Permanent Snowflake tables carrying more than 1 GiB of Fail-safe storage

Seven days of history you pay for but cannot query

Every permanent table gets Fail-safe: a non-configurable 7-day period that starts when the Time Travel retention period ends. Only Snowflake can recover data from it, on a best-effort basis, and it is not a way for you to read old versions. The bytes held there are still billed to the table as storage.

For a curated source-of-truth table that trade is usually worth it. For a staging or intermediate table that a pipeline truncates and reloads, each reload pushes the replaced micro-partitions through Time Travel and then into Fail-safe, so the table can carry several copies of data it could rebuild from source in minutes. Transient tables have no Fail-safe period and therefore no Fail-safe cost.

Listing permanent tables by Fail-safe bytes

SELECT table_catalog, table_schema, table_name,
       ROUND(active_bytes    / POWER(1024, 3), 2) AS active_gib,
       ROUND(time_travel_bytes / POWER(1024, 3), 2) AS time_travel_gib,
       ROUND(failsafe_bytes  / POWER(1024, 3), 2) AS failsafe_gib
FROM SNOWFLAKE.ACCOUNT_USAGE.TABLE_STORAGE_METRICS
WHERE is_transient = 'NO'
  AND deleted = FALSE
  AND failsafe_bytes > 1073741824
ORDER BY failsafe_bytes DESC;

In TABLE_STORAGE_METRICS IS_TRANSIENT reads ‘YES’ for transient and temporary tables, which have no Fail-safe. A Fail-safe figure much larger than active bytes is the signature of a high-churn table.

The two conditions

ZopNight skips any table it knows to be transient. For the rest, it averages the table’s Fail-safe bytes over the last 14 days and fires when that average is above 1 GiB (1,073,741,824 bytes). There is no condition on active size, edition or how often the table is queried.

When the check stays quiet

A table with no Fail-safe series, or no datapoints in the 14 days, is not flagged. A table whose Fail-safe footprint sits at or under 1 GiB is left alone because a rebuild would not be worth the risk. The rule cannot tell whether your data is regenerable; that judgement stays with you.

No computed saving, and what you would recover

The finding is advisory and shows no dollar figure. Converting to transient removes the Fail-safe bytes for new changes, but Snowflake bills storage per TB at a rate that depends on edition, region and contract, and the old table’s history still has to age out. The value is roughly the table’s Fail-safe bytes multiplied by your storage rate.

Rebuilding the table as transient

A permanent table cannot be switched to transient in place. Build a copy and swap names:

1. Create the transient copy: CREATE TRANSIENT TABLE stg.events_t AS SELECT * FROM stg.events;
2. Check grants, clustering keys and dependent objects on the new table before going further.
3. Swap the two in one transaction: ALTER TABLE stg.events SWAP WITH stg.events_t;
4. Drop the old permanent copy once the pipeline has run cleanly against the new one.