The dataloader package ships with caching on. For the GraphQL resolver tree it was designed for that makes sense — stable reads within a single request are what you want, and the loader is garbage-collected with the request. Reuse a loader across requests or hold one for the lifetime of a server though and the cache becomes a Map that grows forever, never evicts, and quietly hands out stale reads.

I default to cache: false everywhere and only turn it on when the underlying data is genuinely small, immutable, and unambiguously worth the trouble — a lookup table of statuses, a tiny set of feature flags. Anything that can change while the process is running, the cache is off.