Example library

Build a graph of 10,000 people and 49,985 follows, then use SQL to explore communities, propagate state, trace exposure, and try changes transactionally. Every recipe includes runnable SQL and explains how to interpret its results.

Start with the dataset setup. You can then run any recipe in any order. Each recipe also has a complete SQL download that includes setup and runs in a fresh database, through either psql or pgAdmin’s Query Tool. The examples use functions available in pg_warren 0.7.0.

Example	What you will build
Traverse from your application data	Look up a person, explore their neighbourhood, and join results back to your tables.
Find communities with Louvain and Leiden	Label every person with a community and inspect the largest groups.
Count walks with a semiring	Count all six-hop walks from a person using `plus_times`.
Compute distances with min plus	Implement shortest distances with `min_plus` and check them against BFS.
Propagate a boolean through the graph	Spread an alert and save a reached flag for every person.
Trace exposure to the nearest taint	Start from 100 tainted people, save nearest distances, and reconstruct an explanatory path.
Try a change and roll it back	Build a staged version, query the proposed change, and undo it.

Use your own data

The graph uses your original table ids. A node_id returned by BFS or a vector is a people.id, not an internal graph index. Replace the node and edge queries in the setup with your application’s queries, then replace the sample name or taint selection with a business identifier or rule.

The graph is directed: src follows dst. Traversal defaults to following those edges outward. Louvain and Leiden instead use the undirected view. The dataset is synthetic, with no planted communities; discovered groups demonstrate the algorithms rather than established real-world structure.

Results and freshness

Queries read a built graph version. After changing people or follows, call warren.refresh('example_social') and rerun analyses for that new version. Vectors and saved route tables are computed results, not automatically maintained annotations. Changing a seed selection also requires rerunning the corresponding analysis.

The examples show correctness and composition with SQL. They are not performance benchmarks; timings depend on hardware and workload. For measured comparisons, see performance.