Skip to main content
In the normal path a dataset’s storage names its catalog and table, and deploying it binds the dataset and starts the export; see the overview. This page covers the catalog itself and the cases that need an explicit binding.

​
Register a catalog

Provision the catalog, namespace and storage with your provider. Store the catalog’s JSON property map in the deployment’s private credential inventory and register it once per deployment:
Keep catalog-properties.json private and out of source control; declarations and binding files contain only the credential reference. bijection dataset catalog status analytics shows the catalog’s setup, its placements and any outstanding work; add --watch to follow it. Bijection creates missing tables and validates existing ones against your deployed declarations. It never creates namespaces or drops tables.

​
Bind a dataset explicitly

Bind explicitly when a dataset declares no storage, when you choose its table at deploy time, or for an external dataset that a job writes:
--check contacts the catalog and reports the access the binding needs, without submitting it. It creates no tables and does not test storage writes; submission checks the catalog again. --wait observes the accepted binding until every selected dataset has a published version, for at most the given duration; publication continues after the command exits. A dataset that declares storage binds as declared with bijection dataset bind orders; a binding that differs from the declaration is refused. A keyed external dataset also needs --buckets, its fixed logical bucket count from 1 to 32. To bind many datasets at once, put them in a file:
binding.json
Repeat an interrupted bind with the same configuration. To require a particular existing table, add its tableUuid to the binding.

​
Follow publication

bijection dataset status is the entry point for every dataset:
Without a dataset it lists every declared dataset with its target, serving state and publication summary. --watch refreshes once a second until interrupted; --format jsonl emits one observation per line. With --binding-revision or --after <receipt>, it waits for that exact requirement for at most --wait. Unknown freshness means there is no source observation yet; a successful check does not establish freshness. A publisher can be offline while its last complete version remains readable. Attention distinguishes automatic recovery from an action you must take. Each observation still checks current access. bijection dataset pause orders and resume stop and restart publication without changing the binding. bijection dataset resnapshot orders --request-id <key> establishes a new source basis when the source history can no longer be continued.

​
Drain a catalog

Draining stops publication to the catalog and waits for recorded work to settle; retained checkpoints must be released first. It preserves customer tables, catalogs and infrastructure. Each drained table keeps the property bijection.publication-state set to stopped and bijection.publication-stopped-boundary set to the last complete source position it reflects. Binding a table again removes both. To move a dataset to another table, drain its catalog, then bind it again.