Skip to content

Spark, Databricks: port upstream interval PRs #2614 and #2616 - #19

Open
moshap-firebolt wants to merge 8 commits into
firebolt/v0.63.0-patchesfrom
moshap/port-upstream-intervals
Open

moshap-firebolt wants to merge 8 commits into
firebolt/v0.63.0-patchesfrom
moshap/port-upstream-intervals

Conversation

@moshap-firebolt

@moshap-firebolt moshap-firebolt commented Oct 14, 2026 •

Copy link
Copy Markdown

Replaces #16 and #17 with the versions under upstream review, apache/datafusion-sqlparser-rs#2614 and #2616, which went further after review: INTERVAL is a column name when no literal follows (SELECT interval FROM t, interval - 1), a double-quoted string is an interval value, a bare number still requires a unit, and a string window frame bound stays a string literal.

Reverts #17 and #16 on top of #18, then cherry-picks the six upstream commits unchanged (four from apache#2614, two from apache#2616). The src/ and test changes are identical, line for line, to the upstream PRs; only conflict context in imports and test-file tails differs. The fork's full suite (--all-features), fmt and clippy are clean.

🤖 Generated with Claude Code


Note

Medium Risk
Changes core expression parsing and dialect-specific INTERVAL disambiguation; misclassification could affect AST shape for Spark/Databricks SQL and any consumer relying on old interval or identifier behavior.

Overview
Spark and Databricks now follow Spark-style INTERVAL rules: string literals can carry units inside the value (INTERVAL '2 months'), bare numeric literals still need an explicit unit, and value arithmetic inside the literal (INTERVAL 1 + 1 DAY) is rejected.

A new dialect flag supports_interval_string_without_qualifier drives parse_interval: literal-only values (signed numbers/strings, including double-quoted strings), optional qualifiers only for non-string values, and parse_interval_literal_value for the restricted value grammar. INTERVAL is no longer always reserved as an identifier—SELECT interval FROM t and interval - 1 parse as column references unless a number or string immediately follows INTERVAL. Window frame bounds keep plain string literals when that flag is off; other dialects still coerce '1' DAY to INTERVAL '1' DAY.

Shared and dialect-specific tests cover the new cases and narrow existing interval tests to dialects that require a unit but do not use string-without-qualifier mode.

Reviewed by Cursor Bugbot for commit a70bb84. Bugbot is set up for automated code reviews on this repo. Configure here.

Spark's multi-units interval syntax lets the string carry its own
units, as in INTERVAL '1 YEAR 2 DAYS 3 HOURS', so no unit follows the
literal. Databricks Runtime parses with the same grammar. A unit after
a literal is still read; the expression form this gives up
(INTERVAL 1 + 1 DAY) is not Spark syntax.
Address review: turning require_interval_qualifier off accepted
INTERVAL 1 and read SELECT interval FROM t as SELECT INTERVAL FROM AS t.
Keep it on and add supports_interval_string_without_qualifier: the value
must be a literal, and only an unsigned string may omit the qualifier.
INTERVAL is not reserved for these dialects, so SELECT interval FROM t
reads a column.
…ter interval is arithmetic

Address review: a double-quoted string is a Spark string literal, so it
may be an interval value and may omit the qualifier (INTERVAL "2 months",
INTERVAL "1" DAY). Only a bare number or a string after INTERVAL keeps
the interval error; a signed number falls back to the identifier, so
interval - 1 and INTERVAL -1 are arithmetic on the column, as in Spark.

A string window frame bound stays a string literal in these dialects
(RANGE BETWEEN '1 day' PRECEDING), as in Spark, rather than being read
as an interval without a qualifier.
Address review: Spark accepts INTERVAL '1' DAY TO as an alias, so assert
the error on TO x instead. parse_interval_number_without_unit_error and
parse_interval_expression_value repeat test_interval_numeric_requires_unit
and parse_interval_require_qualifier. The guard comment now covers
strings as well as numbers.
Spark's multi-units interval syntax lists several value / unit pairs,
INTERVAL 10 YEAR 20 MONTH, documented as the same as the string form
INTERVAL '10 YEAR 20 MONTH'. A new supports_interval_multi_units hook,
enabled for SparkSqlDialect and DatabricksDialect, reads the extra pairs
into that string form, so the AST is unchanged.
A double-quoted string is a Spark string literal, so it may be any value
of a multi-units interval, e.g. INTERVAL "1" DAY "2" HOUR.
@moshap-firebolt
moshap-firebolt force-pushed the moshap/port-upstream-intervals branch from c0c01aa to a70bb84 Compare October 20, 2026 01:02
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant