Combine data models

Goal. Model something that needs more than one shape, without running more than one database. About five minutes.

They are already in the same database

There is nothing to enable. A database has all five models from the moment it exists:

strata ./app kv put session:42 "active"
strata ./app json set user:ada '$' '{"name":"Ada","role":"engineer"}'
strata ./app vector collection create docs 3
strata ./app graph create org

strata ./app describe

StrataDB 1.2.2 · durable_local
branch default · branches: default · spaces: default
capabilities: arrow event graph_core inference json kv vector vector_index
kv 1 · json 1 · events 0
vector collections: docs (dim 3, cosine, 0 vectors)
graphs: org (0 nodes, 0 edges)

Why that matters more than it sounds

The usual version of this application has Postgres for the documents, Redis for the sessions, a vector database for the embeddings and a graph store for the relationships. Four systems, four connection strings, four backup stories, and no moment at which they are all consistent with each other.

Here they share three things:

One commit timeline. A write to any model advances the same version counter, so --as-of 7 means the same instant to all of them:

strata ./app json get user:ada '$' --as-of 7
strata ./app graph bfs org ada --as-of 7

That is a consistent snapshot across shapes, which is the thing four systems cannot give you.

One set of branches. Forking forks everything:

strata ./app branch fork default experiment

The documents, the sessions, the vectors and the graph all have an experiment version now. There is no coordination to write.

One directory. One thing to copy, one thing to back up.

A worked shape

An agent application might use all four:

• JSON for the entities it reasons about, because it updates one field at a time.
• Key-value for opaque blobs and session state, where the value means nothing to the database.
• Vectors for the text it needs to search by meaning.
• Graph for how the entities relate, when the question is “what depends on this”.
• Events for the log of what it did, which Record tool activity covers.

Pick per question, not per entity. The same user can be a JSON document, a vector and a graph node at once.

The one thing that does not compose

Merging does not carry events or graph data. Key-value, JSON and vectors move; the other two do not, and the merge receipt says success either way.

strata ./app branch preview experiment default

The capabilities_unsupported line is the authoritative list. If your design puts graph writes on a branch that has to merge, change the design rather than discovering this later. Branches covers why.

Failure modes

A cross-model --as-of read returns nothing for one model. That model had no data at that version yet. Versions are shared, so a model added later starts later.

Graph or event data missing after a merge. See above.

You want a join. There is not one. Reads are per model, and combining them is your code’s job.

Next

Working with data for choosing between the models in the first place.