When NOT to Use a Dial

Dials has a sharp boundary, and using it on the wrong side produces subtle production incidents rather than error messages. The test:

A value is a dial when the global default should flow through to every scope that hasn't explicitly overridden it.

That flow-through is what makes a dial useful — and what makes it dangerous on the wrong value — so check each candidate against the cases below.

Dial-shaped values

• A checkout fee that is 250bps everywhere, except two negotiated markets.
• A delivery threshold tuned per market as data comes in.
• A kill switch.
• A batch size, a timeout, a retry count that ops adjusts under load.

The common trait: when someone updates the global, you want every non-overridden scope to move with it, immediately.

Deliberately-priced bundles are NOT dials

Suppose each market's merchant pricing is negotiated: the fee, the partner's revenue share, and the settlement terms are decided together, per market, by a human looking at that market. Two properties disqualify it:

1. Flow-through is wrong. If the "default" fee changes, a negotiated market must NOT silently move. Snapshot semantics — full-copy rows per market — are the intent, not a smell to refactor away.
2. The columns are a bundle. Fee, share, and terms are one decision; a dial holds one value. Modeling a bundle as N independent dials invites changing one leg of a negotiated package in isolation.

Keep these as plain domain tables. If you're unsure which case you have, ask the operational question: "when the default changes, should priced markets move?" If anyone says "no, obviously not" — domain table.

Not configuration, not secrets, not flags

• Deploy configuration (database URLs, API endpoints, per-environment tuning) belongs in ENV / credentials — it changes with deploys and differs by environment, not by operator decision. ultra_settings is the good tool for that layering.
• Secrets never belong in a dial. The change log stores old and new values in plaintext by design.
• Feature flags overlap at the edges (a boolean dial is a fine kill switch), but percentage rollouts, actor targeting, and A/B experiments are a different problem — use Flipper when you need those.
• High-cardinality per-entity values (a price per SKU, a limit per user) are domain data. Dials scopes are for small enumerable dimensions — markets, platforms, locales — a table an operator can read, not a million-row store.

Rule of thumb

If the value's natural home is a spreadsheet an operator owns, it's probably a dial. If its natural home is a migration, a contract, or a vault — it isn't.