Snowflake stores tables as immutable micro-partitions. A write creates new partitions rather than overwriting bytes in place. That physical choice is why Time Travel (query AT / BEFORE a timestamp or statement id) and zero-copy clones (CREATE TABLE … CLONE) can exist as first-class SQL instead of a weekend restore project. It is also why storage bills confuse teams who think clone means "free forever."

Time Travel is retained metadata plus retained partitions

For a configurable retention window (often 1 day standard, up to 90 on enterprise editions for some objects), Snowflake keeps the partitions that a table needed to reconstruct older versions. UNDROP is the same machinery. After Time Travel expires, Fail-safe (for Snowflake-managed storage) keeps a non-customer-readable recovery window that only Snowflake support can use — you still pay for it in the storage model, and you cannot query it as AT.

sql
SELECT * FROM orders AT (TIMESTAMP => '2026-08-10 12:00:00'::TIMESTAMP);
CREATE TABLE orders_dev CLONE orders;

Clones copy pointers, not bytes, at creation time. As orders and orders_dev diverge, each side keeps partitions the other no longer references. The clone becomes expensive when both sides churn. A nightly clone of a 200 TB table that then gets rewritten in QA is not a trick; it is a second table with delayed storage realization.

What this is not

This is not application-level CDC. Time Travel is for recovery, audits of "what did this table look like," and spinning environments. It is a poor change-data-capture bus: use streams and tasks, or an external log, if you need incremental processing. It is also not a substitute for RBAC; a clone inherits structure, but you still manage who can read PII in the copy.

Transient and temporary tables have shorter or no Fail-safe; they exist to stop you from paying disaster-recovery storage for scratch data. Teams that build ETL landing tables as permanent with 90-day Time Travel discover a warehouse-sized version of git history they did not want.

Operational habits

Set Time Travel high on gold tables you would page to restore. Set it low on raw dumps. Clone for CI and developer sandboxes, then drop clones. Monitor TABLE_STORAGE_METRICS (or account usage views) for FAILSAFE_BYTES and TIME_TRAVEL_BYTES, not just active bytes. A merge that rewrites every partition ages out old ones only after retention — a large UPDATE is a storage event.

Cross-cloud and replication features have their own copy semantics; do not assume a clone in one account is a DR plan.

Snowflake's documentation on cloning considerations is the honest page: unique keys, sequences, stages, and pipes do not all clone the way a junior engineer hopes. Read it before you script CLONE DATABASE as the onboarding flow.

Time Travel is a gift of immutable partitions. Treat retention as a product decision with a budget owner, not as a default left at maximum because "storage is cheap" — in a warehouse, storage is a line item that grows with every rewrite you forgot to expire.