On this page
A working analytical engine is only one part of a product. Once people are using an application, the less visible responsibilities become just as important: upgrading it, diagnosing a fault, recovering from a bad release and preserving the information customers have already entrusted to it.
That led to a distinction in determinIQ which I now regard as fundamental: the state of a software release is not the state of a customer. Treating them as though they are the same makes recovery unnecessarily dangerous.
Two kinds of state
An application has versioned code, configuration of its release, dependencies and compatibility assumptions. A customer has users, organisational settings, uploaded material, reports, access arrangements and a history of analytical work. Reverting one should not automatically revert the other.
Suppose a new application release introduces a defect in a report view. The desired correction may be to restore the previous compatible version of the code. It would be unacceptable for that operation to discard a report created since the upgrade, undo an organisation setting or restore last week’s copy of an analytical Ledger.
The reverse is also true. If a dataset or database record is damaged, returning to an earlier plugin ZIP does not restore the missing information. Those are different incidents requiring different recovery authorities.
The boundary between rollback and recovery
The development approach therefore separates application release management from customer-data and infrastructure recovery. The update and rollback path concerns a verified code package and its compatibility with the state it will read. Database and storage backup belong to a separate operational responsibility.
Research figure 01
A release is not customer state
Code recovery and data recovery are separate responsibilities
- TWO KINDS OF STATEIdentify what failedApplication release or retained customer information
- RELEASE AUTHORITYVerify the application packageVersion, integrity and compatibility matter
- CODE ROLLBACKChange the compatible application versionKeep customer records and reports in place
- SEPARATE RECOVERYRestore data only under its own authorityA damaged dataset is not fixed by replacing code
- POST-CHANGE ASSURANCECheck continuity and accessValidate representative tenant and analytical journeys
This illustrates responsibility boundaries, not a required order of incident-response steps.
This separation is straightforward to say and more demanding to implement consistently. It affects migrations, report persistence, tenant settings, access controls and every assumption made about which state survives an update. A code-only rollback is safe only when the preceding release can still interpret the retained data and schema.
The documentation expressly treats an incompatible schema change or data corruption as a reason to stop calling the operation a simple rollback. At that point, a controlled recovery plan is required. This is precisely where a reassuring-looking “Restore” button can be misleading if it obscures which kind of state it will restore.
Integrity is necessary, but it proves something specific
A release package should have a stable identity. determinIQ’s documented build process includes a file-level SHA-256 manifest, a ZIP checksum and archive checks. Those controls help answer whether a staged package has the intended files and whether its bytes have changed.
They do not prove that the application behaves correctly on every dataset, that its host is trustworthy or that customer records are intact. This is why release verification includes representative application journeys rather than relying on a checksum alone.
The same discipline applies to an update that appears technically successful. A plugin may load while an important user journey is broken. Checks therefore need to cover access, prepared data, analyses, evidence, reports, sharing and the relevant privacy boundaries. A release is not finished merely because a version number has changed.
An operational contract is not a deployed disaster-recovery service
The maintained documentation distinguishes implemented application update and rollback functions from the larger responsibilities of infrastructure backup and disaster recovery. It specifies what a production backup and recovery design should cover, including customer state, protected uploads, organisation configuration and report history. It does not establish that an automated estate-wide disaster-recovery service, a tested recovery-time target or off-site backup policy is already provided by the plugin.
That difference matters. Writing down the recovery boundary is useful architecture; operating and testing the full recovery service is a further piece of engineering and assurance. It would be misleading to present one as proof of the other.
In a multi-tenant environment this distinction becomes even more important. A failure concerning one customer should not lead to an unnecessarily broad restore that changes unrelated tenants. Recovery must take account of ownership, isolation and consistency, not merely the availability of a ZIP or database snapshot.
Why I include operations in the research
Early product work naturally concentrates on whether the central idea can be made to work. Later work reveals a different question: can the resulting system be changed safely while people are relying on it?
The determinIQ research paper describes the transition from analytical architecture to an integrated product. The current codebase and operational documents show the corresponding concern with update integrity, rollback eligibility, diagnostics and the separation of authorities.
This study does not claim that the whole production operating environment has been independently tested or certified. It documents an architectural principle, supported application mechanisms and the remaining external recovery obligations.
The useful outcome is not a promise that software will never fail. It is a design in which the response to failure can be controlled without treating a customer’s history as disposable.
Connected research
The wider research programme examines how calculations, evidence and decisions remain governed. This study extends the same reasoning into product operation: evidence must survive presentation, and customer state must survive appropriate changes to the application that presents it.
Research context
Research programme: Making data analysis understandable, reproducible and accountable
Topic: Evidence and verification
Explore further
Supporting material is listed below.
Further reading is listed below.