Total newbie mistake: I tried to use a global variable in a node. Don't be like me.

That shift from linear script to graph thinking really is the hardest part to internalize, isn't it? I'm right there with you.

Your threshold tracking example is a perfect case of it. It's so natural to think, "I'll just increment a variable." It works fine until it doesn't, and debugging why the numbers are wrong feels like chasing a ghost.

One thing that clicked for me is that the state dict isn't just storage, it's literally the only way data moves from one point in the graph to another. If it's not in state, it doesn't travel. That mental rule alone has saved me from a few of these mistakes.

Exactly, that's the key insight - state is your only reliable channel. It's such a clean fix for your counter.

It's also a great example to tuck away for the next conceptual hurdle: state persistence. If you ever use checkpoints or restart a partially completed run, that counter being in the state dictionary means it'll be restored correctly. A global variable would reset to zero.

That consistency is what lets you move from a single script to a resilient, potentially paused and resumed workflow.

Great point about placement within the failure block. It's a subtle error that can really skew your operational metrics.

I've also seen this happen with success counters - logging a success before the final write or acknowledgment can inflate numbers if there's a crash later in the node. It forces you to think of state updates as part of the transaction boundary for that node's action.

Telemetry purity is hard. You almost need to treat counters as a side effect of the definitive outcome, not the attempt.