Petra Novak

My interest is in automation that fails in a way you can debug at 3am without reading the vendor's documentation from the top. A workflow is a graph. Nodes do work, edges decide order, and the two failure modes that matter are a node that errors and a run that stops halfway with partial work already done. The second is the one that bites, because whether the run rolled back depends on settings nobody reads. Error handling has to be chosen per node. Some nodes should stop the run. Some should continue and let later nodes work around the gap. Some should retry, and a retry needs a backoff or it reproduces the load condition that caused the failure. I write these pages around that decision rather than around the node list. Idempotency is the other recurring theme. An automation that fires on a schedule and also on a webhook will run twice unless something in the design deduplicates, and the field to add for that is usually an idempotency key. Expressions are where workflows acquire a reputation for fragility. A string concatenated into a request body is an injection point, a path built by splitting a string breaks when the shape changes, and a numeric comparison on a value that may arrive as a string fails quietly rather than loudly. I end every page with what the run log shows, because a workflow you cannot inspect is a workflow you cannot fix. A workflow that touches anything a person can see needs a way to stop. A pause node waiting for approval is the cheapest control there is, and the alternative is discovering the mistake from an email.