Build in Public

Share a SaaS Restore Test While Building in Public

Share a SaaS restore test with its recovery point, checked data, timing boundaries, and untested dependencies. Use a fictional example to keep the public…

By Uriel Bitton · · 9 min read

Illustrated cover for “Share a SaaS Restore Test While Building in Public”.

The short answer

Share a SaaS restore test by naming the selected recovery point, data restored, customer task checked, and work still unverified. Keep data age separate from recovery duration. A completed restore job or database check does not establish whole-product recovery, zero data loss, or a guaranteed future recovery time.

Share a SaaS restore test by naming the recovery point, the data restored, the customer task checked, and the work still unverified. Put those limits beside the result. A completed restore job can be useful progress without proving that the whole application can recover, that no data would be lost, or that a future incident would take the same time.

When you build in public, this work belongs in the product story before an emergency. You can show how you are examining a reliability promise, invite criticism of a specific gap, and return with the next finding. The public value comes from making the evidence understandable, including an unfinished result.

Say which recovery question the test answered

A backup being created, a saved resource being restored, and a customer finishing a task are different observations. Write down which one you have reached before choosing a headline. “We tested recovery” is too open-ended if the exercise stopped when a database became available.

AWS Backup's validation documentation makes a useful distinction: a restore job can reach COMPLETED before a separate validation workflow examines the restored resource. The documentation describes an AWS capability, not a test of your product. Its publishing lesson is that a job status and an application result need separate evidence.

For a founder update, identify the ordinary customer action that matters. Can a person open an existing project, retrieve its file, and complete the expected approval? Checking that a server answers a request may support a narrower statement than checking that whole task.

This article concerns reporting a planned exercise. During an actual interruption, affected customers need current impact and instructions first. The guide to communicating a SaaS outage while building in public covers that separate job. A progress post about yesterday's test should not become a substitute for today's incident notice.

Keep data age separate from recovery duration

There are two time questions in a restore update: how far back the recovered state reaches, and how long the chosen recovery work took. A quick restore of an old copy answers those questions very differently from a slower restore of a recent one.

Google Cloud's disaster recovery planning guide defines a recovery time objective, or RTO, as the maximum acceptable application downtime. A recovery point objective, or RPO, concerns the maximum acceptable time interval of potential data loss. These are planning objectives. One successful test duration does not establish a universal recovery promise.

In your record, keep the selected recovery point, test start, resource completion, and application validation completion as separate timestamps. Give the timezone. Explain anything excluded from the timer, such as choosing the recovery point or preparing the destination.

If the application never passed its chosen task check, there may be no full recovery completion time to announce. You can still publish the duration of the completed stage. Describe it as that stage, and keep the unfinished task beside it.

A fictional proofing SaaS shows the gap

Nina builds Page Otter, an invented SaaS for reviewing packaging proofs. Its customers would open a project, inspect a proof file, and leave an approval. In this paper exercise, project records, comments, and file references live in a database; the proof files themselves live separately.

The people, product, architecture, times, counts, and results below are fictional. No backup, restore, infrastructure change, production access, or customer-data inspection was performed for this article. The example uses an imagined synthetic dataset, not a real customer's work.

Nina's hypothetical exercise starts at 09:00 UTC using a database recovery point representing 08:45 UTC. Destination preparation happened before the timer began. The restore job finishes at 09:18. Checks of the restored database finish at 09:26. The six proof files have not been restored to the test destination.

Part of the exercise

Illustrative observation

What the public update can establish

Recovery point

Database state at 08:45 UTC; request at 09:00 UTC.

The selected state is 15 minutes older than the test request.

Restore job

Completed at 09:18 UTC.

The database restore job took 18 minutes from the request.

Database checks

At 09:26, all 10 expected projects and 120 expected comments match the synthetic fixture, including their project associations.

Those defined database checks passed 26 minutes after the request.

File references

Six expected references appear in the restored database.

The database contains those references; this does not establish that the files open.

Proof files and approval

Files were outside the exercise; the complete proof-review task was not validated.

Whole-product recovery remains unverified.

The database checks are stronger than a row count alone because the imagined check also compares expected records and associations. They still cover only their stated criteria. The table deliberately gives the missing files the same visibility as the successful database work.

Nina cannot turn 09:18 into “our SaaS recovers in 18 minutes.” Even 09:26 marks the end of database checks, not a completed customer approval. There is no validated whole-task finish time in this example.

The 15-minute gap also needs care. It describes the selected point relative to the request. No actual customer activity or loss was observed. The exercise does not establish that a production incident would lose exactly 15 minutes of work, or that every write before the selected point would be recoverable under different conditions.

Turn the record into a useful public update

Lead with the scope and result, then the limitation that changes its meaning. The following is model wording for Page Otter's fictional exercise. It is not a statement about Buildside, Uriel's infrastructure, or a real recovery test:

In our synthetic-data exercise, the database restore job completed in 18 minutes. Checks of the expected projects, comments, and associations finished 26 minutes after the request. The selected recovery point was 15 minutes older than the request. Proof files were outside this exercise, so we have not validated the complete proof-review task or a whole-product recovery time. Our next test will include those files and check whether an existing proof can be opened and approved.

The short version preserves the important qualification. If Nina shares only a cropped screenshot of the completed job, readers lose the distinction between database progress and product readiness. A written caption can explain that gap without requiring them to interpret a provider dashboard.

Keep the next test in future tense until it happens. A planned file check is useful evidence of the next decision, but it does not extend the completed exercise. Link the later result back to this update so someone can follow what changed.

Publish the method without exposing the recovery environment

Keep a fuller private record behind the public statement. Include the exercise date, relevant software versions, selected recovery point, dataset scope, defined checks, results, and timing boundaries. That record helps you answer a reader's question without relying on the wording of the social post alone.

For the public version, prefer a synthetic example or a carefully reviewed summary. Do not upload backup archives, customer records, credentials, internal access links, or raw logs just to make the work look inspectable. A readable explanation can show the method without granting access to the environment.

State when synthetic data limits the conclusion. A small fixture can help examine selected behavior, but it does not establish how recovery behaves with the size, dependencies, permissions, and history of a production workload. If you have not examined those conditions, keep them open.

A demonstration video needs the same care. Editing out a wait changes the viewer's impression of duration. Explain cuts and keep the recorded timer boundaries in the accompanying note. Use a sample project whose content is safe to show rather than blurring a live customer's screen after recording.

Ask peers about the missing dependency

A focused question gives readers a practical way to improve the next exercise. For Page Otter, “What else should we test?” is less useful than explaining that the database references exist while the corresponding proof files were outside the restore.

Nina could ask founders with file-heavy products which checks reveal a broken connection between restored records and separately stored files. Another useful question concerns the end of the timer: does opening a proof and saving an approval cover the relevant task, or does a later export belong in that definition?

Treat replies as suggestions to evaluate. A comment proposing a file checksum is not evidence that Page Otter has passed such a check. If the suggestion changes the next exercise, say what was added and credit the contributor appropriately. Report the result only after the defined check is complete.

Keep the discussion about decisions rather than asking strangers to inspect private production material. The useful public exchange can happen around the shape of a test, the meaning of a claim, and an intentionally shareable example.

A historical incident explains why ownership belongs in the story

GitLab's February 10, 2017 postmortem reported that logical database backups had been failing, failure emails were rejected, and regular recovery testing lacked an owner. GitLab also described sharing public recovery notes and streaming the procedure during the January 31 incident. These are GitLab's historical findings, not claims about its current systems.

The communication lesson is more specific than “be transparent.” A public account can connect an operational gap to an owned follow-up. Announcing a backup mechanism alone leaves unanswered who will check it, what successful recovery means, and how a failed check will reach someone responsible.

For a small SaaS, name the next exercise owner privately and make the public commitment proportionate. “We will report the file-inclusive test after completing it” is a clear follow-up. A guarantee that future recovery will always succeed would require evidence this example does not provide.

Return with the changed result

Your next post should explain what the new exercise established, including any remaining failure. Keep its date, conditions, and covered dependencies visible. Do not silently extend an earlier test's scope because the newer one went further.

If the earlier announcement was accurate at the time, link the later finding as an update. If you had described a database-only restore as whole-product recovery, correct that claim where it appeared. The guide to correcting a build-in-public update explains how to keep a correction connected to the original statement.

Before publishing your next reliability update, write the customer task at the top of your private test record. Under it, list what passed and what remains outside the exercise. Build the public claim from those two lists. That gives the next conversation a concrete starting point.

Questions founders ask next

Can I share a failed restore exercise?

Yes, if the account accurately describes the exercise, checked result, and remaining work without exposing private material. A failed test can support a useful progress update. Keep any actual customer incident and its required communication in the appropriate response process.

Should I call one test duration my RTO?

Keep the objective and observation separate. State the intended recovery objective and the duration of the defined exercise, including its start, finish, and exclusions. A database-stage timer does not establish a whole-application recovery time when the customer task remains unverified.

Does a synthetic test prove customer backups work?

A synthetic test establishes only what was checked under its stated conditions. It can help examine a method or selected behavior. Claims about actual customer backups require evidence about those backups and the relevant recovery conditions; do not infer that evidence from a fictional or synthetic example.

Sources

Frequently asked questions

Can I share a failed restore exercise?

Yes, if the account accurately describes the exercise, checked result, and remaining work without exposing private material. A failed test can support a useful progress update. Keep any actual customer incident and its required communication in the appropriate response process.

Should I call one test duration my RTO?

Keep the objective and observation separate. State the intended recovery objective and the duration of the defined exercise, including its start, finish, and exclusions. A database-stage timer does not establish a whole-application recovery time when the customer task remains unverified.

Does a synthetic test prove customer backups work?

A synthetic test establishes only what was checked under its stated conditions. It can help examine a method or selected behavior. Claims about actual customer backups require evidence about those backups and the relevant recovery conditions; do not infer that evidence from a fictional or synthetic example.

A note from Uriel Bitton

I share practical lessons from building Buildside and learning alongside SaaS founders.

Keep building with us

Get practical founder lessons and product updates in your inbox.

Subscribe or Join Buildside.

Keep reading

Browse more articles from Buildside