I reach for graph data when the relationship is the question. If the business asks who is connected to whom, how risk spreads, or which path links two entities, I recommend modeling edges as first class data.
The path is the data
Graph data is the right shape when the relationship is more important than the row.
How relationships reveal hidden behavior
Graph data is the right shape when the relationship is more important than the row.
Where connected data lives
The test is whether the data is trustworthy enough for a business action, not just stored correctly.
Neo4j, Amazon Neptune, Cosmos DB graph APIs, JanusGraph, TigerGraph, RDF stores, graph tables in lakehouses, Elasticsearch for graph backed search projections.
Managed graph options
| Provider | Typical services |
|---|---|
| AWS | Neptune, Glue, S3, OpenSearch, SageMaker. |
| Azure | Cosmos DB graph APIs, Azure Databricks, AI Search, Synapse. |
| Google Cloud | Cloud Spanner Graph, BigQuery, Dataflow, Vertex AI, partner graph databases. |
Relationship pattern to remember
For graph data, I keep the source relationship and timestamp beside every edge so the graph can explain why two things are connected.
Pinterest is a public example: its Taste Graph connects people, pins, topics, and interests so recommendation and advertising systems can reason about relationships rather than isolated rows.
Fraud ring example
The point here is to move from concept to operation: who uses it, what breaks, and what decision changes.
A fraud platform links accounts, cards, devices, addresses, and login IPs. A graph query finds new accounts that share a device with previously blocked accounts within two hops.
Graph review query
The point is to show who uses the data and what decision it supports.
MATCH (a:Account {id: $accountId})-[:USED_DEVICE]->(d:Device)<-[:USED_DEVICE]-(other:Account)
WHERE other.status = 'BLOCKED'
WITH a, count(distinct other) as blocked_neighbors, collect(d.id) as shared_devices
RETURN a.id,
blocked_neighbors,
shared_devices,
CASE WHEN blocked_neighbors >= 2 THEN 'REVIEW' ELSE 'ALLOW' END as decision;The query follows shared device relationships, counts blocked neighbors, and returns a review decision. The graph is useful because the risky pattern is in the connection path.
Graph traps to watch
- Define edge direction and meaning carefully.
- Prevent graph traversals from becoming unbounded.
- Keep graph features explainable for reviewers.