A deleted dataset is not just a storage problem. It can remove the tables behind financial reports, break scheduled transformations, interrupt downstream applications and leave analysts unsure which numbers can be trusted. The immediate question is usually, “Can we get the data back?” The more important management question is, “Can we restore a usable, authorized, validated reporting service before the business deadline?”
BigQuery provides meaningful recovery capabilities, including time travel and dataset undelete. Yet those features are bounded by retention windows, object types, region, permissions and operational dependencies. A reliable recovery design therefore needs more than a checkbox labelled “backup.” It needs an inventory, tiered recovery objectives, a documented restore procedure, independent protection for critical data and regular tests.
This guide explains what the platform can recover, where native recovery stops, and how to build a proportionate policy that a data team can actually operate.
The executive answer
For an accidentally deleted BigQuery dataset, first preserve the incident timeline and identify the original project, dataset ID and region. If the deletion falls within the dataset’s configured time travel window, evaluate BigQuery’s dataset undelete capability before recreating anything under the same name. Then restore access, refresh dependent objects and validate data against trusted controls. If the native window has expired, recovery depends on whatever independent copies, snapshots, exports or reproducible source pipelines you arranged beforehand.
The policy should separate four questions:
- How much data can be lost?
- How quickly must reports return?
- What exactly must be rebuilt besides table contents?
- Who has the authority and tested procedure to do it?
Answers to those questions determine whether native undelete alone is sufficient.
What BigQuery can recover today
1. Dataset undelete within the time travel window
BigQuery documents a dataset recovery operation that restores a deleted dataset to its state at deletion, including objects, dataset properties and security settings, subject to documented limitations. You can invoke it with the SQL statement UNDROP SCHEMA or the datasets.undelete API. This is a major improvement over manually reconstructing each table, but it is not an unlimited backup service.
The recovery window is the time travel setting in effect when the dataset was deleted. Google documents a configurable range of two to seven days, with seven days the default. Increasing the setting after an incident does not retroactively extend the deleted dataset’s recoverability. Confirm the actual configured duration instead of assuming every environment has seven days. The official dataset restore procedure explains prerequisites, syntax and errors.

2. Time travel for historical table versions
Time travel allows access to earlier table data within the configured window. It can help reverse a mistaken overwrite, bad transformation or deletion even when the dataset itself remains available. Restoring a deleted table uses a time decorator and a copy operation to create a recoverable table; the details differ from undeleting an entire dataset.
Time travel is short-lived protection against recent changes, not a retention policy for a quarterly or annual recovery point. It also does not reproduce every kind of table metadata when restoring historical data. Review the time travel and fail-safe documentation before relying on it for a regulated or critical workflow.
3. Fail-safe is an emergency layer, not an ordinary restore path
Google documents an additional seven-day fail-safe period for certain deleted or replaced table data after time travel ends. Current documentation describes an emergency table-level recovery process that requires Cloud Customer Care to grant temporary access, followed by a table copy operation. It is not a self-service dataset undelete extension. Large restores may take time and retries, and the window cannot be extended.
The operational implication is important: do not set a seven-day recovery objective on the assumption that fail-safe will behave like a normal backup. Treat it as a last-resort escalation with uncertain turnaround, subject to eligibility and support engagement, while maintaining separate recoverable copies for critical workloads.
4. Table snapshots for recovery points beyond time travel
A BigQuery table snapshot preserves table contents at a chosen point in time and can be retained longer than the time travel window. Snapshots are read-only, and a standard table can be created from one for restoration. They are useful for important financial, reference and curated tables where a previous known-good state must remain available for weeks or months.
Snapshots are not a complete dataset image. They do not directly snapshot logical views, materialized views or external tables. Some streaming-buffer data is excluded; metadata and expiration behavior need separate review. Snapshot storage cost depends on changes to underlying storage blocks and can increase substantially after rewrites. See Google’s table snapshot guidance for limitations and billing behavior.

Why “we have time travel” is not a recovery strategy
Consider a finance dashboard built from an orders fact table, customer dimensions, an authorized view, scheduled SQL, a semantic model and a BI refresh. Undeleting the dataset may recover the core tables, yet the report can still fail because an authorization must be re-established, a materialized view needs refreshing, a downstream job was disabled or the BI connector is pointing at an unexpected object.
Google specifically lists recovery limitations: linked dataset subscriptions may need to be recreated; business tags are not restored; authorized views, datasets and routines may require reauthorization; materialized views need manual refresh; and restored CDC-enabled tables do not automatically resume background apply jobs. Dataset search results can also take time to reflect the restored resource.
A useful recovery plan defines “restored” as a business state, not merely an API success response. It should prove that the right data is available, the right users can access it, sensitive users cannot, critical transformations run and published reports reconcile.
Define recovery objectives by business impact
Recovery point objective: how much data can you afford to lose?
Recovery point objective (RPO) is the maximum acceptable gap between the last recoverable data state and the incident. For an hourly operations dashboard, an RPO of several hours may be tolerable. For financial closing tables, a completed and signed-off daily snapshot may be more useful than the most recent unvalidated load. For a near-real-time customer workflow, minutes of missing updates may be unacceptable.
Do not confuse time travel duration with RPO. A seven-day window tells you how far back some historical states remain accessible, not that every source transaction is recoverable or that a particular snapshot exists at the required point.
Recovery time objective: how quickly must the service return?
Recovery time objective (RTO) is the maximum acceptable outage duration. The recovery clock includes detection, triage, authorization, data restore, dependency repair, reconciliation, report refresh and sign-off. A two-hour RTO is meaningless if only the table-copy step has been timed.
Assign targets to services, not only datasets. A revenue reporting service may depend on three datasets, a transformation repository, scheduled jobs, service accounts and the BI semantic layer. The slowest critical dependency often determines actual recovery time.

A practical tiering model
Tier 1: Board reporting, statutory reporting, customer-facing analytics and operational decisions with material impact. Define named owners, explicit RPO/RTO, protected recovery copies, immediate alerting and frequent restore drills.
Tier 2: Departmental reporting and important recurring analysis. Use scheduled protection, documented rebuild instructions and periodic recovery tests, with longer tolerances.
Tier 3: Exploratory, temporary and reproducible datasets. Prefer source re-ingestion or transformation reruns when economical, and document any exceptions involving irreplaceable data.
These tiers are illustrative, not universal thresholds. Set the final targets with business owners who understand the consequences of missed decisions and incorrect numbers.
Inventory the assets that must survive deletion
Start with a service-to-data dependency map. Record project, dataset, region, table and view inventory, data classification, retention setting, table expiration rules, partitioning, clustering, encryption configuration, source lineage and downstream consumers. Include the last successful load and last trusted reconciliation date.
Then inventory operational dependencies outside the dataset: SQL and Dataform repositories, scheduled queries, orchestration definitions, connectors, service accounts, IAM bindings, authorized resources, policy tags, BI models, dashboard credentials, alert rules and external exports. Version control can reconstruct definitions, but only if the definitions are actually committed and deployable.
Keep a small recovery manifest outside the dataset being protected. If the only inventory table lives inside the deleted dataset, the incident removes the map needed to repair it. Store the manifest under separately controlled access and ensure the recovery team can reach it during an outage.
Choose a layered protection architecture
Layer A: Fast operational undo
Use BigQuery’s configured time travel and dataset undelete for recent accidental deletion or erroneous changes. Standardize a time travel policy appropriate to criticality and cost. Establish a fast incident path that identifies the original region and checks the deletion timestamp before anyone recreates a dataset.
Layer B: Scheduled known-good recovery points
For selected critical tables, create scheduled snapshots in a protected dataset with explicit expiration and retention. Use names that encode the source, timestamp and environment. Verify that each snapshot job completed and that expected row counts and business totals are sensible. A snapshot that captures corrupt data faithfully is still a bad recovery point.
Separate permissions so ordinary pipeline identities cannot delete both production data and its recovery copies. Review the snapshot limitations, especially streaming buffers and metadata. For tables that cannot be snapshotted, define another suitable recovery mechanism.
Layer C: Independent copies or source replay
Where a shared-project or shared-organization incident, malicious deletion or extended retention requirement matters, consider independent copies or exports under separately governed access. Evaluate location, encryption, retention, legal requirements, export format, re-import cost and the operational ability to restore. A second copy that shares the same broad delete permissions may offer little protection.
Source replay is often the most economical recovery path for deterministic derived tables. Preserve raw inputs and transformation code, with immutable batch identifiers and reproducible business logic. However, a source API that only retains a short history or a connector that cannot replay deletes is not a dependable rebuild source.
Layer D: Recovery of business service
Protect deployment configurations and dependency definitions, not just data bytes. Maintain scripts or infrastructure-as-code for dataset settings, views, authorized resources, orchestration, scheduled jobs and monitoring. Keep a concise post-restore checklist that verifies access and business totals before re-enabling dashboards.
Actiknow’s business intelligence consulting and integration services cover the data integration, modeling, reporting and refresh work that surrounds the warehouse. Recovery planning should treat those components as one service boundary.
An incident runbook for an accidentally deleted dataset
Step 1: Declare the incident and freeze risky changes
Open an incident record with time, affected project, dataset ID, region, suspected cause, downstream services and incident commander. Stop automated deployment or cleanup processes that might recreate the same dataset ID or delete recovery assets. Preserve audit logs and avoid running speculative repair scripts in production.
Step 2: Determine eligibility before creating a replacement
Confirm when deletion occurred and the time travel setting in force at that moment. Verify the original region and the IAM permissions of the restoration identity. Google warns that only the most recently deleted dataset for a given ID can be undeleted. Recreating a dataset with the same name can remove the ability to undelete the original, even though some tables might still be recoverable.
Step 3: Restore the dataset using the supported operation
For an eligible dataset, follow Google’s documented UNDROP SCHEMA or datasets.undelete procedure, explicitly targeting the correct location where necessary. The following is an illustrative SQL shape, not a command to run without reviewing the production project and region:
UNDROP SCHEMA `example_project.finance_mart` OPTIONS(location=”us-central1″);
Check current SQL syntax, permissions and region support in the official documentation before execution. A name collision, wrong location or expired window can produce ALREADY_EXISTS or NOT_FOUND errors. Never treat a failed undelete as permission to overwrite an existing dataset.

Step 4: Repair non-data dependencies
Review restored IAM bindings and principals, then reauthorize views, datasets and routines as needed. Refresh materialized views, inspect linked dataset subscriptions, restore missing business tags and restart CDC background apply behavior only after validating its requirements. Re-enable scheduled jobs in a controlled order, with writes paused until you know the target data state.
Step 5: Reconcile before releasing reports
Compare recovered row counts, distinct business keys, date coverage, source control totals, financial sums and late-arriving updates against a trusted baseline. Confirm the intended recovery point and document any data gap. Test access for both authorized and unauthorized roles. Run critical transformations in dependency order, refresh BI models and have the business owner sign off on representative reports.
Step 6: Reopen service and learn
Resume ingestion and downstream jobs with idempotent or deduplicated processing so retries do not double-count. Monitor freshness, job failures, unexpected deletes and dashboard accuracy. Record actual detection time, restore duration, lost-data interval, manual interventions and any gap between the documented runbook and what happened.
Recovery validation: prove the numbers, not just the objects
A recovered dataset can look complete while containing wrong business results. Use at least three levels of checks. Structural checks compare expected objects, schemas, partitions, access settings and required dependencies. Data checks compare row counts, key uniqueness, null rates, date coverage and control totals. Service checks verify scheduled transformations, dashboard refresh, representative user access and agreed report metrics.
For a finance mart, a useful acceptance test might compare total booked revenue by accounting date, distinct invoice IDs, reversal totals and the latest closed period against a separately stored ledger extract. For a customer analytics mart, compare distinct customer keys, segment populations and opt-out exclusions. These are illustrative test designs, not claims about any Actiknow client.

If you need an observability model for such checks, Actiknow’s guide to data pipeline observability explains freshness, volume, quality, lineage and downstream business reconciliation. The same signals should become recovery acceptance criteria.
Prevent recovery copies from sharing the original failure mode
A backup dataset that every production service account can delete is vulnerable to the same accidental or malicious operation as the source. Use least-privilege roles, separate administrative identities, change approval for destructive operations and alerts for dataset deletion and permission changes. Consider organization policies and controlled break-glass access appropriate to the environment.
Test permissions in both directions. Can the recovery identity restore without unnecessary broad administration? Can the ordinary pipeline identity delete a snapshot, alter its expiration or remove the protection policy? Can an analyst query sensitive restored data without the original row-level controls? Recovery must not accidentally bypass governance.
Protect encryption keys and the people or systems that administer them. A retained copy is not useful if its required key becomes unavailable. Keep the ownership, escalation route and access process current as employees and vendors change.
Cost planning: compare protection with rebuild expense
Storage is only one component of recovery economics. Include snapshot retention and changed-block storage, copies or exports, cross-region movement, scheduled query costs, monitoring, engineering effort, restoration testing and the business cost of downtime. Compare those costs with the time and reliability of rebuilding data from upstream sources.
A simple decision model is: annual protection cost versus expected annual incident loss plus the cost of an unplanned rebuild. The incident-loss estimate should reflect report downtime, delayed decisions, staff time, data reconstruction, regulatory exposure where relevant and potential customer impact. Do not claim a universal payback period; the economics vary sharply by data criticality and replayability.
For physically billed storage, time travel and fail-safe storage can create separate charges. Snapshot storage can grow after updates or reclustering. Track actual usage and review current pricing rather than applying a generic percentage to table size.
A four-week implementation plan
Week 1: Map risk and define targets
Identify the five to ten most business-critical datasets or reporting services. Capture their owners, regions, time travel settings, dependencies, last trusted recovery points and current ability to replay sources. Agree on RPO and RTO for each service and name the person authorized to declare a recovery complete.
Week 2: Implement protective controls
Standardize time travel settings and alerting. Configure protected snapshots or independent copies for selected tables. Version critical SQL, orchestration and permission configuration. Restrict destructive privileges and establish a break-glass recovery identity. Document exceptions for external tables, streaming data and objects that cannot be snapshotted.
Week 3: Rehearse in an isolated environment
Delete a disposable test dataset and exercise the documented restore process within its configured window. Separately restore a table from a snapshot, reconstruct a view and re-enable a representative scheduled transformation. Verify correct location, access policies, report totals and downstream refresh. Record real elapsed time and every manual step.
Week 4: Operationalize and report
Fix runbook gaps, establish quarterly or risk-based restore drills, and add a recovery-readiness dashboard. Track protected datasets, last successful recovery copy, snapshot age, tested RTO, unresolved permissions, failed protection jobs and owners without a recent sign-off. Review the policy after material architecture changes.
Executive recovery readiness scorecard
For every Tier 1 service, leadership should be able to answer yes to these questions:
- Is its owner named?
- Are the RPO and RTO approved?
- Do we know the exact time travel window?
- Are protected copies recent and readable?
- Is the recovery identity tested?
- Are dependencies and security settings documented?
- Have we successfully restored in an isolated environment?
- Can the business reconcile the restored numbers?
- Are alerts and escalation routes live?
- Is the next drill scheduled?
A service that passes only the “we have a snapshot” question is not recovery-ready. The scorecard is intentionally about evidence: last successful test date, observed restore duration, measured data gap and documented acceptance, not confidence expressed in a meeting.
Common mistakes that make a small deletion a major outage
Recreating the deleted dataset immediately can interfere with undelete eligibility. Assuming fail-safe is a normal self-service restore path can consume precious incident time. Keeping backups under the same unrestricted delete identity can defeat separation. Snapshotting only tables while ignoring views, permissions and orchestration can produce a technically successful but unusable recovery. Restoring the latest data without checking whether it was already corrupted can reintroduce the original problem.
The remedy is straightforward but disciplined: preserve the timeline, choose the correct recovery point, protect copies independently, restore dependencies, reconcile business totals and practice before a real incident.
Frequently asked questions
1. Can I undelete an entire BigQuery dataset?
Yes, when it is eligible within the configured time travel window. BigQuery supports UNDROP SCHEMA and the datasets.undelete API, subject to permissions, location, name-collision and recovery limitations.
2. How long after deletion can a dataset be recovered?
The dataset’s time travel duration at deletion controls the ordinary undelete window. Google documents a configurable two-to-seven-day range, with seven days the default. Confirm the actual dataset setting.
3. Does BigQuery fail-safe extend dataset undelete by seven days?
No. The additional fail-safe period concerns certain deleted table data and uses an emergency support-assisted recovery process. It is not an ordinary extension of dataset undelete.
4. What happens if someone recreates the dataset with the same name?
Google warns that recreating a dataset with the same ID can prevent undeleting the original dataset. Some table-level recovery may still be possible. Investigate before creating replacements.
5. Will dataset undelete restore all permissions and metadata?
It recovers many dataset objects, properties and security settings, but documented exceptions include business tags, some linked dataset relationships, authorized resources and materialized-view refresh requirements. Verify each dependency.
6. Are table snapshots a replacement for dataset backups?
No. They preserve eligible table contents at chosen points, not every dataset object or operational dependency. Combine them with configuration backups, permissions documentation and service recovery procedures.
7. Can I snapshot views and external tables?
Google’s table snapshot documentation says snapshots cannot be created directly from logical views, materialized views or external tables. Protect their definitions or underlying source data through other mechanisms.
8. Can I recover an overwritten table?
Time travel can help recover a prior table state within the configured window. A suitable earlier snapshot or independent copy can provide a longer-lived recovery point.
9. What is the difference between RPO and RTO?
RPO measures the acceptable data gap. RTO measures the acceptable time to restore usable service. Both should be agreed with business owners and validated through drills.
10. Should backup copies be in another region or project?
That depends on residency, threat model, cost and operational requirements. Separation of permissions and failure domains matters more than simply adding a second dataset name. Cross-region copies have their own costs and governance implications.
11. How often should recovery drills be performed?
Use a risk-based cadence. Critical services merit more frequent exercises and tests after material changes. Every drill should measure elapsed time, recovered data state, permission integrity and business sign-off.
12. How do we know a recovered dashboard is trustworthy?
Reconcile recovered data with independently stored control totals, verify critical business metrics and freshness, test security boundaries and confirm the BI refresh and downstream jobs operate correctly.
Conclusion: recover the reporting service, not merely the dataset
BigQuery’s dataset undelete and time travel features make recent accidental deletion less catastrophic, but their short windows and operational limitations are not substitutes for recovery engineering. A sound policy pairs native undo with protected recovery points, reproducible transformations, controlled access, tested runbooks and business-level validation.
The strongest proof of resilience is not the existence of a backup. It is a recent, documented exercise showing that the organization can restore accurate, authorized reports within an agreed time and data-loss tolerance.
If your team is reviewing BigQuery recovery readiness alongside data pipelines and executive reporting, contact Actiknow to discuss your analytics architecture. The starting point can be a focused review of critical datasets, dependencies, recovery objectives and the evidence needed to test them.

