Technical Brief · Data Engineering Perspective · Published: October 27, 2026
Water quality sensors are small producers. Each one logs around the clock, and a hundred of them quietly convert time into data at a rate most teams never stop to calculate. The total is smaller than IoT marketing implies and more consequential than “we’ll sort out storage later” assumes. Volume is trivial. The governance decisions around retention, compression, and downsampling are what decide whether that data becomes an asset your models can use or a compressed afterthought nobody can query.
Here are the four numbers that frame the conversation.
Table of Contents
Key Takeaways
- Number 1 — 1,440: one parameter at a one-minute interval produces 1,440 readings a day (60 × 24). That falls straight out of the sampling definition.
- Number 2 — ~2.1 million: a four-parameter instrument (pH, conductivity, ORP, temperature in one probe) generates about 2,102,400 readings a year at that cadence. Stored as 4-byte floats — the encoding used in published time-series capacity examples [4] — that is roughly 8.4 MB of values per instrument-year, still tens of megabytes with timestamps attached.
- Number 3 — under a dollar a month: a 100-instrument fleet produces around 210 million readings a year, which even with generous raw-data overhead stays in the low single-digit gigabytes. At S3 Standard pricing of $0.023 per GB-month for the first 50 TB [2], a full year of fleet history costs less than a monthly coffee. S3 Glacier Deep Archive sits at $0.00099 per GB-month [2].
- Number 4 — megabytes, not gigabytes, on the wire: a 4-byte payload is about 5.8 KB per instrument-day, so a hundred units move a couple of megabytes a day in binary and something in the low tens of megabytes once text framing is added. Cloud egress at $0.09 per GB after the first 100 GB free [3] stays a rounding error.
- The cost lever is architecture, not storage pricing: tiered retention (raw data short, aggregates long) is the standard industrial pattern [6], and downsampling plus compression routinely cuts long-term storage by an order of magnitude [5].
- Sensor choice shapes the pipeline. Instruments that publish synchronized multi-parameter readings over a digital interface eliminate the alignment and bookkeeping overhead that multiplies engineering cost long before storage shows up on a bill [1].
Number 1 — 1,440 Readings per Day, per Parameter
Start with cadence, because everything scales from it. A one-minute interval is a common and defensible choice for process water quality: fast enough to catch regeneration events, cleaning cycles, and dosing upsets, slow enough to keep volumes and radio duty cycles friendly. At one minute, one parameter yields:
1,440 readings/day = 60 × 24
That is the arithmetic of the definition, no source needed. The encoding of each reading is a sourced quantity: published IoT time-series capacity planning uses 4-byte float values per sample [4], and that is the number to carry forward.
Worth noting what a one-minute cadence does not capture: sub-minute transients. If your process needs them, you probably already know — and you will pay for them in every number below, proportionally.
Number 2 — About 2.1 Million Readings per Instrument-Year
A four-parameter probe measuring pH, conductivity, ORP, and temperature simultaneously produces:
4 parameters × 1,440 readings/day × 365 days ≈ 2,102,400 readings/year
At 4 bytes per value [4], the raw values total about 8.4 MB per instrument-year. Add timestamps and metadata and you are still in the tens-of-megabytes range before compression.
This is the number that surprises people. A full year of high-resolution, four-parameter history from one instrument is smaller than a smartphone photo library. The engineering significance is not the volume. It is that there is no excuse for not keeping it. Data discarded for storage reasons at this scale is a governance failure, not a cost necessity.
Modern time-series databases do better than raw floats. Published capacity examples show Gorilla-style compression achieving roughly 10:1 on smoothly varying sensor data [4], which brings a four-parameter instrument-year down to roughly a megabyte stored.
Number 3 — A 100-Unit Fleet Costs Less Than a Coffee per Month to Store
Scale the second number to a hundred instruments:
100 instruments × 2.1 million readings/year ≈ 210 million readings/year
With realistic raw-data overhead — timestamps, tags, indexing — call it a few gigabytes a year uncompressed, single-digit gigabytes with typical time-series compression. Now price it.
Amazon S3 Standard starts at $0.023 per GB-month for the first 50 TB, with volume discounts above that. S3 Glacier Deep Archive offers archival storage at $0.00099 per GB-month [2]. A 5 GB fleet-year at Standard rates is about $0.12 per month; archived at Deep Archive rates it is about half a cent. Multiply raw volume tenfold out of caution and the bill for a hundred-instrument fleet is still loose change.
The pricing nuance worth respecting is not the per-GB rate but the request and retrieval structure. S3 request charges run about $0.005 per 1,000 writes and $0.0004 per 1,000 reads on Standard [3], and cold classes carry retrieval fees and minimum durations. High-frequency small-object writes can cost more in requests than in storage — one more reason to batch readings into files or use a time-series layer rather than posting one cloud object per sample.
Number 4 — Transfers Measured in Megabytes, Egress Measured in Cents
Data motion is the other half of the ledger. At 4 bytes per reading in a binary payload [4], one parameter at one-minute cadence moves about 5.8 KB a day. A four-parameter instrument moves about 23 KB a day. A hundred-unit fleet lands near 2.3 MB a day in binary — call it the low tens of megabytes a day with JSON-style text framing, and comfortably under half a gigabyte a month for the whole fleet even with retries and handshakes.
Outbound transfer pricing makes even pessimistic cases cheap: after the first 100 GB per month free, AWS charges about $0.09 per GB for data transfer out [3]. A fleet uploading 1 GB in a month pays nine cents.
The lesson is directional. Bandwidth is not the constraint for water quality telemetry, so choosing sensors and architectures around bandwidth anxiety is a category error. Choose around data quality and synchronization instead — which is where digital multi-parameter output earns its keep, feeding clean, time-aligned streams to SCADA, historians, and cloud platforms [1].
The Governance Decisions That Actually Matter
If storage is cheap and transfer is cheap, why do organizations still drown in sensor data costs? Because the money leaks in architecture and governance, not in gigabytes. Three practices separate a tidy time-series estate from a messy one:
- Tiered retention. Keep full-resolution raw data only as long as it is operationally useful — days to weeks — then retain aggregates (min/max/mean per minute or hour) for years. Industrial time-series reference architectures follow exactly this pattern: raw high-frequency data for a short window, longer retention for coarser aggregates [6].
- Deliberate downsampling. Aggregating detailed points into lower-resolution summaries — hourly averages replacing per-minute records — is the standard way to cut time-series storage cost while preserving analytical value [5]. Do it by policy, not by panic.
- Compression before deletion. Modern time-series engines compress smoothly varying sensor data at ratios around 10:1 in published examples [4]. Compression first, downsampling second, deletion last, in that order.
| Number | Magnitude | Key assumptions | Governance action |
|---|---|---|---|
| 1,440 | Readings per day, per parameter at 1-min cadence | One-minute sampling interval (60 × 24) | Confirm cadence matches the decisions the data must support |
| ~2.1 million | Readings per 4-parameter instrument-year | 4 parameters × 1,440 × 365; 4-byte float values [4] | Never discard at this scale; keep raw + aggregates |
| ~$0.12/month | 100-instrument fleet storage for a year (≈5 GB) | S3 Standard $0.023/GB-month [2]; Deep Archive $0.00099/GB-month [2] | Adopt tiered retention; batch writes to control request costs [3] |
| ~10–12 MB/day | Fleet transfer volume, 100 instruments, text payloads | 4-byte readings [4]; text framing multiplies binary volume | Keep egress trivial; spend attention on data alignment instead [3] |
Why the Sensor Decision Still Dominates
None of these storage numbers argue for cheap instrumentation. They argue for smart instrumentation.
When four parameters arrive time-aligned from a single multi-parameter probe over a digital interface, the pipeline downstream is simple: one stream, one clock, one calibration story, one archive [1]. When the same four parameters come from four separately cabled single-parameter devices, the fleet’s data volume is identical but the engineering cost — alignment, correlation, failure isolation — multiplies.
Buy the sensor, own the outcome. The storage bill will be a rounding error either way. Whether those 2.1 million readings a year are usable is decided at the probe, not in the cloud.
References
- Shanghai ChiMay — Shanghai ChiMay 4-in-1 Sensors for Industrial Process Control. Product documentation for multi-parameter sensing (conductivity, pH, dissolved oxygen, ORP in one platform) with digital output for unified calibration and data correlation. https://www.chimaycorp.com/shanghai_chimay_4-in-1_sensors_for_industrial_process_control/810.html
- Amazon Web Services — Amazon S3 Pricing. S3 Standard starts at $0.023 per GB-month for the first 50 TB; S3 Glacier Deep Archive at $0.00099 per GB-month. https://aws.amazon.com/s3/?did=ft_card2&trk=ft_s3
- CostGoat — Amazon S3 Pricing Calculator. Request pricing of about $0.005 per 1,000 PUT and $0.0004 per 1,000 GET on S3 Standard; data transfer out of $0.09 per GB after the first 100 GB free per month. https://costgoat.com/pricing/amazon-s3
- EmbeddedRoad — Time-Series Data in IoT: Characteristics & Storage. Worked capacity examples using 4-byte float samples and Gorilla-style compression at roughly 10:1 on smoothly varying data. https://embeddedroad.com/branches/iot-connectivity/iot-connectivity-time-series-data-characteristics-and-storage-needs
- The Computer Science Professor — Implementing Data Downsampling in InfluxDB to Reduce Storage Costs. Explains downsampling as aggregation of detailed points into lower-resolution summaries to cut time-series storage costs. https://thecomputerscienceprofessor.com/implementing-data-downsampling-in-influxdb-to-reduce-storage-costs
- Cohera — Time Series Data Management for Industrial IoT. Reference tiered-retention architecture: raw high-frequency data for a short window, longer retention for coarser aggregates, cold tier on object storage. https://cohera.co/insights/time-series-data-industrial-iot
About the author: Written by the Shanghai ChiMay Technical Editorial Team — the team behind Shanghai ChiMay’s online water quality analyzer and multi-parameter sensor documentation, covering Modbus/SCADA integration, historian and cloud data pipelines, and instrumentation for AI-ready water models.
