects with overloaded TRUNC (date vs numeric):
- DateTrunc: first arg is TEMPORAL_TYPES and (second arg exists or default_date_trunc_unit set), or second arg is TEXT_TYPES
- Trunc: first arg is NUMERIC_TYPES, or second arg is NUMERIC_TYPES, or (date_trunc_requires_part=False and no second arg)
- Anonymous: otherwise
Oracle: Uses shared builder with unabbreviate=False, default_date_trunc_unit="DD"
Exasol: Uses shared builder; handles DATETIME (no default unit, single-arg temporal falls back to Anonymous)
Snowflake: Uses shared builder with date_trunc_requires_part=False (single-arg TRUNC is always numeric)
MySQL: Generates TRUNCATE(...); TRUNC alias normalizes to TRUNCATE
T-SQL: Generates ROUND(x, n, 1)
SQLite: Single-arg only; decimals arg warns via @unsupported_args
Other dialects (PostgreSQL, DuckDB, BigQuery): Auto-supported via _sql_names
Date-only dialects (Hive, Spark): Unchanged

Tests:

Oracle: type inference tests (5 discrimination cases for shared build_trunc), identity and cross-dialect transpilation
Exasol: Type assertion, DATETIME handling, single-arg temporal fallback to Anonymous, cross-dialect transpilation
Snowflake: Numeric TRUNC identity, single-arg untyped is numeric, type assertions, cross-dialect transpilation, Anonymous fallback
MySQL: TRUNCATE identity and TRUNC alias normalization
T-SQL: ROUND(x, n, 1) generation
SQLite: Single-arg identity, decimals warning test
Spark/Hive: Assertions that TRUNC remains DateTrunc/TimestampTrunc
Cross-dialect date/timestamp TRUNC: validate_all tests for various units / dialects:
- Units DAY, WEEK, MONTH, QUARTER, YEAR, HOUR, MINUTE, SECOND
- Across Oracle, Snowflake, Postgres, BigQuery, DuckDB, TSQL, Spark

Out of scope / potential follow-up:

T-SQL ROUND(x, n, mode): third arg still passed through verbatim (pre-existing), producing invalid SQL in most dialects. Literal ,1 could now parse as Trunc, literal ,0 could simplify to 2-arg Round; non-literal / invalid mode would need a warning.

Add exp.Trunc expression class for numeric truncation, filling a gap alongside existing Round, Floor, and Ceil functions (in AST since 2021). Implementation: - New exp.Trunc class with _sql_names = ["TRUNC", "TRUNCATE"] - Oracle: Type-annotation-aware builder (DATE/TIMESTAMP/DATETIME -> DateTrunc, numeric -> Trunc) - Exasol: Updated to use exp.Trunc instead of Anonymous; fixed to handle DATETIME type and single-arg date TRUNC (defaults to 'DD' like Oracle) - MySQL: Generates TRUNCATE(...); TRUNC alias normalizes to TRUNCATE - T-SQL: Generates ROUND(x, n, 1) - Other dialects (PostgreSQL, Snowflake, DuckDB, BigQuery): Auto-supported via _sql_names - Date-only dialects (Hive, Spark) unchanged Tests: - Oracle: exp.Trunc identity and type assertion (vs DateTrunc), single-arg TRUNC - MySQL: TRUNCATE identity and TRUNC alias normalization - T-SQL: ROUND(x, n, 1) generation from other dialects' TRUNC - Exasol: exp.Trunc type assertion, DATETIME handling, single-arg date TRUNC Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

SQLite's TRUNC() only accepts one argument. This change: - Simplifies TRUNC(x, 0) to TRUNC(x) for SQLite - Warns when non-zero decimals are used (unsupported)

VaggelisD

Hey @doripo, thanks for the contribution!

Can we increase test coverage e.g in dialects like Spark to ensure that TRUNC(foo) remains exp.DateTrunc?
Snowflake's TRUNC is also overloaded, so we'd need to approach it similar to Oracle & Exasol. See this comment

sqlglot/dialects/exasol.py

sqlglot/dialects/oracle.py

sqlglot/dialects/sqlite.py

- SQLite: Use @unsupported_args("decimals") for best-effort transpilation, move trunc_sql inside Generator for auto-discovery, update tests - Spark/Hive: Add tests to verify TRUNC remains DateTrunc/TimestampTrunc Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

- Add build_trunc to dialect.py for Oracle/Exasol/Snowflake - Uses type annotation to distinguish date vs numeric truncation - Returns Anonymous if type cannot be determined - Add Snowflake support for overloaded TRUNC/TRUNCATE - Add comprehensive Snowflake TRUNC tests Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

tests/dialects/test_oracle.py

tests/dialects/test_exasol.py

Add validate_all tests for: - Oracle: Date truncation (DAY, WEEK, MONTH, QUARTER, YEAR), Timestamp truncation (HOUR, MINUTE, SECOND, DAY, MONTH, YEAR) - Exasol: Date/time truncation with Exasol-specific units (YYYY, MM, DD, HH, MI, SS, WW, Q) - Cross-dialect: Oracle, Snowflake, Postgres, BigQuery, DuckDB, TSQL, Spark Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

sqlglot/dialects/snowflake.py

VaggelisD

Thank you for the quick & accurate iterations! Leaving a few more comments your way, should be close to merging soon (also, appreciate the a-b sorting of the imports as a cherry on top 😆)

sqlglot/dialects/dialect.py

tests/dialects/test_exasol.py

tests/dialects/test_oracle.py

tests/dialects/test_snowflake.py

tests/dialects/test_spark.py

tests/dialects/test_snowflake.py

add .assert_is(exp.Trunc) as chained function call Co-authored-by: Vaggelis Danias <daniasevangelos@gmail.com>

- Use TEMPORAL_TYPES instead of DATE/TIMESTAMP/DATETIME to correctly handle all temporal types (e.g. TIMESTAMPTZ, TIMESTAMPLTZ) - Group build_trunc conditions by return type (DateTrunc/Trunc/Anonymous) - Chain .assert_is() to validate_identity calls - Merge redundant validate_all tests - Add fallback tests for Anonymous case - Remove duplicate Spark TRUNC test Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

- test_trunc_type_inference: tests build_trunc discrimination logic (shared across Oracle, Exasol, Snowflake) - 5 cases covering temporal+?, ?+string, numeric+?, ?+int, ?+? fallback - test_trunc: Oracle-specific identity and transpilation tests Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

doripo · 2026-02-03T14:45:20Z

@VaggelisD modified according to your suggestions, thanks, please also let me know if the trailing comment in the main PR description ("Out of scope / potential follow-up:") is appropriate or better remove it from the commit message.

VaggelisD · 2026-02-03T15:37:39Z

Thank you!

please also let me know if the trailing comment in the main PR description ("Out of scope / potential follow-up:") is appropriate or better remove it from the commit message.

It is fine for future reference, thanks for documenting it; The PR has already grown so it's great if we can now limit its scope to what's already there

VaggelisD

Hey @doripo, I believe I have one final comment and a few nits left, really appreciate your cooperation. If anything else comes up after that we'll take it to the finish line by merging this PR and applying the fixes afterwards

tests/dialects/test_exasol.py

sqlglot/dialects/dialect.py

tests/dialects/test_hive.py

tests/dialects/test_snowflake.py

tests/dialects/test_spark.py

sqlglot/dialects/dialect.py

fivetran-amrutabhimsenayachit · 2026-02-03T16:33:36Z

Hey @doripo, I believe I have one final comment and a few nits left, really appreciate your cooperation. If anything else comes up after that we'll take it to the finish line by merging this PR and applying the fixes afterwards

I agree, we can definitely increase the test coverage by adding more TRUNC specific tests in oracle, snowflake & exasol files. We can add it as part of a separate PR since this is a big PR.

Something like this for Oracle:

# Year variants
    for unit in ["YEAR", "YYYY", "YY", "Y", "SYYYY", "SYEAR"]:
        self.validate_identity(f"TRUNC(SYSDATE, '{unit}')")
    
    # Quarter
    for unit in ["Q", "QUARTER"]:
        self.validate_identity(f"TRUNC(SYSDATE, '{unit}')")
    
    # Month variants
    for unit in ["MONTH", "MM", "MON"]:
        self.validate_identity(f"TRUNC(SYSDATE, '{unit}')")
        
        ... and so on for all the date/time parameters.

Similarly we can do the same for snowflake (https://docs.snowflake.com/en/sql-reference/functions-date-time#label-supported-date-time-parts)

date_units = [
            ("YEAR", ["YEAR", "YEARS", "Y", "YY", "YYY", "YYYY", "YR", "YRS"]),
            ("QUARTER", ["QUARTER", "QUARTERS", "Q", "QTR", "QTRS"]),
            ("MONTH", ["MONTH", "MONTHS", "MM", "MON", "MONS"]),
            ("WEEK", ["WEEK", "W", "WK", "WEEKOFYEAR", "WOY", "WY"]),
            ("DAY", ["DAY", "DAYS", "D", "DD", "DAYOFMONTH"]),
        ]

        time_units = [
            ("HOUR", ["HOUR", "HOURS", "H", "HH", "HR", "HRS"]),
            ("MINUTE", ["MINUTE", "MINUTES", "M", "MI", "MIN", "MINS"]),
            ("SECOND", ["SECOND", "SECONDS", "S", "SEC", "SECS"]),
            ("MILLISECOND", ["MILLISECOND", "MILLISECONDS", "MS", "MSEC"]),
            ("MICROSECOND", ["MICROSECOND", "MICROSECONDS", "US", "USEC"]),
            (
                "NANOSECOND",
                [
                    "NANOSECOND",
                    "NANOSECONDS",
                    "NANOSECS",
                    "NS",
                    "NSEC",
                    "NANOSEC",
                    "NSECOND",
                    "NSECONDS",
                ],
            ),
        ]

Exasol:

# Basic numeric truncation
    self.validate_all(
        "TRUNC(123.456, 2)",
        write={
            "exasol": "TRUNC(123.456, 2)",
            "oracle": "TRUNC(123.456, 2)",
            "postgres": "TRUNC(123.456, 2)",
            "mysql": "TRUNCATE(123.456, 2)",
        }
    )

- Snowflake: date_trunc_requires_part=False (single-arg is numeric) - Oracle: default_date_trunc_unit="DD" (single-arg temporal uses DD) - Update tests for new behavior - Use validate_identity where SQL round-trips Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

- Use "truncation" (concept) instead of "TRUNC" (function name) - Consistent patterns: "Numeric truncation identity", "Date truncation with typed column and unit", "Cross-dialect numeric truncation transpilation" Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

VaggelisD · 2026-02-04T12:18:25Z

The PR grew to more dialects so I attempted to aggregate their docs for our reference during reviewing:

Dialect	Date Trunc	Decimal Trunc
Oracle	TRUNC(datetime_expr, [fmt])	TRUNC(numeric_expr, [n])
Snowflake	TRUNC[ATE](datetime_expr, fmt)	TRUNC[ATE](numeric_expr, [n])
Exasol	TRUNC[ATE](datetime_expr, [fmt])	TRUNC[ATE](numeric_expr, [n])
BigQuery	[DATE/TIME/DATETIME]_TRUNC(datetime_expr, fmt)	TRUNC(numeric_expr, [n])
Spark/DBX	TRUNC(datetime_expr, fmt)	❌
Presto/Trino	DATE_TRUNC(fmt, datetime_expr)	TRUNCATE(numeric_expr, [n])
Postgres	DATE_TRUNC(fmt, datetime_expr, [zone])	TRUNC(numeric_expr, [n])
DuckDB	DATE_TRUNC(fmt, datetime_expr)	TRUNC(numeric_expr)
MySQL	❌	TRUNCATE(numeric_expr, n)
SQLite	❌	TRUNCATE(numeric_expr, n)
Clickhouse	dateTrunc(fmt, datetime_expr, [zone])	❌

georgesittas

Great work

VaggelisD

Final comments from me too, thank you!

VaggelisD · 2026-02-04T12:49:52Z

tests/dialects/test_exasol.py

+        # Fallback to Anonymous (Exasol requires unit for date truncation)
+        self.validate_identity("TRUNC(CAST(x AS DATE))").assert_is(exp.Anonymous)


According to Exasol docs unit is optional so we could infer that this is date truncation here, right?

@VaggelisD

I looked into Exasol's TRUNC default behavior. I can't say that the docs are 100% explicit, but:

TRUNC/TRUNCATE docs say format is optional and "if a format is not specified, the value is truncated to days"

DATE_TRUNC is described as "PostgreSQL compatible"

The ROUND function docs reference "days from noon" as the rounding boundary, implying day-level granularity

This suggests DD (day truncation to midnight) matches Oracle's default, though not stated verbatim. Are you comfortable adding default_date_trunc_unit="DD" for Exasol, or would you prefer keeping the Anonymous fallback until we have firmer confirmation?

I have the fix ready (4 changed lines) just wanted to double check.

(continuing 77830a3#r2759813244)

Related: opened new issue #6959 for DD / DAY unit transpilation

tests/dialects/test_exasol.py

sqlglot/dialects/dialect.py

doripo · 2026-02-04T15:00:45Z

@VaggelisD that's a very good reference! I believe that these minor corrections to your table apply:

SQLite => instead of TRUNCATE(numeric_expr, n), should be:
TRUNC(x) - no decimals parameter, no TRUNCATE function
ClickHouse => instead of ❌ (none), should be:
Actually has trunc(x[, N]) with alias truncate
Presto/Trino => Correct, but note: returns double, not integer
Spark/DBX => Correct, but note: may be worked around e.g.:
IF(x >= 0, FLOOR(x, d), CEIL(x, d))

Based on my understanding. the current gaps are therefore quite small (about 20 lines of code, not including the tests):

Dialect	Current	Correct	Status
DuckDB	`TRUNC(x, d)`	`TRUNC(x)` only	Needs `@unsupported_args("decimals")`
Presto/Trino	`TRUNC(x, d)`	`TRUNCATE(x, d)`	Needs `rename_func("TRUNCATE")`
ClickHouse	Not handled	`trunc(x, d)`	Needs `rename_func("trunc")`
Spark/Hive	Not handled	No native support	Needs `IF(x>=0, FLOOR(x,d), CEIL(x,d))` workaround

I can address them either now or in a follow-up PR, as they are incremental improvements.

georgesittas · 2026-02-04T15:15:29Z

Let's get any minor fixes in and then improve the parsing / transpilation in a follow-up to avoid increasing the scope more.

doripo · 2026-02-04T15:17:03Z

Or, for Spark/Hive probably better to just cast to BIGINT with unsupported_args("decimals") - the 'smart' workaround is somewhat weird..

If I do that now, then all the code fixes are really minor - and just a few more tests. @georgesittas does that seem acceptable at this stage of the PR?

georgesittas · 2026-02-04T15:20:06Z

Let's do it in a follow-up PR @doripo, we can apply any last one-liners here.

fivetran-amrutabhimsenayachit

Thank you!

Style fixups Co-authored-by: Vaggelis Danias <daniasevangelos@gmail.com>

doripo added 4 commits February 2, 2026 17:09

style: fix line length in exasol and oracle

2ac9378

fix(sqlite): handle TRUNC decimals argument limitation

2736621

SQLite's TRUNC() only accepts one argument. This change: - Simplifies TRUNC(x, 0) to TRUNC(x) for SQLite - Warns when non-zero decimals are used (unsupported)

style: fix line length in tsql

b4cb97d

VaggelisD reviewed Feb 2, 2026

View reviewed changes

doripo added 2 commits February 2, 2026 19:45

fivetran-amrutabhimsenayachit reviewed Feb 2, 2026

View reviewed changes

tests/dialects/test_oracle.py Outdated Show resolved Hide resolved

fivetran-amrutabhimsenayachit reviewed Feb 2, 2026

View reviewed changes

tests/dialects/test_exasol.py Show resolved Hide resolved

doripo requested review from VaggelisD and fivetran-amrutabhimsenayachit February 2, 2026 19:03

fivetran-amrutabhimsenayachit reviewed Feb 2, 2026

View reviewed changes

sqlglot/dialects/snowflake.py Outdated Show resolved Hide resolved

Merge branch 'tobymao:main' into feat-numeric-trunc

6aeb57c

VaggelisD reviewed Feb 3, 2026

View reviewed changes

VaggelisD mentioned this pull request Feb 3, 2026

feat(duckdb): Add transpilation support for TRUNC date/time function in snowflake #6935

Closed

doripo and others added 4 commits February 3, 2026 14:48

Apply suggestions from code review

2f10a68

add .assert_is(exp.Trunc) as chained function call Co-authored-by: Vaggelis Danias <daniasevangelos@gmail.com>

Merge branch 'tobymao:main' into feat-numeric-trunc

9f6021c

doripo requested a review from VaggelisD February 3, 2026 14:43

Merge branch 'tobymao:main' into feat-numeric-trunc

0c72ac3

VaggelisD reviewed Feb 3, 2026

View reviewed changes

sqlglot/dialects/dialect.py Outdated Show resolved Hide resolved

doripo added 3 commits February 3, 2026 21:08

Merge branch 'tobymao:main' into feat-numeric-trunc

3dab08c

test: add missing assert_is(exp.Trunc) checks

3d836f2

Signed-off-by: Dori Polotsky <doripo@riverpool.ai>

doripo requested review from VaggelisD and fivetran-amrutabhimsenayachit February 3, 2026 21:33

Merge branch 'tobymao:main' into feat-numeric-trunc

4dd28c4

Merge branch 'tobymao:main' into feat-numeric-trunc

2f5d4c9

georgesittas approved these changes Feb 4, 2026

View reviewed changes

VaggelisD approved these changes Feb 4, 2026

View reviewed changes

fivetran-amrutabhimsenayachit approved these changes Feb 4, 2026

View reviewed changes

Apply suggestions from code review

36b9583

Style fixups Co-authored-by: Vaggelis Danias <daniasevangelos@gmail.com>

georgesittas merged commit 8725010 into tobymao:main Feb 4, 2026
9 checks passed

This was referenced Feb 4, 2026

Date part units (DD/DAY) don't transpile correctly across dialects #6959

Closed

feat(generator): Add numeric TRUNC output for additional dialects #6961

Merged

		# Fallback to Anonymous (Exasol requires unit for date truncation)
		self.validate_identity("TRUNC(CAST(x AS DATE))").assert_is(exp.Anonymous)

Comments

Conversation

doripo commented Feb 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

VaggelisD left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

VaggelisD left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

doripo commented Feb 3, 2026

Uh oh!

VaggelisD commented Feb 3, 2026

Uh oh!

VaggelisD left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

fivetran-amrutabhimsenayachit commented Feb 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

VaggelisD commented Feb 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

georgesittas left a comment

Choose a reason for hiding this comment

Uh oh!

VaggelisD left a comment

Choose a reason for hiding this comment

Uh oh!

VaggelisD Feb 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

doripo Feb 4, 2026

Choose a reason for hiding this comment

Uh oh!

doripo Feb 4, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

doripo commented Feb 4, 2026

Uh oh!

georgesittas commented Feb 4, 2026

Uh oh!

doripo commented Feb 4, 2026

Uh oh!

georgesittas commented Feb 4, 2026

Uh oh!

fivetran-amrutabhimsenayachit left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

doripo commented Feb 2, 2026 •

edited

Loading

fivetran-amrutabhimsenayachit commented Feb 3, 2026 •

edited

Loading

VaggelisD commented Feb 4, 2026 •

edited

Loading

VaggelisD Feb 4, 2026 •

edited

Loading