Understand Cognite Data Fusion

1.1 Three islands of data in every plant

Walk into any power plant, refinery or factory and you will find that the data about a single piece of equipment — say Boiler Feed Pump BFP-1A — is scattered across at least three separate worlds:

Each world has its own software, its own naming (the pump is BFP-1A in the P&ID, 21-PA-001A in SAP, and U2.BFP1A.DISCH.PRESS in the historian), and its own owner. Nobody can look at one screen and see the whole truth about BFP-1A.

1.2 What "contextualization" means

You will hear this word constantly in the Cognite world. It sounds complicated; it is not.

Once the lines are drawn, you can ask questions a plain historian can never answer: "Show me every pump whose vibration went up in the 30 days after a bearing replacement." That needs OT (vibration), IT (work orders) and ET (which pumps have that bearing) joined together.

1.3 Who is Cognite?

• Founded in Oslo, Norway by a team from the industrial group Aker ASA (incorporated October 2016; Cognite's own materials date the company to 2017, when it began operating). Co-founder John Markus Lervik was CEO until 2022; Girish Rishi has been CEO since. The oil company Aker BP was the anchor customer — and Cognite's largest customer and a shareholder until 2022 — and still runs its offshore operations on CDF.
• Scale (2025): revenue above $170 million, ARR bookings up 36%, 800+ employees; earlier investors include TCV (a $150 million round in 2021 valuing Cognite at $1.6 billion) and Saudi Aramco.
• June 30, 2026 — Schneider Electric agreed to acquire Cognite for $3.1 billion in cash. Cognite is to be integrated with AVEVA (Schneider's industrial-software arm), with Data Fusion and Atlas AI becoming the data-and-AI layer inside AVEVA's CONNECT platform. Closing is expected "in the coming quarters", subject to regulatory approvals. Module 6.4 discusses what this means for the skill.
• Product: Cognite Data Fusion® (CDF) — a software-as-a-service (SaaS) platform. Nothing to install on premises except small "extractor" programs that push data up.
• Newer layer: Cognite Atlas AI™ — a workbench for building AI agents that answer questions using the knowledge graph.
• Industries: oil & gas (Aker BP, BP, Saudi Aramco; SLB as a partner), power generation & grid (e.g. Statnett), chemicals (Celanese), manufacturing (Alfa Laval, Mitsubishi Heavy Industries), renewables. Sold directly and through cloud and consulting partners (Microsoft, AWS, Google Cloud, SLB, Accenture and others).
• The category they created is called Industrial DataOps — applying the discipline of DevOps to industrial data: automated, monitored, governed pipelines instead of one-off spreadsheets.

2.1 The definition

2.2 The official architecture diagram

This is Cognite's own architecture picture from the documentation. Read it left to right: data enters on the left, gets integrated, becomes the knowledge graph in the middle, and is consumed on the right. The bottom band is the foundation it all sits on.

2.3 The 8 layers, one by one

Here is the same diagram flattened into a stack so you can read each layer. The chips are the exact capability names Cognite uses.

Layer 1 — Data sources (what goes in)

Everything from Module 1: historians and control systems (OT), business systems (IT), drawings and models (ET), plus two newer sources — field workers (checklists, photos, observations captured on tablets) and robots/drones (inspection images, thermal scans).

Layer 2 — Data integration (how it gets in and gets linked)

• Extractors — small programs (Cognite-provided or your own) that read from a source system and push data to CDF. The Extractors page lists 102 entries — 23 Cognite-built extractors plus 79 documented source integrations (Cognite's marketing says "90+ connectors"): OPC UA, OSIsoft/AVEVA PI, SAP, databases, MQTT, Kafka, Azure Event Hub, file shares…
• Connectors — two-way links to cloud platforms (e.g. Microsoft Fabric) so data can flow both directions.
• Transformations — SQL-based (or Python) steps that reshape raw rows into the data model: rename columns, join tables, compute fields.
• Data workflows — the orchestrator: "run extractor → then transformation → then function, every hour; alert me on failure."
• Functions — your own Python code, hosted and scheduled inside CDF (like AWS Lambda, but with the CDF SDK built in).
• Entity matching, Diagram parsing, 3D contextualization — the AI tools that draw the relationship lines (Module 1.2). Entity matching links tag names to assets; diagram parsing reads P&IDs and finds every tag on the drawing; 3D contextualization maps CAD nodes to assets.

Layer 3 — Industrial knowledge graph (where it lives)

The heart of CDF. Not just a database — a graph: things (nodes) and relationships (edges), each with properties. It natively stores time series (billions of data points), files (PDF, images), 3D models and point clouds, geospatial shapes, annotations (e.g. "this box on page 3 is tag 21-PT-1019"), and units of measurement. Data modeling is how you define the shape of your graph (Module 4).

Layer 4 — Data management (how you govern it)

Explore and search everything, catalogue it (which data set, from which source, with what quality), and control who can see and change what via access management tied to your company's identity provider.

Layer 5 — Atlas AI (the agents)

Low-code agent builder to create AI assistants that answer questions using the knowledge graph ("why did the compressor trip last night?"), a library of ready-made agents (root-cause analysis, troubleshooting), and LLM benchmarking to pick the best language model for each job.

Layer 6 — Industrial tools (the apps people use)

• Search — Google-like search across assets, time series, documents, 3D.
• Canvas — a free-form whiteboard where you drop P&IDs, charts, 3D, photos and work orders side by side, comment, and ask the AI copilot questions (see Module 5).
• Charts — trend and calculate on time series without code; set alerts.
• InField (field operations) — tablet app for operator rounds, checklists, observations.
• Maintain (maintenance planning) — build and optimise turnaround/maintenance scopes.
• InRobot (robotics) — plan and run robot inspection missions.

Layer 7 — Consume & integrate (how developers and BI reach it)

Everything in CDF is available through a REST API, a GraphQL endpoint for the data model, official Python and JavaScript SDKs, and ready connectors for Power BI (OData) and Grafana. If you can call an API, you can build on CDF.

Layer 8 — Foundation

CDF runs as SaaS on AWS, Azure, or Google Cloud (you pick the region/cloud), authenticates with OpenID Connect against your own identity provider (Microsoft Entra ID etc.), and is certified to ISO 27001. More in Module 7.

2.4 Vocabulary you will meet everywhere

A fuller glossary is at the bottom of this page.

3.1 The six stations

3.2 What each station produces

3.3 Worked example — a boiler feed pump in a coal-fired unit

Let us follow one tag and one pump through the whole journey, using names you would recognise from a thermal power plant.

Notice what happened at the Link step: three different names (U2.BFP1A.DISCH.PRESS, BFP-1A, and the SAP functional location 21-PA-001A) were resolved to the same node. From then on, anyone who opens BFP-1A sees the trend, the drawing, and the work-order history together — no more hunting.

3.4 Batch vs. streaming

• Streaming (seconds): time-series datapoints from OPC UA / PI / MQTT go straight to the time-series store.
• Scheduled batch (minutes to daily): SAP work orders, equipment master data, documents — extract → RAW → transform on a schedule.
• One-off: 3D models, historical drawing archives — uploaded once, revised occasionally.

Data workflows tie these together and give you a single screen showing whether every pipeline ran on time.

3.5 Watch the flow

4.1 What a knowledge graph is

A relational database stores tables. A knowledge graph stores things and their relationships. Each thing is a node with properties; each relationship is an edge. The power comes from being able to walk the edges: from a pump to its sensors, from a sensor to the drawing that shows it, from the drawing to the other equipment on the same page.

4.2 The Core Data Model (CDM)

You could design your own graph from scratch, but Cognite ships a standard, ready-made schema called the Core Data Model. Using it means every Cognite tool (Search, Canvas, Charts, InField, Atlas AI) understands your data out of the box. The main types (all prefixed Cognite):

4.3 The four building blocks of a data model (overview only)

You will go hands-on with these in Level 2. For now, just know the names and the picture:

Then you ingest instances (nodes and edges) into the model with transformations or the SDK, and query them with GraphQL, the REST API, or the Python SDK. Cognite also offers industry-specific extensions of the Core Data Model, e.g. the Process Industries Data Model.

5.1 The layout

Log in at your company's CDF URL (it looks like https://<org>.fusion.cognite.com). You land on a Home page with a left navigation rail. Exact labels move around between releases, but they always group into these areas — and they map straight onto the layers from Module 2:

5.2 Integrate — the Extractors library

Names you will recognise: Cognite PI AF Extractor (AVEVA/OSIsoft PI), OPC Classic and OPC UA, DB Extractor (any ODBC database), Kafka, Azure Event Hub, Microsoft Fabric Connector, Rockwell FactoryTalk, plus oil & gas specific ones (OSDU, Petrel, EDM).

5.3 Industrial Canvas — the collaborative whiteboard

Canvas is where the knowledge graph becomes visible. Because everything is linked, clicking the pump on the P&ID offers Show in 3D, Find related data, Create isolation. This is the tool most domain experts start with — it needs no training beyond "drag things onto the board".

5.4 The other Industrial Tools in one line each

5.5 Where you can practise — four options

Option B in detail — get into the free sandbox

1. Create a free account on Cognite Hub: hub.cognite.com (email or Google login). This one account gives you the Hub community, Cognite Academy, and the sandbox.
2. Open the Open Industrial Data group and read the pinned "What is Open Industrial Data?" post (5 min) — it describes the Valhall first-stage compressor system you will be looking at.
3. Go to publicdata.fusion.cognite.com and sign in with the same account. You will land in the CDF UI of the publicdata project.
4. Two projects share the data: publicdata (classic asset hierarchy) and publicdatacdm (the same data loaded into the Core Data Model from Module 4 — use this one to see CogniteAsset / CogniteEquipment / CogniteTimeSeries for real).
5. The connection details for code (Option D) are on the group's "OpenID Connect on Open Industrial Data" post — copied below. Licence: free for learning and non-commercial use within your organisation; commercial use and redistribution are not permitted, and academic publication was still an open question on the Hub as of late 2025 (terms thread).

Option C — the click-through demos (no login)

Option D — hello world from your laptop (15 min)

Python 3.10+ is enough. In a terminal:

pip install cognite-sdk msal pandas

Then in a Python file or Jupyter notebook — this opens a browser window for you to sign in with your Cognite Hub account, and lists the first assets and time series of the sandbox:

from cognite.client import CogniteClient

client = CogniteClient.default_oauth_interactive(
    project="publicdata",                                   # or "publicdatacdm"
    cdf_cluster="api",                                      # -> https://api.cognitedata.com
    tenant_id="48d5043c-cf70-4c49-881c-c638f5796997",       # Cognite Hub tenant (from the OIDC post)
    client_id="1b90ede3-271e-401b-81a0-a4d52bea3273",       # public client for Python / Postman
    client_name="hnm-learning",
)

print(client.assets.list(limit=5).to_pandas())
print(client.time_series.list(limit=5).to_pandas())

For your own company project, replace project, cdf_cluster, tenant_id and client_id with the values your CDF admin gives you (they are shown under Admin → Access management → Identity provider). You will meet this code again, properly, in Level 2 Topic 5.

5.6 Watch the tools in action

6.1 The business case in numbers

Forrester Consulting has produced Total Economic Impact studies commissioned by Cognite. The 2022 study reported 400% ROI and $21.6M net present value for a composite organisation (30 facilities, 25,000 employees, $2B revenue). Cognite's current product and customer pages cite an updated study with the figures below — all are modelled composite results over three years, not one customer's audited numbers:

Where do these come from? Four mechanisms:

1. Time to find data → time to fix problems. Engineers typically spend the majority of their analysis time locating and cleaning data. When the pump's trend, drawing and history are one click apart, a root-cause analysis that took days takes hours.
2. Condition-based instead of calendar-based maintenance. Linking sensors to work orders lets you see which equipment actually degrades — and stop over-maintaining the rest.
3. Production optimisation. Cross-system calculations (heat rate, specific energy, efficiency curves) become live time series instead of monthly spreadsheets.
4. Reuse. A data foundation built once serves hundreds of use cases; the cost of the 50th dashboard is near zero.

6.2 What real customers report

6.3 CDF versus what you already have

6.4 Should I invest in this skill? An honest view

7.1 Under the hood

• SaaS, multi-cloud. Cognite operates CDF on Microsoft Azure, Google Cloud and AWS. Your company's project is created in a regional cluster (e.g. westeurope-1, az-eastus-1) — you choose the cloud and the region for data residency.
• Micro-services on Kubernetes. Per a Cognite Hub answer by Jason Dressel of Cognite (July 2023), CDF is "300+ micro-services which are continuously deployed as many as 1000 times per week across development, staging and production environments". You never see this — you just get new features.
• Polyglot storage. Each data type uses the right engine: time series use a combination of FoundationDB, Kafka, PostgreSQL and Elasticsearch; files use cloud object storage; the graph has its own store. Again invisible to you, but it explains why CDF can hold billions of datapoints and still answer searches in milliseconds.
• You own the data. Cognite's terms state the customer retains ownership; Cognite processes it to provide the service.

7.2 What runs where

7.3 Security: who may do what

• No separate CDF passwords. CDF trusts your company's identity provider (OpenID Connect). Turning off someone's corporate account turns off their CDF access.
• Capabilities are fine-grained (per resource type × per action) and scoped (to a data set, a space, or specific IDs). "Read time series in the Unit 2 data set" is a normal rule.
• Extractors and scripts use client credentials (service principals), not personal logins.
• Certifications and continuity details are published on Cognite's trust portal (ISO 27001 appears on the architecture diagram).

7.4 Watch: CDF on the clouds

8.1 Ten-question self-check

8.2 Earn the official certificate

Now take the free official path — it repeats what you learned here with Cognite's own slides and a hands-on part, and ends with a badge and a certificate for LinkedIn. (The full recommended order for the whole Academy portal is on the course home page.)

8.3 Recommended reading & watching before Level 2

• docs.cognite.com/cdf — read the "About CDF" page and skim the left menu so you know where things are.
• "Cognite Data Fusion Explained: Solving the Industrial Data Problem" (Medium) — a good independent summary.
• The Definitive Guide to Industrial DataOps (free e-book, registration).
• Full video library, grouped by level: Resources → Videos.

A preview of where this is going

Glossary