Data Residency
The region a workspace's data lives in, how every storage reference resolves to it, and what a region change moves versus leaves in place.
A workspace chooses a region, and the objects that are its apps' sources of truth live at rest in that region once each app is created there or moved there: database files, git repositories, build artifacts, private files, snapshots. This chapter covers the region model, how every storage reference knows where it lives, what is provisioned per region, and what happens when a workspace changes its mind.
#Regions
A region is a branded slug that maps to an AWS region. The slug, not the AWS region, is what gets stored, so the platform can move a slug's backing region without rewriting a single reference. The regions are configuration, and outside the cloud a region can be backed by any S3-compatible store instead (Portability).
| Slug | Location |
|---|---|
us-1 | United States (Oregon) |
eu-1 | European Union (Frankfurt) |
ca-1 | Canada (Central) |
au-1 | Australia (Sydney) |
Each region has its own pair of S3 buckets, its own private-files CloudFront distribution, its own access-log bucket, and its own SNS topic feeding the file-scan pipeline. The Terraform for a region is one module instantiated with a region-pinned provider, mirroring the base region's configuration exactly, so a new region is a stamp.
The residency promise covers data at rest. Compute runs in one region. Processing, including the vector index cache, is not covered by the promise, and the hot working-copy tier described in App Databases is bounded by an hourly checkpoint back to the home region.
#Self-Describing References
Every S3 read and write in the platform goes through one storage accessor operating on a storage reference: a value that names the logical bucket, the region and the key. Its canonical string form carries all three. A region is never part of an object's key; it selects the bucket. Objects come in two kinds:
- Self-describing. The stored reference or public URL carries the region, so a read recovers it from the value. Uploads, generated media, oversized values and feedback dumps are written this way, and they never move.
- Re-derived. The key is bare, and every read rebuilds the region from a durable record. Database files, git repositories, builds, static bundles and an app's file stores are written this way, and these are what a move copies.
Three authorities decide a region, and they are not interchangeable:
- The workspace's current region for creating a resource, and for uploads that belong to no app with a repository: the workspace's own.
- The app's pinned region, recorded on its repository row, for every object belonging to an existing app, its new uploads included.
- The database's pinned region, recorded on its catalog row, for the one SQLite file. A cloned database inherits its source, and the two migrate independently.
Each region's buckets are on a store, the service its driver names (S3 in an AWS region, or any S3-compatible service). The accessor runs every operation on a region's store through one interface, so nothing above it asks which kind of store a region has. Each store builds its clients lazily, one per flavor: a plain regional client, an accelerated client for browser uploads, and a retry-tuned client for high-frequency git backups.
#Changing Region
A change of region takes effect at once for new apps and for the workspace's own uploads. An existing app, its new uploads included, stays in the region it was created in until it is moved, and the settings page says so.
A workspace can move. The migration runs from a persisted status row, swept by a cron every five minutes, and moves the workspace's apps one move per app, at most ten at a time. Each move is a row that a cron advances every minute, one idempotent step at a time, so a partial run resumes cleanly.
Moved: everything read through a region re-derived from a durable record. App databases, the git repository tar and its file mirror and commit diffs, method builds, static bundles, presentation artifacts, diagnostics, and the app's file stores in both buckets.
Left in place: objects the app published, whose URL someone may have stored: step outputs, upload grants, and the like. They stay where they were published so those URLs keep resolving, and they are deleted with the app. Objects published before published keys had a segment of their own share names with the app's file stores, so anything under those names written before 2026-09-24 is copied and also left in place.
Deleted at the old region: pure caches, which re-render on demand at the new region.
An app is offline for its own move. The move locks it: requests receive a 503 response saying the app is moving, and queued work for it waits and runs after. The move then waits out work that began before the lock and writes the app's hot working copies back to storage. The object mover copies each object with size verification. At cutover the app's records are flipped to the new region and the app is unlocked. Only after that is the old region's copy deleted, key by key from the copy's own manifest, never by prefix. A move that fails before cutover deletes what it copied and unlocks the app where it was.
#Beyond Storage
Region shows up in three more places a reviewer will ask about.
- Public URLs for regional public files carry the slug as a path segment, so an object's public address says where it lives.
- Model routes carry data-boundary facts, so a workspace can see and pin where inference runs. See Models and Routing.
- Dedicated retrieval instances are provisioned per customer resource and rebuilt from that customer's artifacts in S3. See Retrieval.