changefeedccl: disable drain watcher by default

#102717 introduced a mechanism allowing for an orderly shutdown of changefeed
aggregators and an up-to-date frontier information to reduce duplicates during
a node restart in 23.2. However, we've recently identified a bug that could
lead to unexpected behavior. It's unclear where the bug is yet. This patch disables
this feature by default.

Fixes: #123371
Release note: Disabled a changefeed optimization on reducing duplicates during
aggregator restarts due to its bad performance.

pkg/ccl/changefeedccl/changefeed_processors.go

@@ -137,7 +137,7 @@ var aggregatorEmitsShutdownCheckpoint = settings.RegisterBoolSetting(
 	settings.ApplicationLevel,
 	"changefeed.shutdown_checkpoint.enabled",
 	"upon shutdown aggregator attempts to emit an up-to-date checkpoint",
-	true,
+	false,
 )
 
 type drainWatcher <-chan struct{}

pkg/ccl/changefeedccl/changefeed_test.go

@@ -6207,6 +6207,7 @@ func TestChangefeedHandlesRollingRestart(t *testing.T) {
 
 	db := tc.ServerConn(1)
 	sqlDB := sqlutils.MakeSQLRunner(db)
+	serverutils.SetClusterSetting(t, tc, "changefeed.shutdown_checkpoint.enabled", true)
 	serverutils.SetClusterSetting(t, tc, "kv.rangefeed.enabled", true)
 	serverutils.SetClusterSetting(t, tc, "kv.closed_timestamp.target_duration", 10*time.Millisecond)
 	serverutils.SetClusterSetting(t, tc, "changefeed.experimental_poll_interval", 10*time.Millisecond)