---
title: "Industrial AI Data Feeds: Sampling, Queues and Missing Changes"
description: "Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20."
canonical: "https://www.leviathansystems.co/articles/industrial-ai-opcua-sampling-queue-publishing"
last-updated: "2026-10-20"
---

# Industrial AI Data Feeds: Sampling, Queues and Missing Changes

Leviathan Systems — Technical content
Published 2026-10-20; updated 2026-10-20.

Compare an industrial AI data feed by the observations and changes it actually delivers for the intended task. An advertised update interval does not establish independent fresh process observations or a retained change sequence. Source updates, server sampling, filters, item queues and publication determine different parts of that coverage. Two feeds with the same publishing cycle can return the same latest value while preserving different histories.

## Choose the information the data task needs

A latest-state dashboard and a dataset for studying brief changes can need different inputs from the same industrial Variable. The dashboard might need the newest available value at each refresh. The dataset might require the observed sequence within that refresh period. Calling both feeds a 500 ms update service leaves that difference unresolved.

Start the comparison with the intended consumer's question. Does it need the latest retained value, or every qualifying captured change over a defined period? If the latter, specify what qualifies and what observation coverage the task requires. Complete history of captured changes and complete history of physical process behavior are different claims.

The protocol example here uses OPC UA Variables and their DataValues, with meanings checked in OPC10000-4, Part 4: Services, v1.05.07 on October 20, 2026. Its MonitoredItem model separates sampling interval, monitoring mode, filter and queue. Events have separate semantics; the DataValue queue and overflow discussion below is confined to Variable monitoring.

References: [OPC4](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.13.1.1)

## Sampling a source does not create fresh source data

Clause 5.13.1.2 describes the assigned sampling interval as a best-effort cycle. If the server does not support the requested interval, it assigns an appropriate interval and returns it to the client. The requested number therefore leaves the assigned interval and actual observation behavior to establish.

The clause also describes access to a decoupled underlying system. That system may update its data more slowly than the OPC UA server samples it. Repeatedly reading an unchanged source image does not create independent new observations of the physical process. Unsynchronized source-update and server-sampling cycles can add detection delay.

For an offered feed, distinguish the underlying source's update behavior from the server's requested and revised intervals. A claim about one of those stages cannot establish the others. Also establish whether the offered source is periodically sampled or supplies exception-based changes; the same clause accommodates both models.

References: [OPC1](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.13.1.2)

## A captured sample still has to become a retained Notification

The MonitoredItem overview explains that filters determine whether monitored Variable changes generate Notifications. Clause 5.13.1.5 then places qualifying Notifications in the item's queue for Subscription transfer. A deliberately excluded change never enters that queue.

A finite full queue introduces another selection boundary. For DataValues, the configured policy either removes the oldest Notification or replaces the last queued value. A queue of one instead keeps the newest Notification and ignores the discard policy. For a queue larger than one, discarding a DataValue sets the specified Overflow bit in a retained DataValue's statusCode.

The absence of that greater-than-one overflow indication on a queue of one is no evidence that every earlier change survived. Nor would such an indication identify an observation missed before queueing. Interpret the evidence at the stage it describes.

References: [OPC4](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.13.1.1), [OPC2](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.13.1.5)

## Four changes can produce one returned value or four

Fictional example OPC-AI111 follows one fixed identified industrial Variable, EX-PV1. Its preceding publishing cycle is at 0 ms and its next return is stipulated at 500 ms. Exactly four distinct sample changes occur: V1 at 100 ms, V2 at 200 ms, V3 at 300 ms and V4 at 400 ms. Each passes the filter and immediately produces one Notification. V1 through V4 are value labels, rather than physical units or thresholds.

Both offers begin the interval with an empty item queue. There is no initial Notification, extra change or other queued item. Monitoring and publication are enabled. Sufficient Publish requests, message limits and transport are stipulated to permit the complete available queue to reach the client at 500 ms, with no downstream loss. These are invented actual sample and return times; they are not timing promises derived from requesting a 100 ms sampling interval.

Offer Latest has queue capacity one. After 100 ms it holds V1; after 200 ms it holds V2; after 300 ms it holds V3; after 400 ms it holds V4. At the stipulated return, only V4 remains available. Three earlier Notifications have been displaced.

Offer Sequence has queue capacity four. Its retained contents grow from V1 to V1, V2, then V1, V2, V3, and finally V1, V2, V3, V4. It returns those four in item-queue order. No overflow occurs under these exact assumptions.

Both offers return V4 as their newest value on the same cycle. Only Sequence also supplies the earlier three captured changes. Latest may satisfy the dashboard's stated latest-state need. It cannot satisfy an unchanged requirement for all four stipulated changes. Queue capacity four is sufficient for this example alone; the example supplies neither a general queue-sizing rule nor recommended settings.

References: [OPC2](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.13.1.5)

## A larger queue cannot recover an unobserved pulse

Consider a separate invented process timeline with observations only at 0 ms and 100 ms. The value is at its baseline at both observations. A pulse begins at 40 ms, ends at 60 ms and returns fully to that baseline. No intervening observation or other record is supplied.

Neither observation contains the pulse. Even an unlimited downstream store of those two observations would contain no observation of it. Copying the baseline into additional rows would increase the row count without supplying the missing evidence.

This differs from OPC-AI111: its four changes were observed and qualified, but Latest discarded three Notifications before sending. It also differs from intentional filtering, where an observed change produces no Notification. These mechanisms leave different gaps. Their separation is a reasoning aid for evaluating coverage, rather than a measurement of a product or a prescription for process sampling.

## Publication and retransmission act on what reached them

Clause 5.14.1.1 describes the publishing interval as a Subscription execution cycle that attempts to send a NotificationMessage. Sending uses available Publish requests; with Notifications waiting but no request, the server waits for one. That cycle alone establishes no delivery deadline to an AI consumer.

The same clause describes a separate retransmission queue for sent NotificationMessages and notes that profile support can make it optional. Its message sequence numbers concern missing messages. They do not count every physical change or establish that every qualifying Notification reached publication.

In OPC-AI111, retransmitting the message containing V4 would still provide only V4 from Latest. The earlier three Notifications were displaced before that message was formed. In the pulse example, the missing observation never existed in the supplied input. Message recovery cannot reconstruct either omission. Any separate historian or other recording source would need its own coverage evidence.

References: [OPC3](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.14.1.1)

## Compare the delivered sequence with the consumer's requirement

Use the four-change case to make an offer's return concrete. Have the data provider explain which of V1 through V4 its proposed feed would make available under the stated conditions, and why. A latest-value display is a useful answer to a latest-state question. It leaves a four-change dataset requirement unanswered if it shows only V4.

Then connect that explanation to the offered product and the consumer's actual input. The unresolved issues include the source observation coverage, changes excluded by filtering, retention when publication is delayed, and whether later delivery or storage keeps the sequence or reduces it to the latest value. A server-side queue explanation alone does not establish what reaches the model's dataset. Resolve those specific gaps against the task, rather than infer coverage from an update-rate label.

A transient-feature task can depend on an earlier change absent from the newest value. More GPU throughput cannot recover that missing input history. Reconstructing a possible trajectory would be an inference, rather than the observed sequence the buyer requested. Determine the offered capture and retention coverage before treating additional compute as the remedy.

Once values are delivered, asset attribution and time interpretation still matter. The related operational digital-twin guide explains asset identity and changing relationships; the Redfish guide separates report time, metric time and resource-health meaning. The trace-sampling guide addresses selected diagnostic populations. Those decisions complement this feed-retention question without supplying its missing observations.

## Connect a separate physical cabling need to its service scope

If the equipment project also needs physical network cabling, Leviathan's Structured Cabling & Fiber Installation service covers agreed fiber and copper installation against approved pathways, endpoint schedules and labeling rules. Use the related project inquiry with a nonconfidential outline of the equipment, physical connections, supplied materials and required connection records to discuss scope, requirements and availability.

The data-feed comparison is educational. Feed software, OPC UA integration, industrial automation and operating decisions belong to the separately appointed platform and process teams. A physical cabling installation does not establish the observation or retention coverage discussed here.

References: [L1](https://www.leviathansystems.co/services/structured-cabling)

## Primary sources

- [OPC1 — OPC Foundation - OPC10000-4 v1.05.07, 5.13.1.2: Sampling interval](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.13.1.2). Checked 2026-10-20. Assigned best-effort intervals, underlying source updates and unsynchronized observation cycles. Applies to the Variable sampling discussion.
- [OPC2 — OPC Foundation - OPC10000-4 v1.05.07, 5.13.1.5: Queue parameters](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.13.1.5). Checked 2026-10-20. Notification retention, DataValue overflow and per-item return order. The worked example applies these rules to stipulated inputs, without transferring Event behavior.
- [OPC3 — OPC Foundation - OPC10000-4 v1.05.07, 5.14.1.1: Subscription description](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.14.1.1). Checked 2026-10-20. Publishing attempts, Publish-request availability and sent-message retransmission. Message delivery is a later stage than item-queue retention.
- [OPC4 — OPC Foundation - OPC10000-4 v1.05.07, 5.13.1.1: MonitoredItem overview](https://reference.opcfoundation.org/Core/Part4/v105/docs/5.13.1.1). Checked 2026-10-20. Separates sampling, mode, filtering and queueing, including filter-dependent Notification generation for monitored Variable values.
- [L1 — Leviathan Systems - Structured Cabling & Fiber Installation](https://www.leviathansystems.co/services/structured-cabling). Checked 2026-10-20. Physical installation to agreed pathways, endpoint schedules and labels; useful for a separate equipment-connection requirement.

## Related Leviathan resources

- [Operational digital twins: asset identity, time and relationships](/articles/operational-digital-twin-identity-time-relationships)
- [Redfish telemetry: report time, metric time and resource health](/articles/redfish-metric-report-time-versus-resource-health)
- [Sampled AI traces and request-metric populations](/articles/ai-observability-trace-sampling-versus-request-metrics)
- [Structured Cabling & Fiber Installation](/services/structured-cabling)
- [Discuss a separate physical network-cabling project](/contact?service=structured-cabling&from=%2Farticles%2Findustrial-ai-opcua-sampling-queue-publishing)

## Discuss a separate physical cabling project

Share a nonconfidential equipment, physical-connection and handoff outline to discuss installation scope, requirements and availability.

[Discuss the scope](/contact?service=structured-cabling&from=/articles/industrial-ai-opcua-sampling-queue-publishing)