Why Querykeep

Applications that poll an external API often grow a timer, a variable for the last value, and a boolean for “request in flight”. That works for one call. It gets expensive when several stops, feeds, or devices share one upstream budget: overlapping requests, retry storms, a dropped snapshot on the first error, and a poll loop that dies when the last view unmounts.

Querykeep is the piece that owns that loop for one process.

• One canonical key has one in-flight call. Readers share the snapshot.
• The client caps concurrency, spaces starts, and records a provider cooldown. A ticket blocked on one provider does not block another.
• The last accepted success stays in place through retries and admission failures. receivedAt tells you how old it is.
• The query keeps running with zero subscribers. pause, resume, and dispose are explicit.

You still write queryFn. The library does not know HTTP, GraphQL, or your schema. You parse Retry-After and pass the delay to client.cooldown.

What to use instead

Need	Fit
One process polls a rate-limited source, several readers share the result, and the loop should outlive any one view	Querykeep
A component cache with stale time, focus or reconnect revalidation, mutations, and React or Vue bindings	TanStack Query
An RxJS poll that should stop when the subscriber unsubscribes, and should error the stream when retries run out	rx-polling
A queue of tasks, with no snapshot, no key, and no retained last success	p-queue plus p-retry
A single setInterval around one call	A timer is enough

Querykeep assumes an RxJS peer dependency. The quickstart does not use operators. refreshOn$, enabled$, and select are the places RxJS shows up in your code. If you do not want that dependency, TanStack Query core is the more usual choice, and it does not share a provider admission budget across queries.

What this alpha will not grow into

Limits stay inside one client. There is no cross-tab leader, no durable job runner, and no storage port. Freshness is a comparison you make against receivedAt. A busy cycle drops extra triggers instead of queueing a trailing run. Those are deliberate. The limits page lists them.