Restore benchmarks: measure the time until the business can work
Throughput is one component of recovery. Metadata, identity, keys, and application validation can dominate elapsed time.
“Several gigabytes per second” sounds like a fast restore until the team starts rebuilding the service around the copied files. Time the dependency setup, data transfer, validation and application recovery together. The service owner needs to know when people can use it again.
Define the finish line
Choose a service and define the recovery-time objective with its owner. Separate data availability from application readiness and business acceptance. Record the recovery point and acceptable data-loss window independently.
Use a representative dataset: large files, small files, directory depth, permissions, and application metadata. A single large sequential file is an inadequate model for a namespace containing millions of small objects.
Establish a lower bound
A simple transfer-time calculation is useful as a sanity check:
100 TiB / 2 GiB per second = 51,200 seconds = about 14.2 hours
This is arithmetic, not a measured product result. It excludes initialization, metadata work, protocol overhead, retries, and application checks. Use consistent units and actual sustained throughput from the intended recovery route.
If the business requires recovery in four hours, a fourteen-hour byte-transfer lower bound already demonstrates a design mismatch. No dashboard wording can close that gap.
Time the complete sequence
| Milestone | Record |
|---|---|
| Recovery authorized | Start timestamp |
| Credentials and keys available | Access preparation delay |
| Destination ready | Provisioning delay |
| Namespace available | Initial mount or restore completion |
| Working set read | Cold-access completion |
| Application accepted | Business validation timestamp |
WEKA’s Snap-To-Object documentation notes on-demand recovery considerations. Similar distinctions should be investigated for any product that exposes data before all of it is local. Do not equate a quick mount with a fully recovered working set.
Avoid cache-driven optimism
Run the first representative read separately from repeat reads. Record whether source or destination caches were warm and whether object-store retrieval was involved. A second run can measure a different system path.
Measure file-count and metadata progress as well as bytes. Preserve failure and retry counts. A benchmark that silently skipped unreadable files can report excellent throughput while failing its purpose.
Validate data and authority
Compare a pre-recorded manifest, checksums for the test corpus, ownership, and expected access. Then execute an application transaction. Include a denied-access test for an unrelated user so the restored service does not trade availability for excessive exposure.
Report the actual result
Publish the dataset, tool versions, topology, elapsed milestones, throughput, errors, and manual interventions. Label estimates as estimates and lab results as lab results. Do not extrapolate a small trial linearly without explaining the limits.
Give the service owner the measured recovery time and the work that dominated it. If it misses the objective, that identifies what the next exercise needs to improve.