CREATE TABLE ... CLONE

CREATE TABLE ... CLONE creates a dataset as a fork of another one, copying no data. The new dataset's first manifest lists the same files the source's manifest listed, so the cost is the same whether the source holds a megabyte or a terabyte.

What lands is an ordinary dataset. You query it, write to it, time-travel it and drop it exactly as you would any other — and writes to it land in its own storage, never touching the files it borrowed. The source records that the fork exists, and will not expire the snapshot the fork rests on.

Syntax

sql

CREATE TABLE <target> CLONE <source>;
CREATE COLLECTION <target> CLONE <source_collection>;

The collection form forks every dataset in the source collection in one statement — which is how a set of related tables, like a benchmark schema, is taken in one go rather than one table at a time.

Parameters

<target> — the dataset to create. It must not already exist.
<source> — the dataset to fork. It must be a dataset with a manifest, so one projected from an external catalog cannot be cloned — use CREATE TABLE AS SELECT for those.

CLONE cannot be combined with AS SELECT, with column definitions, or with OR REPLACE: a clone takes the source's schema whole, so there is nothing for those to describe.

Examples

Fork a dataset

sql

CREATE TABLE personal.alice.lineitem CLONE samples.tpch_sf1.lineitem;

Fork a whole collection

sql

CREATE COLLECTION personal.alice.tpch CLONE samples.tpch_sf1;

Refused wholesale if any one dataset would be refused — the membership is read and every target name checked before the first fork is made, because a half-cloned collection is a state nobody can act on. Views are not cloned: a view is a query over dataset names, and copying one where those names mean something else would silently read the wrong data.

Query the fork, and edit it

sql

SELECT COUNT(*) FROM personal.alice.lineitem;

DELETE FROM personal.alice.lineitem WHERE l_quantity < 5;

The delete writes to the fork alone. samples.tpch_sf1.lineitem is unchanged, and so are every other fork of it.

Permissions

READ on the source. A fork exposes nothing a SELECT * would not, so forking is a read-tier act.
CREATE on the target, the same tier as any other CREATE TABLE.
Forking out of another workspace is refused while that workspace has egress_protection on. A fork is exactly the standing, systematic copy that guard exists to stop — a full mirror kept off the back of a single reader grant — so the source workspace's owner opens it deliberately with ALTER WORKSPACE <source> SET egress_protection TO OFF. No SECURE exemption applies: SECURE sanctions a named task or materialized view, and a hand-run statement is neither. Forking within one workspace is not egress and is never refused on this ground.

Notes

The source cannot be dropped or renamed while forks of it exist. Detach them first — see ALTER TABLE.
A fork of a fork is allowed, and inherits its parent's obligations: every dataset in the chain keeps the snapshot the fork below it rests on.
The fork occupies no storage of its own until you write to it, and you are billed for only what you add.