Azure SQL Database Hyperscale is designed to deliver fast, storage‑optimized backup and restore operations. According to Microsoft documentation, most Point‑in‑Time Restore (PITR) operations for Hyperscale should complete within 10 minutes, regardless of database size, because the service uses metadata-based restore techniques rather than copying full data files.
However, some real‑world scenarios can lead to unexpectedly long restore times, even when the source database is Hyperscale and even when no obvious configuration changes are made. This article explains one such scenario, outlines what happened, and provides guidance to avoid similar delays in the future.
Expected Behavior for Hyperscale PITR
Hyperscale databases use a unique architecture that separates compute and storage. Backups are taken from storage snapshots and do not require data copying during a typical PITR restore.
From Microsoft Learn:
“Most restores complete within minutes, even for large databases.”
Ref: https://learn.microsoft.com/azure/azure-sql/database/hyperscale-automated-backups-overview?view=azuresql#backup-and-restore-performance
This performance expectation applies as long as the Backup Storage Redundancy remains the same between the source DB and the target restore.
Customer Scenario Overview
A customer initiated PITR restore operations for Hyperscale DB:
- Source DB: Hyperscale (SLO: HS_PRMS_64)
- Target DB: Hyperscale (SLO: HS_PRMS_128)
- Same logical server
- Source DB Backup Storage Redundancy: Standard_RAGRS
- Customer enabled Zone Redundancy for the target database during restore
The customer therefore expected the restore to finish within the normal Hyperscale window (~10 minutes).
Instead, the restore took significantly longer.
Why the Restore Took Longer
Although the source database used Standard_RAGRS, enabling Zone Redundancy at restore time introduced a configuration change that affected the underlying Backup Storage Redundancy (BSR) for the newly restored database.
🔍 Key Point: Changing BSR Creates a Full "Size-of-Data" Restore
When the target DB uses a different BSR type than the source DB, Azure SQL Database cannot perform a fast metadata-based restore.
Instead, it must perform a full data copy, and the restore becomes proportional to the database size:
- More data → longer restore
- Effectively behaves like a physical data movement operation
- This overrides Hyperscale’s normally fast PITR workflow
This behavior is documented here:
https://learn.microsoft.com/azure/azure-sql/database/hyperscale-automated-backups-overview?view=azuresql#backup-and-restore-performance
In the customer’s case:
- Customer-enabled Zone Redundancy changed the restore workflow.
- As a result, the system selected a backup storage redundancy configuration different than the source:
- Restore workflow chose: Standard_RAGZRS
- Source database actually used: Standard_RAGRS (non‑zone‑redundant)
This mismatch triggered a size-of-data restore, leading to the observed delay.
Summary of Root Cause
✔ Hyperscale PITR is fast only when BSR is unchanged
✔ Customer enabled Zone Redundant configuration during restore
✔ This resulted in a different Backup Storage Redundancy from the source
✔ Target restore had to perform a full data copy, not metadata-based restore
✔ Restore time scaled with database size → leading to long restore duration
Key Takeaways
1. Do not change Backup Storage Redundancy during PITR unless necessary
Any change (e.g., RAGRS → RAGZRS) converts the restore into a size‑of‑data operation.
2. Restores that involve cross‑region or cross‑redundancy conversions always take longer
This applies equally to:
- PITR restore
- Restore to another server
- Restore with SLO changes
- Restore involving ZRS/RA‑GZRS transitions
3. Hyperscale PITR is extremely fast—when configuration is unchanged
If the source and target BSR match, Hyperscale restores usually complete in minutes.
4. Enabling Zone Redundancy is valid, but do it after the restore
If the customer wants ZRS for the restored DB:
- Perform PITR first (fast restore)
- Then update redundancy post‑restore (online operation)
Conclusion
While Hyperscale PITR restores are typically very fast, configuration changes during the restore—especially related to Backup Storage Redundancy—can trigger a full data copy and significantly increase restore duration.
To get the best performance:
- Keep the same BSR when performing PITR
- Apply redundancy changes after the restore completes
- Use metadata-based restores whenever possible
Understanding these nuances helps ensure predictable recovery times and aligns operational processes with Hyperscale’s architectural design.
Microsoft