cheatsheets

Build the workflow first with workflow.build(), then call create_cycle() on the built workflow and define connections with field mappings, max_iterations, and a converge_when condition. Note that cyclic workflows are a planned feature not yet implemented in SDK v0.9.31.

Use the node-selection-guide cheatsheet, which provides a decision guide for choosing between nodes like PythonCodeNode for data transforms, HTTPRequest for API calls, and LLMNode for AI tasks. It is marked as a critical reference for workflow design.

The NameError happens because input_data does not exist on the first iteration of a cycle. Wrap access in a try/except block and provide default values, as shown in the cycle-debugging cheatsheet.

Yes, the DistributedTransactionManagerNode automatically selects between Saga and Two-Phase Commit patterns based on your consistency and availability requirements. If any participant lacks 2PC support, it defaults to the Saga pattern with compensation actions.

Key mistakes include using raw SQL instead of DataFlow, forgetting to call .build() before runtime.execute(), using relative imports, mocking in tier 2-3 tests, and omitting the result.* prefix when mapping PythonCodeNode outputs in cycles.

Create one pool at application startup and inject it into nodes using the external_pool parameter of AsyncSQLDatabaseNode. The SDK borrows the pool without closing it, preventing connection exhaustion in multi-worker deployments like Gunicorn with FastAPI.