Direct Answer
A leader asks why one database cannot hold devices, points and alarms together. The dual-database division is decided by data characteristics: the time-series database stores high-frequency point history (high write throughput, compression), and the relational database stores device profiles, users and work orders (strong transactional consistency); the two are linked by device ID, and forcing them into one database sacrifices performance on one side. The FEXCloud IoT cloud platform organizes data in a "time-series + relational" dual-database architecture; data characteristics decide where storage lives.
A scene: the first version of a campus energy platform used one relational database — point writes slowed the platform down, and half a year later a history curve query took more than ten seconds — after switching to the dual-database architecture, writes became fast, curves opened instantly, and device profiles and work orders were managed as before. One database cannot hold both kinds of data: high-frequency point data needs the time-series database's write and compression; device profiles and work orders need the relational database's transactional consistency — different data characteristics require separate storage.
1. Load Characteristics of the Two Data Types: One Write-Heavy, One Change-Heavy
Point data (time-series) — high-frequency writes, almost no modification: every device produces a value every second or minute — writes dominate, and reads are mostly history-curve queries; Profile/business data (relational) — low-frequency writes, frequent modification: device profiles, user permissions, work-order status — transactional operations dominate, one change needs locks and consistency; Write-pattern difference — time-series data is "append-only": appends are the most natural pattern; business data changes again and again: frequent update/delete; Read-pattern difference — time-series data is "scan a time range": the curve of a device over a period; business data is "find one by condition": one work order, one user.
In one sentence: point data "only appends and scans by time", profile/business data "changes and queries by condition" — one write model, one transaction model, two kinds of storage by nature.
A load example: 300 devices at 50 points each with a 5-second cycle — 3000 point records written per second — business data changes only a few hundred times a day — one database must both absorb 3000 appends per second and guarantee the consistency of work-order changes — two jobs in one place, the database does neither well. Load characteristics decide the storage division: high-frequency appends go to the time-series database, low-frequency transactions to the relational database, each doing its strength.
2. The Time-Series Database: Write Throughput and Compressed Retention
Write throughput — the time-series database is designed for "high write": batch writes, sequential disk writes — tens of thousands per second is normal, appends never lock tables; Compressed retention — time-series data is partitioned by time and compressed: values with similar trends are stored compressed — after the raw-value retention period, downsampling keeps long-term curves queryable; Retention strategy — "hot data at full precision, cold data at reduced precision": the latest 30 days at raw precision, earlier data aggregated at 5-minute or 1-hour intervals — storage cost stays controlled; Query strength — scanning a time range: curve queries and trend aggregations are the time-series database's specialty — millisecond returns.
In one sentence: the time-series database's signature skills are "writes a lot, stores cheaply, queries fast" — batch writes, time-partitioned compression, curve scans by time — point history was born for it.
A time-series example: 3 billion point records in one year — with daily partitioning and compression, the time-series database used less than 200 GB — the latest 30 days at raw precision, earlier data aggregated at 5-minute intervals — any day's curve opens in seconds. The time-series database's compression makes "affordable to store" — 3 billion raw records compressed to 200 GB, history curves available whenever you want.
3. The Relational Database: Device Profiles and Business Transactions
Profile data — device ledgers, user permissions, project configuration: low-frequency writes, frequent modification — structured and relation-constrained; Transactional consistency — work-order status flow: create → process → close — every step either succeeds or rolls back; no intermediate state allowed; Relational queries — devices to points, devices to alarms, users to permissions — queried by foreign keys; business reports rely on it; Strong consistency — ACID transactions: several tables change together in one operation — the relational database's signature skill, which the time-series database does poorly.
In one sentence: the relational database's signature skills are "transactional consistency and clear relations" — device profiles, work-order flows, permission management — the strong consistency business operations need is what it provides.
A relational example: a work-order flow: alarm triggered → assigned → processed → accepted — every step updates the work-order status, writes an operation log and notifies people — several tables change in one transaction — the relational database guarantees "all succeed or all roll back". Business transactions need strong consistency — work-order flows touch multiple tables; the relational database's ACID keeps every step error-free.
4. Dual-Database Consistency: Whose Side Wins
Device ID is the bridge — point records in the time-series database carry device IDs; the device profile's primary key in the relational database is also the device ID — the two databases are linked by device ID; Consistency strategy — "profiles are the authority, data is the axis": whether a device exists and what it is named follows the relational profile; how much a point value is follows the time-series data — each side is authoritative in its own area; Deletion and change — device retirement: the relational database marks it retired, the time-series database archives its history by device ID — profile first, data second, avoiding orphan records; Synchronization order — a new device is profiled: write the relational database first (device ID generated) → then time-series data writes — this order keeps both sides complete.
In one sentence: the dual database does not demand "the same record exists on both sides" — device profiles follow the relational database, point data follows the time-series database — linked by device ID, each manages its own authority zone.
A consistency example: a device retired — operations marked it retired in the relational database — the time-series database archived the history by device ID — the platform view: the profile shows retired, the history curve still queries — each side manages its own, no contradiction. The key to dual-database consistency is "dividing authority" — profiles belong to the relational database, data to the time-series database; device IDs tie them together without conflict or redundancy.
5. The Minimum Viable Combination for Small and Medium Projects
Sizing by scale — small projects (a few dozen devices): a single-instance time-series database plus a relational database (or built-in) suffices — no cluster needed; Medium projects (a few hundred devices): dual instances of the time-series and relational databases — one server or cloud-hosted both work; Database selection — choose a mature open-source or cloud-hosted time-series product (spec by write volume), and a mainstream relational database (by business complexity); Evolution path — a smooth path from small to large: start with dual databases on one machine → when volume grows, scale or split — the architecture is not torn down.
In one sentence: small and medium projects need no complex architecture — two instances, a time-series database and a relational database, are the minimum viable combination — spec by device count and business complexity, and scale smoothly as it grows.
A combination example: 200 devices on a campus — dual databases: a cloud-hosted time-series database (billed by write volume) and a single-instance relational database — one deployment, one maintenance visit per month — when devices grow to 500, scale the time-series database only. The minimum viable combination is not "simplest" but "sufficient and able to grow" — start with dual databases on one machine; when volume grows, scale the time-series database without redesign.
6. Forcing Them Together: Three Real Costs
Cost one: writes drag down business queries — one database carries both high-frequency writes and business transactions — writes eat the disk IO and work-order queries slow down; Cost two: compression and transactions conflict — time-series needs compressed partitions, business needs transactional locks — the two mechanisms fight inside one database; Cost three: the model becomes awkward — point data needs "scan by time", business needs "query by condition" — one table model serves neither well, and SQL gets stranger.
In one sentence: three costs of forcing them together — writes drag queries, compression conflicts with transactions, the model becomes awkward — the question is not "can it work", it is "neither side works well".
A cost example: the first version of a platform used one database — after three months, point data filled the disk IO — work-order queries went from 1 second to 10, curve queries took over ten seconds — after splitting into dual databases, each returned to its strength. The cost of forcing them together is "both sides slow down" — data characteristics differ, and splitting is not an option, it is a necessity.
7. Dual-Database Ops: No Data Loss, No Slow Queries
Backup strategy — the time-series database backs up by partition (cold data backed up less frequently); the relational database uses full plus incremental backups — backup rhythm follows data temperature; Monitoring metrics — the time-series database watches write latency and disk water level; the relational database watches connections and slow queries — each database has its own gatekeeper; Archiving and cleanup — the time-series database auto-downsamples and expires by retention policy; the relational database archives old work orders periodically — storage cost stays controlled by the retention policy; Disaster recovery — the time-series database reconnects and catch-up-syncs after interrupted writes; the relational database does primary-replica switchover — recovery order: business first (relational), then history catch-up (time-series).
In one sentence: dual-database ops each manage their own strength — the time-series database watches writes and disks, the relational database watches transactions and slow queries — backup, monitoring, archiving and recovery, each done by the database's load characteristics.
An ops example: the time-series database's disk water level hit 80% — by the retention policy, raw data older than 90 days was downsampled to 5-minute aggregation — the water level returned to 60%, and history curves remained queryable — "temperature partitioning" in dual-database ops: cold data should be reduced when due. Ops is not equal effort — the time-series database manages storage by temperature, the relational database guarantees consistency by transactions; each runs at its own rhythm.
8. Back to FEXLINK: How the Dual-Database Architecture Lands
To land the dual-database logic on products: the FEXCloud IoT cloud platform organizes by "time-series + relational" dual databases — the time-series database stores point history (high writes, compressed retention), the relational database stores device profiles and business transactions (strong consistency), device IDs link both sides, each managing its authority zone — small and medium projects start with dual databases on one machine and scale smoothly as they grow. The platform's dual-database path: ① unify device IDs → ② time-series stores data, relational stores profiles → ③ profiles are the authority, data is the axis → ④ spec by scale — four steps done, dual databases in parallel.
One summary: the energy platform's dual databases = division by data characteristics — the time-series database stores high-frequency points (write + compression), the relational database stores profiles and business (transaction + relation) — device IDs link them, each manages its authority zone, and forcing them together sacrifices performance on one side. Different data characteristics require separate storage — the time-series database manages the "data flow", the relational database manages the "business state", each doing its strength.
One reminder: dual databases are not two systems each minding its own — device ID is the bridge, and profiles and data divide authority — device retirement goes profile first, data second; a new device is profiled first, then data collected — with the right order, the two sides never contradict. The key to dual-database parallelism is "making clear who is authoritative" — profiles belong to the relational database, data to the time-series database, and device IDs tie both sides together.
Scope and Limitations
- The conclusions in this article are intended for LV distribution and electrical-safety sites. Criteria are compiled from public codes and FEXLINK product documentation, and product parameters follow the latest selection handbook and product knowledge base.
- Quantitative criteria (protection zone, discharge and residual voltage, earth resistance, thresholds and grading) hold under the preconditions stated in this article. Where the site voltage level, earthing arrangement, environmental category or load profile differs, the figures must be recalculated. Specific configurations involving FEXCloud IoT cloud platform should be verified against actual circuit parameters.
- Specific product models and parameters are subject to the latest selection handbook and product knowledge base; this article does not constitute selection or procurement advice.
- Standard numbers cited here refer to the currently effective editions; when a standard is revised, the related quantitative conclusions must be re-checked.
- Deployment and implementation for specific projects must be carried out by qualified engineers based on site survey and design drawings.
Frequently Asked Questions
Why Energy Platforms Pair a Time-Series Database with a
A leader asks why one database cannot hold devices, points and alarms together