Every year, the data engineering community circulates a list of "architecture patterns you must master." The 2026 list is not much different — the problem is that most of these patterns are explained with web and commerce data in mind. Data on the factory floor is different in character: tens of thousands of time-series events per second, millisecond latency requirements, air-gapped networks, and context-free signals like "TAG_00042." Let's reread the 12 patterns through the eyes of industrial data.
Batch and Stream — Patterns of Flow
1. Lambda architecture — Runs a batch path and a real-time path in parallel and merges the results. As proven as it is heavy. Maintaining two sets of pipelines is a burden for manufacturing IT organizations.
2. Kappa architecture — The simplification that "everything is a stream." Equipment data is a stream by nature, making this the pattern with the best fit for manufacturing. Reprocessing is handled by stream replay as well.
3. Event-driven architecture (EDA) — Publishes state changes as events and reacts through subscriptions. A natural match for manufacturing logic that "reacts to what happened" — alarms, interlocks, quality dispositions.
4. Streaming-first + CEP — Detects patterns on the stream with complex event processing (CEP). Threshold breaches, trend deviations, and compound-condition alarms must be caught in the flow, not stored and queried afterward.
Storage and Structure — Patterns of the Vessel
5. Data Lake — A reservoir that accumulates raw data as-is. Flexible, but left unattended it becomes a "data swamp." In manufacturing, it needs to share the work with a time-series-optimized store.
6. Lakehouse — A compromise that layers the warehouse's management (schemas, transactions) on top of the lake's flexibility. Valid as the long-term retention and enterprise analytics layer for OT history.
7. Medallion (Bronze/Silver/Gold) — Stratifies quality through raw → refined → business stages. Transposed to manufacturing, the three tiers map exactly to "raw signals → standardized tags → KPIs (OEE, RAM, EMS)."
8. Data Vault — A modeling technique that never loses change history. Highly relevant to defense and aerospace quality data, where audit and traceability are everything.
Organization and Meaning — Patterns of People
9. Data Mesh — Domains, not a central team, own data as a "product." For a multi-plant enterprise, this reads as the combination of per-plant data products plus enterprise-wide standard contracts.
10. Data Fabric — An integration layer that weaves scattered data together with metadata and semantics. In manufacturing, the technology playing this role is the Unified Namespace (UNS). The asset hierarchy becomes the real-time addressing scheme.
11. Semantic layer / ontology — The layer that gives data meaning. If you don't know which line, which piece of equipment, and what signal "TAG_00042" refers to, neither AI nor humans can answer. This is the pattern that decides the success or failure of manufacturing AI.
12. Data products & data contracts — Manages data like an API, with versions, quality, and SLAs. This is the direction in which MES and ERP integration evolves from "throwing files over the wall" to "contract-based interfaces."
The Conclusion for the Factory Floor
This does not mean adopting all twelve. The combination that decides the game in industrial data is clear:
- Kappa + CEP — equipment data is born as a stream
- UNS (Data Fabric) — the asset hierarchy as a real-time addressing scheme
- Ontology (semantic layer) — a data foundation that AI can understand
- Medallion-style stratification — quality stages from raw signals to KPIs
The PlantPulse® Platform implements this combination as a single package — CEP at 510,000 events per second, a UNS built on MQTT and Sparkplug B, and an ontology knowledge graph on top of the ISA-95 asset hierarchy.