I build automations, and I review the ones that quietly don't work.
Based in Slovenia. I run Računalniško programiranje Alej Kump s.p., registered 15 September 2026 — so this is new, and I'd rather say that than imply otherwise.
Reliability reviews of n8n workflows. The failure I care about isn't the red execution — it's the green run that did nothing, or did half the job, and told nobody. A node returns a clean 200 while a changed payload shape makes a mapping produce nothing. A duplicate check passes twice because two executions overlapped. The execution list shows green, so there's no error to search for.
Most of that is visible in the workflow's own export, before anything runs.
I wrote a deterministic analyser that reads an exported workflow and reports where it can fail. Three properties matter more than the finding list:
- It never executes anything from the export. No node runs, no request is made, no credential value is read. Code nodes are read as text.
- No parameter value or Code-node body appears in the report — a blanket rule, not a filter, because a filter is a list of the leaks someone thought of.
- Every finding states what it does not prove. "This contains an unhandled failure path" and "this has failed" are different claims, and an export only supports the first.
My lead-intake deduplication passed every test I wrote, for a full session. State survived across separate containers. Retention pruning worked. Every one of those tests was sequential — and the property it actually depended on was a concurrency guarantee.
Ten minutes in n8n's source settled it: static data is loaded per execution and written back as a whole blob after the execution ends. That's a read-modify-write race, and no code inside a Code node can close it, because the node never performs the write. Predicted from the source, then reproduced: two concurrent executions with the same key were both accepted, and three with distinct keys persisted only one — two dedupe keys silently discarded. No error. Three green runs.
Nothing was lost except the memory that those leads had been seen. That's the part that's hard to notice.
I sell the review described above — this is a commercial page in that sense, and I have no clients yet. What exists instead of a client list is a method you can inspect, including the part where my own tests were measuring the wrong thing.
- Site: https://alejkump-oss.github.io/
- Email: alej.kump@gmail.com
Please don't send credentials, API keys or tokens. Describe the workflow first — the export comes later, with secrets stripped.
Računalniško programiranje Alej Kump s.p. · Ulica heroja Marinclja 3, 1330 Kočevje, Slovenia · matična številka 7585942000 · davčna številka 67735509