The database market is organized around nouns. Graph database. Vector database. Document database. Time-series database. Spatial database. Each noun corresponds to a useful computational model—and usually to a new product boundary.
That packaging has an architectural consequence. The same real-world thing becomes a row in one system, a node in another, an embedding in a third and an event in a fourth. ETL keeps the copies approximately aligned. Security teams reproduce policy. Recovery teams reason about several clocks. A query that crosses those interpretations becomes a distributed systems problem before it becomes a question about the data.
8DB begins with a different hypothesis: data modality is a projection over shared typed atoms, not the identity of the underlying store.
A model should describe a question
A table is excellent when a workload needs typed attributes, set operations and familiar transactional access. A graph is excellent when relationships and traversal dominate. A vector index is excellent for similarity. None of those observations requires the corresponding model to own a separate copy of reality.
Under a projection architecture, a durable atom can participate in multiple structures. The projections specify how to interpret, index and query the shared information. They do not require the application to translate one authoritative database into several secondary ones.
This changes the role of a modality from storage ownership to computational contract.
Native means more than accepting a data type
A database can claim support for a modality because it accepts a vector column or serializes graph-shaped JSON. That is compatibility, not necessarily native behavior.
A useful definition of native modality support requires at least four properties:
- Representation fidelity: the important semantics of the modality survive storage.
- Queryable operators: the engine can execute the modality’s defining operations.
- Integrated lifecycle: indexing, policy, provenance, recovery and replication include the modality.
- Cross-projection identity: the same information can be addressed across modalities without an external reconciliation key.
This definition is deliberately demanding. It prevents a universal database claim from becoming a catalog of parsers.
New structures should not require a new database
The projection model is most useful when a system encounters a structure that existing product categories do not express well.
8DB includes familiar table, document, graph, hypergraph, key-value, vector, time-series, spatial and media projections. It also works with structures such as order lattices, geometric manifolds, MERA tensor networks and topological braids.
The point is not that every application needs a braid or tensor network. The point is that adding a new way to understand existing information should not automatically create another storage, replication and security system.
An order lattice can preserve partial order, dominance and incomparability that a total ordering would destroy. A braid can represent equivalence across reorderings of independent events. A manifold can encode local geometry and test whether observations form one coherent global structure. These are computational choices over data—not reasons to duplicate the data estate.
Cross-modal query becomes a local question
Consider a query that asks which vessels:
- are connected to a target owner through a two-hop relationship;
- entered a spatial region during a defined interval;
- produced a time-series anomaly; and
- satisfy a confidence threshold.
In a polyglot design, this can require graph traversal, spatial filtering, time-series analysis and confidence logic across independent stores. The integration work includes identifiers, serialization, network calls, stale copies and policy alignment.
In the 8DB Nightmare workload, those projections resolve against shared state. The repeated cross-modal intersection measured 20.3 microseconds as a ten-run Criterion mean on Apple M1 hardware. That figure is scoped to the recorded workload and maintained intersection; it is not a claim that every arbitrary cross-modal query will complete in microseconds.
It does demonstrate the qualitative change: the engine is selecting and combining local projections, not federating independently authoritative products.
Removing ETL changes security as well as latency
ETL is often treated as a data-engineering cost. It is also a security boundary.
Every new copy can create:
- another credential and service identity;
- a new serialization and parser surface;
- logs or queues containing sensitive fields;
- a second retention and deletion lifecycle;
- backups that do not inherit the source policy;
- a timing window in which two systems disagree;
- a recovery order that can restore stale derived state.
A well-engineered polyglot system can manage those risks. The claim is not that multiple databases are inherently insecure. The claim is that each additional authoritative copy and integration path must be secured, observed, tested and authorized.
When projections share a substrate, policy and provenance can follow the information rather than being reconstructed for every engine.
The specialist-performance objection is correct
The strongest objection to model independence is performance. Specialist databases exist because physical layout, indexes and execution strategies matter. A universal logical model that ignores those differences will underperform.
Projection does not mean one physical layout. It means one durable identity and lifecycle with modality-specific physical structures. An adaptive query router can select the relevant projection, memory tier, index and execution path while preserving cross-projection identity.
That is why evaluation must compare both sides:
- specialist workloads against specialist systems;
- cross-modal workloads against the complete assembled stack;
- end-to-end application behavior with encryption, governance and recovery enabled;
- storage, memory, energy and operational cost, not latency alone.
8DB’s recorded results include 27.6 million operations per second in YCSB Workload B, 296,600 queries per second in LDBC SNB IS-1, 42.2 million metrics per second in a TSBS pipeline and 4,707 queries per second at 0.957 Recall@10 in the SIFT-1M vector comparison. Each result has its own environment and scope. Together they test whether consolidation requires a universal performance tax.
A different unit of database competition
If modality is a projection, the competitive unit changes. The buyer is no longer comparing 8DB with one graph database or one vector database in isolation. The relevant alternative is the assembled system required to answer the same questions with the same security, recovery and deployment boundary.
Sometimes the specialist remains the correct choice. If a workload is dominated by one modality, has modest security requirements and benefits from a mature ecosystem, a focused engine may minimize risk.
The projection architecture becomes more valuable as the number of modalities, copies, policies and disconnected environments grows. Its advantage is not that every database category disappears. It is that new understanding no longer requires new ownership of the data.
Continue the technical conversation
Bring the workload, boundary and objection.
8Braid can map the article’s claims to an acceptance plan for your security, data and platform teams.
Discuss this work