Fail-safe
Fail-safe is a fixed, non-configurable 7-day period that begins the instant Time Travel ends for a permanent table. During it, data is recoverable only by Snowflake Support, for disaster scenarios, never self-service.
The full lifecycle of a micro-partition version
| TABLE TYPE | Active queries see it | Time Travel you: AT · BEFORE · UNDROP | Fail-safe Snowflake Support only | Purged gone for good |
|---|---|---|---|---|
| Permanent standard production | current | 1 day default · up to 90 * | 7 days, fixed | purged |
| Transient reconstructible data | current | 0 or 1 day | no Fail-safe | purged |
| Temporary dropped at session end | current | 0 or 1 day | no Fail-safe | purged |
Billed as storage while the version sits in any of these three stages: Active, Time Travel, Fail-safe
* Up to 90 days needs Enterprise edition or higher; Standard stops at 1 day.
Fail-safe is a last-resort safety net against catastrophic failure or operator error after Time Travel has lapsed. Every edition has it on permanent tables, and its storage is billed like any other storage. Transient and temporary tables get zero Fail-safe, which is precisely why they cost less to store: Snowflake never has to retain that extra week of old micro-partitions for them.
And it does bill. Fail-safe storage is charged like any other storage for its seven days, which is why a high-churn permanent table costs more than its size suggests: every rewritten micro-partition lingers through Time Travel plus seven more days. Thomas's staging tables rewrite themselves nightly, so as permanent tables with 1 day of Time Travel, each one has nine versions in storage at any moment: one active, one in Time Travel and seven in Fail-safe. Longer retention adds more. Converting them to transient is the standard fix, and it's the answer whenever the goal is to minimize the cost of reconstructible, high-churn data. A recovery that Support runs from Fail-safe also uses serverless compute, which is billed too.
Which tool for which disaster
| What went wrong | The recovery | Window |
|---|---|---|
| Bad UPDATE or DELETE | Query or clone with BEFORE (STATEMENT => query_ | Time Travel |
| TRUNCATE by mistake | INSERT INTO t SELECT * FROM t BEFORE (STATEMENT => '<truncate_ (table still exists, nothing to UNDROP) | Time Travel |
| Dropped table, schema, or database | UNDROP, after freeing the name if something took it | Time Travel |
| Need last week's state alongside today's | Zero-copy clone AT a timestamp or offset | Time Travel |
| Retention expired on a permanent table | Snowflake Support, disaster justification | Fail-safe · 7 days |
| Whole region down | An admin promotes the secondary failover group (see Replication) | Business Critical |
| Table type | Time Travel | Fail-safe | Typical use |
|---|---|---|---|
| Permanent | 0–90 days | 7 days, fixed | Standard production data |
| Transient | 0–1 day | None | Staging/ |
| Temporary | 0–1 day | None | Session-scoped scratch work, dropped automatically at session end |
"Time Travel you drive, Fail-safe Snowflake drives." You control Time Travel with AT/BEFORE/UNDROP, self-service, instant. Fail-safe has no user-facing controls or syntax at all, it's a Snowflake Support engagement, and it only exists after Time Travel is gone.

