After a production line stops, the investigation may depend on records spread across a SCADA archive, a drive event log, and a meter export. The engineer needs to establish what changed first, which values are trustworthy, and whether the relevant interval is complete. If those records use different clocks or hide communication gaps, assembling a trend becomes part of the investigation itself.

That is a useful starting point for choosing an industrial data historian. The selection should begin with the evidence your plant needs to retain and retrieve. Define the operational questions, trace each required measurement from its source, and agree on tests for timing, quality, interruption recovery, and retrieval. Then compare products against that specification. A tag count or a list of supported protocols provides only part of the information needed for a purchase decision.

This approach also gives the existing system a fair evaluation. OPC UA’s Historical Access model recognizes that historians vary in functionality while sharing requirements around timestamped history and status information. It does not establish a universal boundary between an archive embedded in SCADA and a separately deployed historian. The practical question is whether the proposed configuration preserves and returns the information required for your workflows. [1]

The guide below follows that evaluation from requirements to pilot acceptance. It explains where source timestamps matter, why stored and calculated values need different treatment, and how to test recovery without assuming every interruption can be repaired. It also separates ordinary process history from evidence that requires specialized source recording. The result is a specification your engineering, operations, and IT teams can use to assess the same deployment, with clear responsibilities and measurable conditions for acceptance.

Define the operational questions your historian must answer

Separate history from control

A historian retains operational records for later access. In the cited OPC UA model, those records include timestamps and status information, while historical events can contain additional fields. For procurement, define which measurements, states, and events belong in your historical service and how users will retrieve them. Do not assume that selecting a historian also defines how process control will be implemented. [1]

Start with a short collection of tasks. An electrical engineer may need to align feeder measurements with equipment trips. An energy manager may need comparable consumption and production periods. A process engineer may need to retrieve the conditions around an off-specification batch. These are proposed evaluation scenarios, each requiring a different combination of resolution, context, and retention.

Give each task an owner who can judge the output. Ask that person to identify the required signals and events, the investigation window, and the uncertainty the decision can tolerate. For a stop investigation, that could mean distinguishing events separated by a fraction of a second. For an operating-period comparison, it could mean reproducing a calculation with the same input records and rules.

Translate tasks into acceptance criteria

Convert each task into an observable result before preparing the request for proposal. State what the user should be able to retrieve, how long it should take under the agreed workload, and what must happen when data is unavailable. These are project requirements to negotiate, rather than universal specifications for every historian.

For example, require a selected equipment trend to show unavailable intervals distinctly and allow the engineer to export its underlying records. If a report excludes bad-quality data, require the output to disclose that treatment and the affected duration. If a remote connection fails, identify which records should remain recoverable and which failure conditions sit outside the proposed protection.

Use these outcomes to organize the source inventory. Include asset identity, measurement name, unit, source interface, timestamp origin, update policy, quality information, and ownership. Where an attribute is unknown, keep it visible as a question for the pilot. Replacing an unknown with an assumed value makes it harder to determine why a later acceptance test failed.

Compare SCADA history, a dedicated historian, and a database stack

Compare capabilities across deployment options

Evaluate complete configurations using the same criteria. An embedded archive, a separate historian, and a custom database-based service are deployment options to investigate. Their labels alone do not establish capacity, recovery behavior, usability, or fitness for a particular plant. The comparison below is a procurement framework, not a ranking of those categories. [1]

Evaluation area Existing SCADA history Separately deployed historian Database with custom services
Acquisition Verify the installed connectors and collection policies. Verify source coverage and any required gateways. Identify who builds and supports collectors.
Timing and status Test what survives the current archive path. Test mapping and query semantics end to end. Define the record model and mapping rules.
Interruptions Locate buffers and test failure boundaries. Inspect collector, link, and archive recovery separately. Establish persistent queues and recovery ownership.
Retrieval Test operator, engineering, and external access. Test investigations, reports, and application interfaces. Include user tools and APIs in scope.
Operations Review dependencies on the existing deployment. Assign administration and integration responsibilities. Budget for maintaining the complete service.
Expansion and exit Request upgrade and export conditions. Request growth, migration, and export conditions. Confirm documentation and continued maintainability.

If the current installation passes the essential tests, retaining or extending it remains a valid option. When a gap appears, describe it precisely: an unavailable source, an inadequate archive window, an unsupported export, or a retrieval task that fails the agreed performance target. That gap gives the next architecture decision a clear purpose.

Trace responsibility through the data path

Draw the path from each source through collection, buffering, storage, query processing, and the consuming application. At each step, identify who configures the behavior and who investigates a failure. Include any intermediate service that changes timestamps, scales values, filters records, or calculates aggregates.

Research on an industrial temperature-monitoring pipeline provides an example of separating ingestion, quality evaluation, and visualization. The implementation used distinct services to store readings and assess their quality. It supports checking these responsibilities separately; it does not establish a universal architecture or demonstrate that one database category outperforms another. [7]

For the buyer, the practical test is ownership of the entire workflow. If the proposal supplies storage but leaves acquisition, quality interpretation, and user access to another team, include that work in the evaluation. Ask for a named owner and an acceptance condition for every required component before comparing the total deployment cost.

Specify time, quality, and storage fidelity

Preserve the meaning of timestamps and quality

OPC UA distinguishes the timestamp assigned by the data source from a server timestamp reflecting receipt or knowledge of accuracy. These fields should not be treated as interchangeable. Establish which time drives the historical trend, which time drives calculations, and whether additional arrival information is retained. The cited historical configuration makes server-timestamp storage optional, so confirm the deployed behavior. [2][4]

When evaluating other acquisition paths, establish where timestamps are assigned and how delayed records are recognized. A protocol or connector name does not answer those questions. The MQTT and OPC UA integration comparison provides context for defining the handoff into a historical workflow.

Quality requires an equally explicit mapping. OPC UA defines status information that indicates a value’s usability. Request an example covering valid data, an uncertain measurement, and a source failure. Verify what appears in storage, exports, trends, and reports. A failed reading must remain distinguishable from a valid measured zero. Also establish the limits of the source’s own quality reporting; a retained status field does not independently verify sensor accuracy. [2]

Distinguish collected samples from returned values

Specify source updates, collector acquisition, storage policy, and query processing separately. OPC UA’s historical configuration describes periodic intervals, exception deviation, and interpolation mode. These settings affect which records exist and how returned history should be interpreted. Define the required behavior for analog signals and discrete states, then test a known input sequence. [4]

A raw historical query returns the records saved in the archive. Under OPC UA Historical Access, those records can already reflect compression and storage rules. Calling an export raw does not prove that it contains every measurement generated upstream. Compare the export with a retained reference sequence to establish what was acquired, filtered, saved, and returned. [3]

Ask separately about lossless storage compression and settings that omit samples. Include a short excursion, a steady interval, and a rapid change in the reference sequence. Where fast disturbances matter, require evidence that the source and acquisition path can capture them. A one-second sample sequence, by itself, cannot establish what occurred between samples. State explicitly when the investigation needs dedicated event or waveform records instead.

For calculated outputs, specify the averaging method, time boundaries, interpolation, and treatment of non-Good data. OPC UA’s TimeAverage2, for example, excludes non-Good duration from its averaging interval. That is a defined method, not a universal default for every product. Ask the supplier to reproduce your expected result from a small test dataset before accepting a report calculation. [6]

Consider a simplified test in which a load stays at 100 kW for nine minutes and then at 200 kW for one minute. The arithmetic mean of those two levels is 150 kW. With valid data and that known stepped behavior throughout the interval, the time-weighted mean is 110 kW: (100 × 9 + 200 × 1) / 10. Both numbers can look plausible on a report. The acceptance dataset must establish which calculation the report is required to return.

Have something in mind to discuss?

We’re here to help you find the answers. Let’s talk.

Book a call

Design for interruptions and verifiable recovery

Place buffers at the relevant failure boundary

Test a failed source connection separately from a failed link between the collector and archive. OPC UA Historical Access distinguishes interruptions that can recover from interruptions that leave gaps. It also specifies reporting of lost historical data. Use that distinction when defining the protection expected from your configuration. [5]

A source-side buffer can preserve acquired records while a downstream link is unavailable, provided acquisition, power, and storage remain available. A buffer after the failed link cannot preserve records that never reached it. This is an architectural inference from the interruption boundary, not a promise that every collector provides persistent buffering. Mark the proposed buffer location on the data-flow drawing and test the actual path. [5]

Size the buffer against a declared workload and outage scenario. As an illustrative calculation, 5,000 stored records per second over two hours requires capacity for 36 million records. That count does not establish the required disk space. Measure the serialized size, persistence overhead, available storage, and growth allowance in the proposed implementation.

Define what happens when capacity is nearly exhausted, when the process restarts, and when the collector loses power. Ask how operations staff will detect a growing backlog or stopped acquisition. Keep unrecoverable conditions visible in the specification. The acceptance test should identify the boundary of protection instead of treating buffering as an unconditional guarantee.

Test replay and restoration separately

After reconnecting, verify whether the backlog retains its intended measurement times and status. Send controlled duplicate and out-of-order records to establish the documented handling. Run live collection during recovery and measure how long the backlog takes to clear. Agree in advance whether the report for the affected interval can change after delayed data arrives.

Restoration addresses another failure class. NIST’s OT backup guidance includes the supporting files, tools, configurations, software, and license dependencies needed to recover an operational environment. For a historian evaluation, request a restore that allows a user to retrieve an agreed historical interval and reproduce a required report. [9]

Perform this exercise on an isolated test system and record the recovery duration, missing dependencies, and checks on the restored output. NIST’s OT security guidance specifically recommends testing restoration during contingency-plan testing. Treat this as implementation guidance, not as an international historian certification requirement. [8]

Preserve the restore instructions with the acceptance evidence. An archive that can be copied but cannot be interpreted by the recovered application does not satisfy a requirement to restore an engineering workflow. Specify who maintains the instructions after upgrades, source changes, or modifications to the calculation configuration.

Make history usable for investigations and reports

Reconstruct a stop without overstating causality

Use a controlled investigation scenario to evaluate retrieval. The following records are illustrative, not a customer case or a product export. All times refer to the same day in UTC. Before interpreting event order in a real installation, establish clock alignment, acquisition delay, and timing uncertainty across the relevant sources. Timestamp display resolution alone does not establish that uncertainty. [2][8]

Source time Arrival time Record Value and status
10:14:21.000 10:14:21.080 Motor current 42 A, valid.
10:14:22.000 10:14:22.090 Motor current 76 A, valid.
10:14:22.650 10:14:22.700 Drive event Overcurrent trip, valid event.
10:14:22.700 10:14:22.750 Run feedback False, valid.
10:14:24.000 10:14:24.100 Feeder voltage Unavailable, communication failure.
10:14:25.000 10:16:05.200 Feeder voltage 400 V, valid buffered record.

Subject to those timing checks, the sequence supports investigating the current increase, drive diagnostics, and mechanical load around the stop. It does not establish the physical root cause. The delayed feeder record belongs at its source time when the selected trend uses that clock. Its later arrival remains useful for investigating the collection path.

Ask the engineer to export the evidence and explain the uncertainty. The unavailable feeder reading must remain visible as a gap or invalid interval under the agreed presentation rules. Neither the current samples nor the isolated voltage reading can rule out a short electrical disturbance between measurements. That question requires a recording path with suitable resolution.

Expose quality and calculation coverage

Research on conveyor temperature data distinguishes completeness of individual record fields from completeness over time. A populated value column did not, by itself, establish a continuous measurement history. The study used fixed collection expectations; its findings should not be applied as a universal missing-record rule for exception-based archives. [7]

Make this distinction part of the user test. For periodic collection, compare observed records with the expected sequence. For change-based collection, verify the defined storage policy and independent communication-health information. Explain how users distinguish an unchanged value from a failed source, especially when opening a long-range trend or exporting data to another application. [4][7]

Give reports their own acceptance dataset. Include known valid periods, a gap, and records arriving after the first calculation. Check the result and its coverage information against the declared method. If an hourly value represents only the valid part of an hour, the report should disclose that basis under your specification. Avoid silently equating a calculated value with a directly measured observation. [6]

Keep the research boundary clear: the cited pipeline study does not validate every machine state or thousands of concurrent streams. Its relevance here is the separation of acquisition and quality assessment. Performance, operating-state behavior, and user workflows still require testing against your plant’s sources and intended deployment. [7]

Size the workload and compare lifecycle costs

Measure ingestion and retrieval together

Prepare a workload profile with active signals, expected change rates, stored intervals, retention, concurrent queries, and recovery traffic. As a simple illustration, 2,000 signals stored once per second produce 172.8 million records per day. This assumes every scheduled value is written. It is an arithmetic input to sizing, not a capacity benchmark or a disk-space estimate.

Use a representative dataset for the proposed configuration. Include a recent trend query, a longer comparison period, an export, and a scheduled report. Run those tasks while collection continues, then repeat the relevant checks during backlog replay. Define acceptable response times and collection behavior before observing the results.

Specify detailed retention separately from aggregate retention and interruption buffering. The OPC UA historical model distinguishes archive availability and configured storage windows; those settings do not establish a backup policy. Ask what remains directly queryable, what requires restoration or another retrieval step, and what is deleted when each retention rule takes effect. [4]

Before accepting a sizing estimate, request its assumptions in writing. Include storage overhead, replication, backups, indexes, temporary query space, and growth headroom where applicable. Require the supplier to identify which figures were measured, which were calculated, and which need validation during the pilot. This keeps commercial comparisons tied to the same workload.

Include ownership, security, and exit costs

Request a lifecycle schedule covering collectors, gateways, archive services, redundancy, user access, APIs, storage, support, upgrades, and migration work. Ask each supplier to define a licensed tag, connection, client, and server in the offered terms. These are commercial questions to resolve, not assumptions about how all historian products are priced.

Include the operational boundary in the same review. NIST OT guidance addresses access control and managed configuration changes. Translate those practices into proposed permissions for viewing records, changing collection settings, administering services, and modifying historical data. Identify required network flows and supporting services with the plant’s OT and IT teams. [8]

Make export and handover part of the purchase discussion. Request a representative export with measurement identity, units, timestamps, and quality information. Have another tool interpret it using the documentation supplied. Ask what additional services, licenses, or transformation work would be needed for a future migration, and include that answer in the comparison.

Finally, name the owner of collection health, failed connectors, storage warnings, restore procedures, and configuration records. Agree on a support escalation path and the information needed to diagnose a missing interval. These responsibilities should remain explicit whether the deployment is embedded in the existing system, separately operated, or assembled from several services.

Try CENTO in action. Run our demo.

Launch demo to explore CENTO features or book a guided walkthrough.

Launch demo

Use a pilot to make the selection

Set failure tests before demonstrations

Give shortlisted suppliers the same source sample, reference records, tasks, and acceptance thresholds. The proposed tests below translate the cited timing, historical-access, gap-reporting, and recovery principles into an evaluation plan. Perform disruptive tests in an approved test environment, with expected outcomes agreed before execution. [2][3][5][9]

Test Controlled condition Evidence to retain Acceptance question
Acquisition Replay a known signal sequence. Input and returned records. Were required changes retained?
Timing Delay delivery of known records. Source and arrival times. Does the selected historical clock remain correct?
Source failure Stop source-to-collector communication. Status and gap output. Is unavailable data distinguishable from valid data?
Link failure Interrupt forwarding while acquisition continues. Buffer and replay records. Does recovery match the specified protection?
Collector restart Restart with a known backlog. Before-and-after record comparison. Does the required buffer survive?
Calculation Query a known interval containing invalid data. Expected and returned results. Are method and coverage correct?
Concurrent load Query and report during replay. Timings, counts, and queue measurements. Are agreed operating limits met?
Restore and export Recover on an isolated system and export. Recovery steps and interpreted output. Can the required workflow be reproduced?

Avoid treating a successful demonstration with a small, clean dataset as evidence for untested workloads. Record the software version, configured policies, hardware or service resources, source behavior, and observed results. Keep the failed checks and their resolutions alongside successful results so the final decision reflects the deployment actually tested.

Document the decision and remaining conditions

Apply non-negotiable requirements before scoring preferences. If an essential investigation depends on a quality indicator or recovery behavior that the configuration does not provide, record the requirement as unmet. A proposed weighted score for price and convenience should not conceal that result.

For configurations that pass, compare lifecycle cost, user effort, support responsibilities, expansion conditions, and migration options. Explain why each criterion matters to the plant. Mark an item verified only when there is evidence; use separate labels for a documented commitment and a test still required before acceptance.

Finish with a short decision record: the chosen scope, accepted limitations, unresolved dependencies, named owners, and final acceptance conditions. This gives procurement a concrete basis for the order and gives engineering a reference for commissioning. It also establishes which changes would require another evaluation, such as adding faster sources, extending retention, or supporting a new reporting workload.

Learn how CENTO turns this architecture into real operational results

CENTO combines historical data workflows with industrial monitoring and reporting. Its Historian product description covers acquisition, buffering, storage, and historical access. [10] Use the criteria here to evaluate a specific CENTO configuration with your sources and interruption scenarios. For wider projects, review the SCADA modernization guide and the integration comparison. Bring a source inventory, retention requirement, and one investigation or reporting task to the evaluation. Explore the demo or request a guided walkthrough to establish which deployment and acceptance tests fit your plant.