Explainer
Sizing a self-hosted Node-RED InfluxDB Grafana historian
Short answer
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.
On this page
- What does the software licence cost for a self-hosted historian?
- How many samples per second is 600 points at a one-second poll?
- Can Node-RED write to InfluxDB only when a value changes?
- How much disk does a year of this data need?
- Do the trend viewers change the sizing?
- What kind of host does the stack need?
- Is Docker required for Node-RED, InfluxDB and Grafana?
- Questions asked
Zero in software licences, and one number decides the hardware: 600 point-samples per second. Polling roughly 600 Modbus TCP and MQTT points once a second and storing every value produces 51.84 million samples a day, and Node-RED, InfluxDB OSS and Grafana OSS run that self-hosted with no licence fee. The real money goes to a PC-class host with an SSD, sized by retention, plus the time you spend backing it up. This page gives the write-volume arithmetic, the change-only policy that reduces it, and the disk and host sizing that follows, with a worked example you can repeat with your own pilot numbers.
What does the software licence cost for a self-hosted historian? #
Nothing: Node-RED, InfluxDB OSS and Grafana OSS are the self-hosted open-source editions, they carry no licence fee, and a 600-point one-second historian sits far inside what they are built for. The three tools install on one host and talk to each other locally; nothing in this workload touches a metered service.
What you pay for starts where you choose it: managed cloud hosting, support contracts and enterprise features. Grafana is also published in an Enterprise edition installed instead of OSS, and that is not the selection here. On the dashboard side, since Grafana v13.2 the InfluxDB data source is a standalone plugin, preinstalled in both Grafana OSS and Enterprise, so connecting Grafana to InfluxDB needs no add-on. The cost that remains is the host, the SSD, a backup target and administration time — the sections below size those.
How many samples per second is 600 points at a one-second poll? #
Six hundred point-samples per second: 600 points stored once every second is 600 samples/s, which is 51.84 million samples a day and about 1.56 billion across a 30-day month before any change filtering. The arithmetic is 600 × 86,400 seconds in a day = 51,840,000 samples, and × 30 = 1,555,200,000 samples.
That assumes every poll becomes a stored sample and collection runs continuously, so it is the conservative ceiling, not a measurement. Keep polling and writing separate in your head: Node-RED reads the registers and topics, and whether each read becomes a database write is a policy you set in the flow, covered next. Size disk against the ceiling until a pilot measures the stored rate. Note the shape of the load too — ingestion runs 24/7 while trend viewing happens a handful of times a day, and the write stream sets every hardware number that follows.
Can Node-RED write to InfluxDB only when a value changes? #
Yes, and it is flow design rather than a product feature: a comparison step in the Node-RED flow between the Modbus TCP or MQTT input and the InfluxDB write node decides which samples are stored. Selecting the stack does not create change-only history by itself; you build the rule and you own what it hides.
Set one of three write policies per point:
- Every poll — store every sample including repeats. Maximum volume, regular spacing, the easiest trends to read.
- On change — store only values that pass the comparison rule. Stable signals go quiet, and the history no longer shows how often they were polled.
- Change plus periodic sample — store changes and force an unchanged write at a fixed interval.
For the comparison rule, use exact equality on discrete signals such as bits, states and run flags, and a deadband on analog values — start at 0.5 % of span and tune after a week of real data. A tight deadband stores noise; a wide one hides real drift, so treat it as a per-signal decision and write the rule down where a later trend reader finds it.
Put the periodic sample at 300 s for on-change analogs: 600 points ÷ 300 s = 2 extra samples/s, about 0.3 % of the ceiling, and it makes a flat line mean "unchanged" rather than "not polling". The sizing decision that follows from all of this: plan disk on the full-poll ceiling, and treat whatever change-only logging removes as margin, not planned capacity.
How much disk does a year of this data need? #
Plan on measured growth with the full-poll ceiling as the bound: an illustrative 1.0 GB per day of database growth at 51.84 million samples a day is 365 GB a year, so a 1 TB SSD holds about two years with headroom. Those are example inputs, not a measurement of your system — the measurement is yours to take.
Run the production schema and retention settings for one representative week covering idle and running shifts, then read the disk growth. Worked example: a test database that grows 7.0 GB in 7 days is 7.0 ÷ 7 = 1.0 GB/day; over a year that is 1.0 × 365 = 365 GB, and two years is 730 GB. A 1 TB drive (1,000 GB, decimal) leaves 270 GB for the OS, database overhead and estimate error — call it two years of retention. Budget the same again on a second disk or host for the backup copy.
The example implies 1,000,000,000 bytes ÷ 51,840,000 samples = about 19 bytes per stored sample. Use that only as a sanity check on your own measurement: timestamp width, tag sets and compression move it, so never take a generic bytes-per-point figure as a sizing input.
Endurance is a non-issue at this rate: 1.0 GB/day of logical writes is 0.37 TB a year, far inside the endurance rating of any SSD you would put in an industrial host. Capacity and retention set the drive; endurance re-enters the conversation only when the point count or the poll rate gains a zero.
Do the trend viewers change the sizing? #
No: two or three users opening five trends a day is negligible next to a write stream that runs every second of every shift, so size for ingestion and retention first. Five query sessions a day is a few minutes of work set against 86,400 seconds of continuous writes each day.
The one exception worth designing for is a dashboard that aggregates months of raw samples every time it opens — summarize or downsample long windows rather than re-reading raw history each time. During the pilot, open the real dashboards while collection runs and watch host resources; at this scale you will find the write path, not the viewers, doing the work.
What kind of host does the stack need? #
A fanless industrial PC with an SSD, a current multi-core CPU and 16 GB of RAM runs this stack with margin, because at 600 samples per second the constraint is disk capacity, not processor speed. Ingestion at that rate is a small fraction of one core on any current fanless industrial PC, and 16 GB covers the OS, the database working set and Grafana with room to spare.
The non-negotiable is the drive: a historian writes continuously, and continuous small writes are the workload that wears SD-card storage out, so specify an SSD and check the host's documented storage interface for the capacity your retention needs. When you reach that choice, a fanless industrial mini PC with an M.2 or SATA SSD bay is the class of host this stack belongs on. Run a long-term-support Linux release or Windows IoT/LTSC, whichever your team already patches, and keep the database on the SSD alongside the OS.
Is Docker required for Node-RED, InfluxDB and Grafana? #
No: for one box on one site, a direct install on Linux or Windows is simpler to back up, patch and troubleshoot on shift, and containers only earn their keep once you repeat the deployment. Docker is a packaging choice, not a requirement, and every container layer is one more thing the person on shift has to know during a failure.
Take containers when a second and third site are coming and you want identical, versioned deployments. Whichever route you choose, two things are not optional: a second copy of the data that survives the host, and a resource dashboard watching disk use and write rate, so a filling disk or a creeping point count is caught before it stops collection.
Questions asked
How do I prove a point is still polling when I only log changes?
Write a heartbeat measurement once per poll cycle for each point group — a counter or timestamp, one write per cycle rather than per point. At a one-second cycle that is a handful of samples per second across 600 points, and a dead input shows up as a flat heartbeat instead of an ambiguous gap in the trend.
Which query language should I choose in Grafana's InfluxDB data source?
Grafana's InfluxDB data source offers both InfluxQL and Flux, and you pick one when you add the data source. The choice changes the query syntax on your dashboards, not the historian's write path or sizing, so pick the one your team can read and stay consistent across dashboards and any export scripts.
Can this historian run from an SD card instead of an SSD?
Not in production at a one-second poll. Continuous small writes are exactly the workload that wears SD-card storage out, and a historian never stops writing. Use an SSD in whatever interface the host documents, and treat high-endurance memory cards as trial hardware for the pilot only.
What happens to the disk plan if I enable change-only logging later?
Nothing needs to grow: change-only can only reduce writes, so a disk sized on the full-poll ceiling stays valid. Re-measure for a week covering idle and running shifts, then re-project retention from the measured rate and bank the difference as margin or extra retention.