expr: preserve operand errors in non-strict AND/OR reduce

Closes: CLU-137

Fixes two user-visible bugs in how `reduce` treats an erroring operand of
a non-strict AND/OR:

* A materialized view or query whose `mz_now()` temporal filter sits under
  an OR of ANDs, e.g.

      WHERE (a AND ... AND mz_now() < t) OR (b AND ... AND mz_now() < t)

  could panic the compute worker (or fail to plan) with "Unsupported
  temporal predicate". The shared `mz_now() < t` conjunct was left buried
  inside the OR instead of being factored out, so temporal-filter
  extraction failed.

* A query that should short-circuit, such as `WHERE col AND (1/0 = 1)`,
  could spuriously fail at runtime (e.g. "division by zero") even on rows
  where `col` is false. AND/OR are non-strict: `false AND <error>` is
  `false` and `true OR <error>` is `true`, so such rows must be filtered,
  not errored.

Mechanism: the generic variadic fold in `reduce` replaced a call with any
operand's literal error unconditionally, which is wrong for a non-strict
function. Guard error-propagation with `propagates_nulls()`, mirroring the
null fold directly above. Keeping the erroring operand then reaches
`undistribute_and_or`, which recombines operands across AND/OR's
short-circuit boundary. That only preserves error semantics for operands
common to every disjunct, so skip undistribution otherwise. A shared
temporal predicate (`mz_now() < t`, whose cast can error) is common to
every disjunct, so it is still factored out and stays extractable, unlike
the reverted #37049's blunt `could_error()` skip.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: def-

src/expr/src/scalar.rs

@@ -849,6 +849,49 @@ impl MirScalarExpr {
         }
     }
 
+    /// Whether [`Self::undistribute_and_or`] can rewrite `self` without changing
+    /// its error semantics.
+    ///
+    /// Undistribution recombines the operands *not* common to every disjunct into
+    /// a new AND/OR, changing the short-circuit context they evaluate in. Only a
+    /// dominating operand (`false`/`true`), not a NULL one, absorbs another
+    /// operand's error, so recombining a could-error operand can change whether a
+    /// row errors (e.g. `a OR (a AND <err>)` raises for a NULL `a`, but
+    /// undistributes to `a`). Operands common to every disjunct are factored out
+    /// unchanged and stay sound, which is what lets a shared temporal filter like
+    /// `mz_now() < <expr>` (whose cast can error) be factored out, as the renderer
+    /// requires.
+    fn undistribution_preserves_errors(&self) -> bool {
+        let MirScalarExpr::CallVariadic {
+            func: func @ (VariadicFunc::And(_) | VariadicFunc::Or(_)),
+            exprs,
+        } = self
+        else {
+            return true;
+        };
+        // Fast path for the common case: with no erroring operand there is
+        // nothing to preserve, so skip the per-operand bookkeeping below.
+        // `could_error` recurses, so this also covers nested operands.
+        if !exprs.iter().any(|o| o.could_error()) {
+            return true;
+        }
+        let inner = func.switch_and_or();
+        // The operands of a disjunct (a non-`inner` disjunct is its own singleton).
+        let operands = |o: &MirScalarExpr| match o {
+            MirScalarExpr::CallVariadic { func, exprs } if *func == inner => exprs.clone(),
+            _ => vec![o.clone()],
+        };
+        let common = exprs
+            .iter()
+            .map(|o| BTreeSet::from_iter(operands(o)))
+            .reduce(|a, b| &a & &b)
+            .unwrap_or_default();
+        exprs
+            .iter()
+            .flat_map(operands)
+            .all(|op| !op.could_error() || common.contains(&op))
+    }
+
     /// AND/OR undistribution (factoring out) to apply at each `MirScalarExpr`.
     ///
     /// This method attempts to apply one of the [distribution laws][distributivity]
@@ -922,6 +965,12 @@ impl MirScalarExpr {
         while old_self != *self {
             old_self = self.clone();
             self.reduce_and_canonicalize_and_or(); // We don't want to deal with 1-arg AND/OR at the top
+
+            // Skip undistribution when it would not preserve error semantics.
+            if !self.undistribution_preserves_errors() {
+                return;
+            }
+
             if let MirScalarExpr::CallVariadic {
                 exprs: outer_operands,
                 func: outer_func @ (VariadicFunc::Or(_) | VariadicFunc::And(_)),
@@ -2475,6 +2524,78 @@ mod tests {
     use super::*;
     use crate::scalar::func::variadic::Coalesce;
 
+    #[mz_ore::test]
+    fn test_reduce_and_or_preserves_operand_errors() {
+        // #37049: AND/OR are non-strict, so a `false`/`true` operand absorbs an
+        // erroring operand at runtime. `reduce` must not fold or absorb an
+        // erroring operand into an error the original never raises.
+        let bool_typ = ReprScalarType::Bool;
+        let types = vec![bool_typ.clone().nullable(true)];
+        let err = || MirScalarExpr::literal(Err(EvalError::DivisionByZero), bool_typ.clone());
+        let col = || MirScalarExpr::column(0);
+        let arena = RowArena::new();
+
+        // `x AND <err>` must not fold to the error: `false AND <err>` is `false`.
+        let mut and = col().and(err());
+        and.reduce(&types);
+        assert!(!and.is_literal_err(), "{and:?}");
+        assert_eq!(and.eval(&[Datum::False], &arena), Ok(Datum::False));
+
+        // `a OR (a AND <err>)` must not absorb to `a`: a NULL `a` surfaces the
+        // error (`NULL OR (NULL AND <err>)` = `<err>`).
+        let mut or = col().or(col().and(err()));
+        or.reduce(&types);
+        assert_ne!(or, col(), "{or:?}");
+        assert_eq!(or.eval(&[Datum::True], &arena), Ok(Datum::True));
+        assert_eq!(or.eval(&[Datum::False], &arena), Ok(Datum::False));
+        assert_eq!(
+            or.eval(&[Datum::Null], &arena),
+            Err(EvalError::DivisionByZero)
+        );
+    }
+
+    #[mz_ore::test]
+    fn test_reduce_and_or_undistributes_common_error() {
+        // CLU-137: undistribution must still factor out a could-error operand
+        // common to every disjunct (e.g. a temporal `mz_now() < <expr>`), so it
+        // becomes a standalone conjunct the renderer can extract. The reverted
+        // #37049 guard skipped undistribution for any could-error expression,
+        // burying such predicates inside the OR.
+        let bool_typ = ReprScalarType::Bool;
+        let types = vec![
+            bool_typ.clone().nullable(true),
+            bool_typ.clone().nullable(true),
+        ];
+        let err = || MirScalarExpr::literal(Err(EvalError::DivisionByZero), bool_typ.clone());
+        let arena = RowArena::new();
+
+        // `(a AND <err>) OR (b AND <err>)` --> `<err> AND (a OR b)`.
+        let original = MirScalarExpr::column(0)
+            .and(err())
+            .or(MirScalarExpr::column(1).and(err()));
+        let mut reduced = original.clone();
+        reduced.reduce(&types);
+        assert!(
+            matches!(
+                &reduced,
+                MirScalarExpr::CallVariadic {
+                    func: VariadicFunc::And(_),
+                    ..
+                }
+            ),
+            "common error operand not factored out: {reduced:?}"
+        );
+        // ...and error semantics are preserved over every assignment.
+        for a in [Datum::True, Datum::False, Datum::Null] {
+            for b in [Datum::True, Datum::False, Datum::Null] {
+                assert_eq!(
+                    reduced.eval(&[a, b], &arena),
+                    original.eval(&[a, b], &arena)
+                );
+            }
+        }
+    }
+
     #[mz_ore::test]
     #[cfg_attr(miri, ignore)] // error: unsupported operation: can't call foreign function `rust_psm_stack_pointer` on OS `linux`
     fn test_reduce() {

src/expr/src/scalar/reduce/variadic.rs

@@ -51,9 +51,20 @@ pub(super) fn reduce_call_variadic(
         *e = MirScalarExpr::literal_null(e.typ(column_types).scalar_type);
         return;
     }
-    if let Some(err) = exprs.iter().find_map(|x| x.as_literal_err()) {
-        *e = MirScalarExpr::literal(Err(err.clone()), e.typ(column_types).scalar_type);
-        return;
+    // Fold to an operand's error under the same condition as the null fold
+    // above: only when the function is strict, so any operand error makes the
+    // whole call error. `propagates_nulls()` is the proxy. It is false for the
+    // non-strict variadics, which must be excluded: AND/OR (a `false`/`true`
+    // operand absorbs an erroring operand at runtime, e.g. `false AND <error>` is
+    // `false`) and ErrorIfNull (its message arg is evaluated only when the first
+    // arg is null). It is also false for the strict `greatest`/`least`, so we
+    // miss their fold, but that is harmless: their eager eval raises the error at
+    // runtime regardless. (Coalesce, also non-strict, bailed out above.)
+    if func.propagates_nulls() {
+        if let Some(err) = exprs.iter().find_map(|x| x.as_literal_err()) {
+            *e = MirScalarExpr::literal(Err(err.clone()), e.typ(column_types).scalar_type);
+            return;
+        }
     }
 
     // Per-function dispatch. Arms are mutually exclusive on discriminant; the

test/sqllogictest/error_semantics.slt

@@ -93,3 +93,16 @@ select coalesce(a, (select a/b from test)) from test;
 
 query error Evaluation error: division by zero
 select *, case when a = 5 then (select a/b from test) else 7 end from test;
+
+# Regression for #37049. AND/OR are non-strict: a `false`/`true` operand absorbs
+# an erroring operand at runtime (`false AND <error>` is `false`), so `reduce`
+# must not fold `x AND <error>` / `x OR <error>` to the error. Before the fix
+# these spuriously errored with division by zero (test.a = 1).
+query I
+select a from test where (a = 2) and (1/0 = 1);
+----
+
+query I
+select a from test where (a = 1) or (1/0 = 1);
+----
+1

test/sqllogictest/temporal.slt

@@ -440,3 +440,44 @@ FROM (SELECT CAST('62143-12-31' AS date) + '200000 YEARS' AS c2
       FROM (SELECT
             FROM (SELECT FROM t) AS subq_0) AS subq_1
       WHERE pg_catalog."current_timestamp"() = pg_catalog.pg_postmaster_start_time()) AS subq_2;
+
+# Regression test for https://linear.app/materializeinc/issue/CLU-137
+#
+# A temporal predicate (`mz_now() < ...`) shared across both branches of a
+# disjunction of conjunctions must be factored out of the disjunction, so that
+# it becomes a standalone conjunct the renderer can extract as a temporal
+# bound. The factoring is done by `undistribute_and_or` during predicate
+# canonicalization. The error-propagation fix in #37049 had skipped
+# undistribution for any expression that could error, and `mz_now()` always
+# counts as could-error, so the temporal predicate stayed buried inside the OR
+# and the renderer panicked with "Unsupported temporal predicate".
+
+# A constant collection (`VALUES`) is readable at any logical time, so the MV
+# can be queried `AS OF` a small timestamp, like the `valid` MV above.
+statement ok
+CREATE VIEW clu137 (active, confirmed, enabled, label, valid_until) AS VALUES
+    (true,  true,  true,  CAST(NULL AS text), 10),
+    (true,  true,  true,  '',                 20),
+    (true,  true,  true,  'x',                30),
+    (false, true,  true,  NULL,               40);
+
+statement ok
+CREATE MATERIALIZED VIEW clu137_mv AS
+SELECT valid_until FROM clu137
+WHERE (active AND confirmed AND enabled AND label IS NULL AND mz_now() < valid_until)
+   OR (active AND confirmed AND enabled AND label = ''   AND mz_now() < valid_until);
+
+query I rowsort
+SELECT * FROM clu137_mv AS OF 5;
+----
+10
+20
+
+query I rowsort
+SELECT * FROM clu137_mv AS OF 15;
+----
+20
+
+query I rowsort
+SELECT * FROM clu137_mv AS OF 25;
+----