IMD SolutionComputing LibraryShop main store

Selection guide

Plant historian selection by scale and support needs

Edge computing and IIoT Published 9 min read

Short answer

Pick the tier first: a control-room or automation-vendor historian for one site's process trends, an enterprise platform only when several sites share one operational history, and a self-built time-series stack only when your team owns connectors, dashboards, backups and upgrades. Size by stored events per second — the sum of each tag group's tags times its stored events per second — times retention, never by tag count alone.

On this page
  1. Which historian tier fits your plant?
  2. Who maintains the historian after go-live?
  3. How many stored events per second will your tags produce?
  4. How much archive storage does three years of retention need?
  5. Why does compression mode change the sizing?
  6. What should the proof of concept prove before you sign?
  7. Where should the historian server physically run?
  8. Questions asked

A historian purchase is decided by three questions, in this order: which tier fits, who will run it, and what event rate it must carry. Tag count is the last input that matters, and a vendor roadmap is not an input at all until the first three are settled. A single site that needs trustworthy operator trends is served by the historian already integrated with its control platform. Multiple sites that must share one operational history justify an enterprise platform. A self-engineered time-series stack is a legitimate third tier, but it turns storage into an ongoing engineering product your team owns. This article gives the decision rule for each tier, a worked sizing example that ends at 440 stored events per second and about 416 GB of three-year archive, and the acceptance tests that separate a real selection from a slide deck.

Which historian tier fits your plant? #

Choose the tier before any product name: a control-room or automation-vendor historian for one site's process trends, an enterprise platform only when several sites must share one operational history, and a self-built time-series stack only when your team can own its connectors, dashboards, backups and upgrades. The table below is the whole selection in one view; each row is a different product you buy and a different set of duties you keep.

Tier Fits when Your team runs afterwards Deciding sizing input
Control-room or automation-vendor historian (the historian integrated with your SCADA platform, or a control-room product such as AVEVA Historian) One site: operator trends, local reporting Configuration, user accounts, backup checks Stored event rate and retention
Enterprise operational platform (the tier PI occupies) Several sites sharing one operational history; cross-site and corporate consumers Data governance, site roll-up design Aggregate event rate across sites and the WAN path
Self-built time-series stack (InfluxDB or QuestDB with Grafana dashboards) A team that accepts engineering ownership for custom analytics Connectors, dashboards, permissions, backups, upgrades Stored event rate and dashboard query load

The decision rule that follows from the table: start in the control-room row when the scope is one site's process history, and move up only when a second site or a corporate consumer of shared data appears, because shared operational history across sites is the requirement that justifies the enterprise tier, not plant size. A single large site can stay in the control-room tier indefinitely. The self-built row is entered by ownership, not by features: it is chosen when your team wants the analytics freedom more than it wants the vendor's maintenance.

Ask every candidate one selection question: which capabilities are released and available in the quoted edition and configuration today. A capability that exists only on a published roadmap is a requirement left unresolved; keep it out of acceptance testing until it ships in a version you can install.

Who maintains the historian after go-live? #

Support ownership is the deciding difference: a packaged historian comes with vendor maintenance, documented interfaces and a trend client, while a time-series database leaves collection drivers, quality handling, dashboards, permissions, backups and upgrade procedures to your own engineers.

Write the duty list with a named owner on each line before you compare products: collection interfaces and their controller and driver updates, user administration, dashboard changes, backup and restore drills, version upgrades, and first-line diagnosis when a trend comes up empty. A packaged historian moves engine defects and interface documentation to the vendor and leaves the rest of the list with you. A self-built stack moves every line to you, starting with collection itself, for which bridging OPC UA PLC data to MQTT with an edge PC is the common first-duty pattern.

The decision: if no engineer can be named against each self-built duty, the self-built tier is out regardless of feature fit. If a packaged candidate's support arrangement cannot answer a dead trend client in the middle of a night shift, it fails the selection regardless of its list price. Cost belongs last in this comparison: subscription and credit-based licensing models make sticker comparisons unreliable, so request a written quote for your tag, site and retention scope instead of comparing published numbers.

How many stored events per second will your tags produce? #

Size by stored events per second, not by tag count: add up every tag group's tag count multiplied by its stored events per tag per second, and the 3,500-tag example plant in this article lands at 440 stored events per second. In one line: stored events per second = sum over tag groups of (tags in the group × stored events per tag per second). The worked example below is a single-line plant; replace each row with your own inventory from the controller and SCADA tag lists, then check the rates against the storage mode configuration in the next section.

Tag group Tags Stored events per tag per second Stored events per second
Slow analogs: levels, temperatures (periodic, 10 s sample interval) 1,200 0.1 120
Fast analogs: control-loop flows and pressures (periodic, 1 s) 300 1 300
Discrete and status (stored on change) 2,000 0.01 20
Total 3,500 — 440

The arithmetic, group by group: 1,200 tags × 0.1 events/s per tag = 120 events/s; 300 tags × 1 events/s per tag = 300 events/s; 2,000 tags × 0.01 events/s per tag = 20 events/s. The plant total is 120 + 300 + 20 = 440 stored events per second, which over a day is 440 events/s × 86,400 s/day = 38,016,000 stored events, about 38 million per day.

Two inputs decide these numbers: the storage mode configured per group, and counts read back from a running archive during the proof of concept. Polling frequency is not the input. For the self-built tier, the same group-by-group numbers feed the sizing walkthrough for a self-hosted Node-RED, InfluxDB and Grafana historian.

How much archive storage does three years of retention need? #

Archive volume equals stored events per second, times seconds per day, times retention days, times bytes per stored event: at 440 events per second, 3 years of retention and an assumed 10 bytes per event, that is about 416 GB of archive, so budget a 500 GB class volume. Run the example numbers through the formula:

  • 440 events/s × 10 bytes per stored event = 4,400 bytes/s
  • 4,400 bytes/s × 86,400 s/day = 380,160,000 bytes/day, about 380 MB per day
  • 380 MB/day × 1,095 days (3 years) = about 416 GB of archive

The 10 bytes per event is an illustrative planning input, not a constant: it moves with timestamp resolution, quality storage and the compression mode in the next section. Measure daily archive growth across the first weeks of the proof of concept, divide by the stored event count, and rerun this arithmetic with your measured value. Doubling the input to 20 bytes per event gives about 833 GB for the same example, which is why the formula, not the example figure, carries the decision.

Decision for the stated inputs: specify roughly 500 GB of archive storage for the three-year case, correct it with your measured bytes per event once the proof of concept reports one, and add the engine's index and backup space on top.

Why does compression mode change the sizing? #

Compression mode decides the stored rate, so read the mode configuration before you trust any per-tag number: periodic modes store at a fixed sample rate regardless of process change, while change-based modes such as swinging door, deadband and delta store only when the value moves by the configured amount. Both families are documented behavior, not sales claims. The tag historian in Ignition stores values in Periodic mode at fixed intervals based on the Sample Rate, and in Tag Group mode at the execution rate of the selected Historical Tag Group. AVEVA Historian documents rate-of-change (swinging door), deadband and delta storage modes.

The consequence for your estimate: a swing-door or deadband group on a slow analog can store far fewer events than its poll rate suggests, because a value that barely moves produces few stored events, while a periodic group stores at its sample rate even when the value never changes at all. Build the event-rate estimate from the configured mode of each group, then verify with event counts read back from the archive during the proof of concept. If measured counts differ from the estimate by more than a fifth, correct the input and rerun the storage arithmetic before signing.

What should the proof of concept prove before you sign? #

Make the proof of concept a measured acceptance test, not a demo: collect at your calculated rate for at least a week, open several concurrent trends, pull a full-resolution month of one fast tag, and verify catch-up after a deliberate collection outage before anyone signs. Numbered tests, each mapped to a failure you will actually meet:

  1. Ingest at your calculated rate — 440 stored events/s in this example, your number in yours — for at least one week, and compare daily archive growth with the storage estimate above.
  2. Retrieval under load: open trends on several clients at once and pull a full-resolution month of one fast tag. When a trend client fails to load, the documented checks in troubleshooting historian trend client connection failures isolate the fault before it becomes a support case.
  3. Recovery: stop collection for one hour, restore the path, and verify buffered catch-up with correct timestamps and quality flags and no gaps.
  4. Restore: put yesterday's backup on a spare machine and record how long the restore takes.

The qualified candidate is the one that passes all four at your event rate with your tags. A demonstration on synthetic data at a fraction of your rate tests nothing you will run. If the project replaces a historian that is already running, treat history migration as its own subproject; backfilling PI from an existing plant historian shows the pattern.

Where should the historian server physically run? #

Run the control-room historian on a server or virtual machine next to the SCADA host, and run a self-built or lightweight edge stack on a dedicated fanless industrial mini PC in the line-side cabinet, because the tier you chose in the first section already decides the host. The control-room and enterprise tiers want a server-class or virtualized host beside the SCADA system, with the archive on server storage and backups leaving the site. The self-built and lightweight tiers want a cabinet box chosen for sustained disk writes, storage endurance and cabinet temperature range rather than peak processor marks; 440 stored events/s is a modest ingest that the storage subsystem defines, and when your tier lands there, the fitting hardware class is an industrial mini PC sized by the archive volume above.

When collection must sit closest to the controller, a lightweight store can run on the controller PC itself; deploying Edge Data Store in Docker on a controller PC is a documented pattern for that layout.

Decision for each case: one site with local trends — control-room historian on a server beside the SCADA host. Several sites sharing one history — site-level buffering that forwards to the enterprise platform, sized on aggregate stored event rate. A team that wants owned analytics — a self-built stack on a cabinet-rated box, with the archive estimate from the storage section and the four acceptance tests as the gate before anyone signs.

Questions asked

How many years of history should we keep online?

Derive retention from the longest question anyone must answer with the data. Two years online covers most quality investigations and year-over-year energy comparisons; regulatory, warranty or insurance cases add a cold archive measured in decades. Write the retention requirement per tag group, then multiply each group's rate and bytes per event into the storage estimate.

Can a time-series database replace a packaged historian?

It stores timestamped values well, but the packaged product is the surrounding application: collection drivers that understand controller data types, quality codes, compression, trend clients, user permissions, backup tooling and upgrade paths. Choose the database route only when your team accepts those duties as ongoing engineering work, with a named owner for each one.

Should every site write to one central historian?

Buffer at each site and forward upward. Local storage keeps operator trends alive when the wide-area link fails, and the enterprise tier becomes the shared system of record without depending on one plant's uptime. Assign the official long-term copy per data set in writing, because a value stored twice with different compression is two different histories.

What happens to old data when we change historian products?

Plan it as its own project. Export formats and quality handling differ between products, so not every archive transfers cleanly; keep the old system readable during an overlap period, migrate the tags that feed live analysis, and backfill only what a current use case needs. Very old data can stay frozen in its original archive format.

Keep reading

  • How-to

    Backfilling PI from an existing plant historian

    Split the historian-to-PI backfill into four independent stages — extract, intermediate store, external compression, destination load — and qualify throughput with a measured pilot. Rows divided by elapsed seconds, not tag count, sets the schedule: 12,000 tags from 2020 at 500 rows/s is about 203 days. Check the native Historian 2023 replication route before writing pipeline code.

    Edge computing and IIoT

  • Explainer

    Sizing a self-hosted Node-RED InfluxDB Grafana historian

    Self-hosted on the open-source editions, the software licence cost is zero: a 600-point historian polled once a second tops out at 600 point-samples per second, 51.84 million a day. Size the SSD from a measured pilot week against that ceiling, and treat change-only logging as margin — at this rate capacity, not endurance, sets the drive.

    Edge computing and IIoT

  • How-to

    Deploying Edge Data Store in Docker on a controller PC

    Run Edge Data Store as a Docker container on the controller-class edge PC and let its OPC UA adapter ingest the local server over localhost. Define egress targets, buffering and forwarding schedules in the JSON configuration, size the buffer from tag count, scan rate and outage target—50 tags at 1 event/s across an 8-hour outage is about 288 MB—and sign off only after a live pulled-cable test.

    Edge computing and IIoT