data: Convert timezone offset to UTC for MySQL DATETIME coercion

MySQL converts timezone-aware string literals to the session
timezone (assumed UTC) before storing as DATETIME. For example,
CAST('2020-01-02 03:04:05+08:00' AS DATETIME) yields
'2020-01-01 19:04:05'.

ReadySet's text-to-DateTime coercion was preserving the raw
offset instead, causing incorrect comparisons in cached queries.

Fix the DfType::DateTime arm in TimestampTz::coerce_to() to
convert to UTC via naive_utc() when a timezone offset is present,
matching MySQL's behavior.

Fixes: REA-6487
Release-Note-Core: Fix MySQL DATETIME lookups with timezone
  offsets in WHERE clause.
Change-Id: I7a7b16c09f58f3656e5315a1d9aed1c2e88cfa00
Reviewed-on: https://gerrit.readyset.name/c/readyset/+/12281
Tested-by: Buildkite CI
Reviewed-by: Marcelo Altmann <marcelo@readyset.io>

Author: jasobrown-rs

logictests/mysql/datetime_lookup.test

@@ -47,6 +47,48 @@ select * from dt_test where dt < '2025-01-01 00:00:00'
 2
 2024-01-15 23:59:59
 
+-- REA-6487: Timezone offset in WHERE clause is converted to session tz (UTC).
+-- MySQL: CAST('2020-01-02 03:04:05+08:00' AS DATETIME) = '2020-01-01 19:04:05'
+
+statement ok
+create table dt_tz(id int, dt datetime);
+
+statement ok
+insert into dt_tz values (1, '2020-01-02 03:04:05');
+
+-- No offset: exact match.
+query IZ
+select * from dt_tz where dt = '2020-01-02 03:04:05'
+----
+1
+2020-01-02 03:04:05
+
+-- +00:00 offset: already UTC, matches.
+query IZ
+select * from dt_tz where dt = '2020-01-02 03:04:05+00:00'
+----
+1
+2020-01-02 03:04:05
+
+-- Insert rows at the expected UTC-converted values so we can verify
+-- positive matches, not just absence.
+statement ok
+insert into dt_tz values (2, '2020-01-01 19:04:05'), (3, '2020-01-02 08:04:05');
+
+-- +08:00 offset: converts to 2020-01-01 19:04:05 UTC, matches row 2.
+query IZ
+select * from dt_tz where dt = '2020-01-02 03:04:05+08:00'
+----
+2
+2020-01-01 19:04:05
+
+-- -05:00 offset: converts to 2020-01-02 08:04:05 UTC, matches row 3.
+query IZ
+select * from dt_tz where dt = '2020-01-02 03:04:05-05:00'
+----
+3
+2020-01-02 08:04:05
+
 -- Subsecond precision
 statement ok
 create table dt_micro(id int, dt datetime(6));

readyset-data/src/text.rs

@@ -976,16 +976,45 @@ mod tests {
     }
 
     #[test]
-    fn text_to_datetime_preserves_timezone() {
-        // DateTime (MySQL) routes through coerce_to which preserves the offset —
-        // MySQL DATETIME has no timezone stripping semantics like PG TIMESTAMP.
+    fn text_to_datetime_converts_tz_to_utc() {
+        // REA-6487: MySQL converts timezone-aware literals to session timezone
+        // (assumed UTC) when storing as DATETIME.
         let dt_type = DfType::DateTime {
             subsecond_digits: 0,
         };
+
+        // +08:00 offset: 03:04:05 in +08 = 19:04:05 UTC on the previous day.
         let val = DfValue::from("2020-01-02 03:04:05+08:00")
             .coerce_to(&dt_type, &DfType::DEFAULT_TEXT)
             .unwrap();
-        assert_eq!(val.to_string(), "2020-01-02 03:04:05+08:00");
+        assert_eq!(val.to_string(), "2020-01-01 19:04:05");
+
+        // -05:00 offset: 03:04:05 in -05 = 08:04:05 UTC.
+        let val = DfValue::from("2020-01-02 03:04:05-05:00")
+            .coerce_to(&dt_type, &DfType::DEFAULT_TEXT)
+            .unwrap();
+        assert_eq!(val.to_string(), "2020-01-02 08:04:05");
+
+        // +00:00 offset: already UTC, no change.
+        let val = DfValue::from("2020-01-02 03:04:05+00:00")
+            .coerce_to(&dt_type, &DfType::DEFAULT_TEXT)
+            .unwrap();
+        assert_eq!(val.to_string(), "2020-01-02 03:04:05");
+
+        // No offset: treated as session timezone (UTC), no change.
+        let val_no_tz = DfValue::from("2020-01-02 03:04:05")
+            .coerce_to(&dt_type, &DfType::DEFAULT_TEXT)
+            .unwrap();
+        assert_eq!(val_no_tz.to_string(), "2020-01-02 03:04:05");
+
+        // Subsecond digits + timezone offset + date/year boundary crossing.
+        let dt_type_6 = DfType::DateTime {
+            subsecond_digits: 6,
+        };
+        let val = DfValue::from("2020-01-01 00:30:00.123456+02:00")
+            .coerce_to(&dt_type_6, &DfType::DEFAULT_TEXT)
+            .unwrap();
+        assert_eq!(val.to_string(), "2019-12-31 22:30:00.123456");
     }
 
     #[test]

readyset-data/src/timestamp.rs

@@ -497,6 +497,7 @@ impl TimestampTz {
                 Ok(DfValue::TimestampTz(ts))
             }
             DfType::TimestampTz { subsecond_digits } => {
+                // NOTE: Assumes session timezone is UTC.
                 // TODO: when converting into a timestamp with tz on postgres should apply
                 // local tz, but what is local for noria?
                 let mut ts_tz = *self;
@@ -505,7 +506,16 @@ impl TimestampTz {
                 Ok(DfValue::TimestampTz(ts_tz))
             }
             DfType::DateTime { subsecond_digits } => {
-                let mut ts = *self;
+                // NOTE: Assumes session timezone is UTC.
+                // MySQL converts timezone-aware literals to the session timezone
+                // before storing as DATETIME.
+                let mut ts: TimestampTz = if self.is_zero() {
+                    Self::zero()
+                } else if self.has_timezone() {
+                    self.to_chrono().naive_utc().into()
+                } else {
+                    *self
+                };
                 ts.set_subsecond_digits(subsecond_digits as u8);
                 Ok(DfValue::TimestampTz(ts))
             }