Merge #145899

145899: roachtest: start workload after initial scan r=jeffswenson a=jeffswenson

Previously, the ldr create test started the TPCC workload before the destination completed the initial scan. This caused the workload to build up backlog. TPCC is an update heavy workload that does not modify indexed columns. As a result, the KV writer is more efficient at replicating it than the SQL writer. The test was failing because the SQL writer took too long to catch up on the backlog.

Now, the initial table creation is allowed to complete before the SQL writer is started. The latency bounds were tightened since it no longer needs to catch up.

Release note none
Part of: #144224
Part of: #143958

Co-authored-by: Jeff Swenson <[email protected]>

Author: craig[bot]

pkg/cmd/roachtest/tests/logical_data_replication.go

@@ -426,6 +426,11 @@ func TestLDRCreateTablesTPCC(
 		option.WithNodes(setup.workloadNode),
 		fmt.Sprintf("./cockroach workload init tpcc --warehouses=%d --fks=false {pgurl:%d:system}", warehouses, setup.left.nodes[0]))
 
+	// Setup LDR before the workload starts. Otherwise the workload generates a
+	// backlog as the initial scan runs and it can take away while for the sql
+	// writer to catch up with the UPDATE heavy TPCC workload.
+	_, rightJobID := setupLDR(ctx, t, c, setup, workload, ldrConfig)
+
 	workloadDoneCh := make(chan struct{})
 	monitor := c.NewMonitor(ctx, setup.CRDBNodes())
 	monitor.Go(func(ctx context.Context) error {
@@ -435,9 +440,8 @@ func TestLDRCreateTablesTPCC(
 		// on data ingested via the offline initial scan.
 		return c.RunE(ctx, option.WithNodes(setup.workloadNode), workload.workload.sourceRunCmd("system", setup.left.nodes))
 	})
-	_, rightJobID := setupLDR(ctx, t, c, setup, workload, ldrConfig)
 
-	maxExpectedLatency := 3 * time.Minute
+	maxExpectedLatency := time.Minute + 30*time.Second
 	validateLatency := setupLatencyVerifiers(ctx, t, c, monitor, 0 /* leftJobID */, rightJobID, setup, workloadDoneCh, maxExpectedLatency)
 
 	monitor.Wait()