← All notes
CMS Implementation6 min

A perspective is not a dataset

When a Sanity site needs “safe preview,” the first instinct is still a second dataset: staging, preview, launch-sandbox. That feels isolated. It is a fork. A perspective — published, drafts, or a Content Release stack — is a query-time view of the same dataset. Same GROQ, same document IDs, different which versions win. We keep seeing handovers where editors publish in production and the staging clone is three weeks behind, or where preview “works” only because someone mirrored tokens by hand. The API already gives you the view. You do not need a second lake to look at unfinished work.

A view is not a copy

A dataset is storage: documents, assets, and the credentials that can read or write them. Cloning it duplicates that surface. You now have two places to migrate schemas, two places to seed fixtures, two places for webhook secrets to drift, and two places for “which ID is the real homepage” arguments.

A perspective is a parameter on the query. published returns the live corpus. drafts applies every in-flight Studio edit as if published. A release ID stack shows a named future state layered over published. The documents stay in one lake; the client chooses which layer the same query reads.

That difference matters the moment an editor asks whether the launch page matches production. A clone answers “matches last sync.” A perspective answers “matches this dataset, through this lens, right now.”

What cloning quietly costs

Schema migrations run twice or they diverge. A field added on production never lands on staging until someone remembers. Preview then lies in the other direction: it shows an old shape while Studio already expects the new one.

Assets and references break the illusion of isolation. Images uploaded in one dataset do not exist in the other unless you copy them. References that looked fine in Studio resolve to missing docs in the clone. Teams “fix” it with another sync and train everyone to distrust the preview URL.

Access control gets noisier, not safer. A second write token for staging is another secret to rotate. A shared token across datasets is a wider blast radius. Neither replaces request-scoped draft access on the real dataset — the pattern we already use so public traffic never inherits previewDrafts.

When a second dataset still earns its keep

Perspectives do not replace every environment. Load tests that write junk documents, destructive migration rehearsals, and vendor sandboxes that must never touch live IDs still belong on a separate dataset or project. Those are write-isolation problems.

Preview is a read-isolation problem. Editors need to see unpublished or release-scoped content without shipping it. That is what perspective, Draft Mode, and a token that can see drafts are for — on the same dataset production already uses.

If your “staging CMS” exists only so marketing can click around before go-live, measure whether a perspective stack would delete the sync job. If the answer is yes, the clone is ceremony.

How we wire preview without a fork

Production clients use perspective: 'published' (the safe default on current Content Lake API versions) and the CDN. Preview requests opt into drafts or a release stack per request — never on a shared module import — and set useCdn: false so editors are not fighting a cached published page.

Launch rehearsals use a Content Release ID in the perspective stack, not a folder of DO NOT PUBLISH drafts and not a mirrored dataset. That pairs with treating a release as a coordinated future state, not as “whatever is currently drafted.”

Handover docs name the dataset once, then name the perspectives each surface uses. If the runbook still says “check staging Sanity,” rewrite it to “open Presentation with this release / drafts mode.” One lake, explicit lenses.

What we check before keeping a clone

Ask what the second dataset protects against that a perspective cannot. If the answer is “editors seeing drafts,” you already have drafts perspective and Draft Mode. If the answer is “accidental publish,” that is a Studio role and release workflow problem — copying the corpus does not remove the publish button on production.

Confirm production queries pin published and never fall back to raw on authenticated tokens. Confirm preview chooses perspective from the request. Confirm nobody is syncing documents between datasets on a schedule “to keep preview honest.”

When those checks pass, delete the unused clone or demote it to a true write sandbox with junk data only. Perspective is the preview tool. A dataset is where the content lives — usually once.

Have something to build?

Tell us what you're working on and we'll tell you honestly whether we're the right fit.

Work with us