changefeedccl: deflake TestChangefeedSchemaTTL

Previously, TestChangefeedSchemaTTL started flaking when
create_table_with_schema_locked was enabled by default in the test
suite. This happened because schema_locked tables can cause historical
versions of descriptors to be cached in the lease manager, as the
changefeed logic intentionally queries prior versions. As a result, even
after the test garbage-collected the descriptors table, these old
versions could remain in the in-memory cache, leading to flakes. To
address this, this patch adds a canary schema change, which bumps the
descriptor's version. When a descriptor's version is bumped, all older
cached versions are purged.

Fixes: cockroachdb#149167

Release note: None

Author: fqazi

pkg/ccl/changefeedccl/changefeed_test.go

@@ -6545,6 +6545,7 @@ func TestChangefeedSchemaTTL(t *testing.T) {
 		// Create the data table; it will only contain a single row with multiple
 		// versions.
 		sqlDB.Exec(t, `CREATE TABLE foo (a INT PRIMARY KEY, b STRING)`)
+		sqlDB.Exec(t, `CREATE USER joe`)
 
 		counter := 0
 		upsertRow := func() {
@@ -6575,6 +6576,15 @@ func TestChangefeedSchemaTTL(t *testing.T) {
 		// table to be deleted, with the middle version being lost to the changefeed.
 		forceTableGC(t, s.SystemServer, sqlDB, "system", "descriptor")
 
+		// Do an unnecessary version bump on the descriptor, which will purge old
+		// versions of the descriptor that the lease manager may have cached for
+		// historical queries. When schema_locked support is active, the act of
+		// detecting polling inside pauseOrResumePolling can cause old versions
+		// to be cached. The historical descriptors are cleared everytime a
+		// version bump occurs, otherwise this test can flake if the prior versions
+		// need to decode a row is already cached.
+		waitForSchemaChange(t, sqlDB, "ALTER TABLE foo OWNER TO joe")
+
 		// Resume our changefeed normally.
 		atomic.StoreInt32(&shouldWait, 0)
 		resume <- struct{}{}