Skip to main content
TutorialApplies to: Unified engine deployments
Read six orders once, summarize them in two parallel paths, and combine both summaries. This example introduces a task graph without external data.

Before you start

You need the same engine and compute access as Your first workload. No catalog setup is needed.

1. Open the graph

The complete workload has four tasks:
Both summary tasks wait for orders. They can become ready together. combine waits for both summaries to succeed.

2. Inspect the connections

Each connection has a result alias. by_category and by_region each see their upstream rows as order_rows. Those aliases belong to different downstream tasks, so they can use the same name. combine receives category_totals and region_totals. These are temporary inputs from this run, not stored catalog tables. Do not run an individual dependent query in Workbench unless you also provide its input relations.

3. Run and verify

Validate, save with a Manual schedule, and start a run. Confirm that orders succeeds first, followed by both summary tasks, then combine. Parallel placement does not guarantee simultaneous execution. Available workers and the workload’s parallelism limit still apply. The final query produces: Inspect task states and attempts in the run. Use the retained-results API if you need the final rows directly; see Results and table commits.

4. Change one connection

In a separate draft, remove result sharing from the connection into by_category. The dependency still controls order, but order_rows is no longer supplied. This shows why an ordering dependency and a result alias have different roles. Restore the alias before saving the working example.

Finish

No tables are created. Continue with Tasks and dependencies, or use this graph to explore Selective reruns.
Last modified on November 4, 2026