SSSamarth Sinha

Note

Cross-Cloud Lakehouse Disaster Recovery

Keep a governed, warm copy of a Delta lakehouse in a second cloud, with recovery objectives you can rehearse and measure.

Oct 2026 · 1 min read

  • Databricks
  • Delta Sharing
  • Unity Catalog
  • Dataplex
Delta Sharing replication from a primary to a secondary cloud Primary · Azure: Databricks workspace, Delta tables, Unity Catalog. Secondary · Google Cloud: Replicated Delta, Failover compute, Catalog mirror, Failover runbook. Databricks workspace to Delta tables. Delta tables to Delta Sharing. Delta Sharing to Replicated Delta. Replicated Delta to Failover compute. Unity Catalog to Catalog mirror (policy mirror). Failover runbook to Failover compute (fail over). Infrastructure as code to Unity Catalog (provisions). Infrastructure as code to Catalog mirror (provisions). Primary · Azure Secondary · Google Cloud policy mirror fail over provisions provisions Databricks workspace primary compute Delta tables tiered by RPO Unity Catalog access policies Delta Sharing scheduled replication Replicated Delta object storage Failover compute Dataproc, scaled to zero Catalog mirror Dataplex policies Failover runbook rehearsed, RTO measured Infrastructure as code Terraform, both clouds
IllustrativeDelta Sharing replication from a primary to a secondary cloudScroll →

Context

A replica in the same region or the same cloud is not a disaster-recovery plan. Region-wide incidents, account-level lockouts and provider outages take the replica down with the primary.

The problem

Cross-cloud recovery is hard for three reasons: data must move continuously and cheaply, governance must mean the same thing on both sides, and the failover path must work on a bad day, not just on paper.

How it works

  1. Classify tables into recovery tiers, each with its own recovery point objective (RPO) and recovery time objective (RTO).
  2. Replicate Delta tables to the secondary cloud with Delta Sharing (or scheduled deep clones), at a cadence that meets each tier's RPO.
  3. Mirror governance: catalog entries, access policies and lineage from Unity Catalog into the secondary catalog (for example Dataplex).
  4. Pre-provision failover compute with infrastructure as code (Terraform), scaled to zero until needed.
  5. Rehearse failover with a runbook: endpoint switch, validation queries, measured RTO and RPO, and a path back.

Trade-offs

  • Egress costs set the replication cadence; tiering keeps them bounded.
  • Governance parity across two catalogs needs automation, or it drifts.
  • Streaming state (offsets, checkpoints) needs its own design; table replication alone does not cover it.
  • Targets are not results: an RTO of four hours is a design target until a rehearsal measures it.

When to use it

Use it when the business needs recovery from a provider-level failure. If a regional replica within one cloud meets the requirement, it is simpler and cheaper.