Status and roadmap
UltraFabric is built in phases. Each phase ends with a gate that has to pass before the next one starts, so the project can change course on evidence.
What works today
| Part | State |
|---|---|
| Storage engine | Built: write-ahead log with group commit, sorted tables, compaction, block cache, garbage collection of old versions, snapshot-isolation transactions, idempotent writes, memory governor. |
Local server (ufd) | Built: serves one engine over a Unix socket with the v0 protocol and Prometheus metrics. Not reachable over the network. |
| Query plan layer | Built: types and values, canonical encoding, fingerprints, validator, text form, reference interpreter and the security verifier. |
| Fast executor and optimizer | Built: a vectorized executor and a rule-based optimizer behind the verifier, both checked against the reference interpreter on tens of thousands of generated queries. |
| Queries over stored data | Built: row and columnar collections in a catalog, and verified, optimized queries through the local server. |
| Machine-checked proofs | Built: eight query rewrite laws and the tenant-isolation guarantee for the core operators are proved in Lean 4. |
| Analytic query speed | Much faster, not yet fast enough: columnar collections bring TPC-H to 11–46 times single-threaded DuckDB, from 140–1 400 times, with every answer equal (indicative, measured on a shared development server). The executor is the next step. |
| SQL and the PostgreSQL protocol | In progress: psql and drivers connect with a token, including prepared statements; tables, inserts, updates, deletes, filtered and grouped queries and joins run, and key lookups read only the rows they need. Transactions, the catalog tables ORMs read, and parallel reads come next. See SQL and psql. |
| Replication, backups, hosted databases | Planned. |
Phases
- Foundations (done). Simulator, fault-injecting disk, model-checked log design, benchmark harness.
- Storage engine (done, pending a long soak test and runs on dedicated benchmark hardware).
- Query plans and executor (correctness done; speed open). Plan language, verifier, executor, catalog, optimizer, ten million differential checks, tenant-isolation testing and proofs are complete. Analytic speed is the open item: the 3× target against DuckDB is missed by far and has its own work plan.
- SQL front end (in progress). PostgreSQL wire protocol, so psql and ORMs connect unchanged; one real application's test suite must pass against UltraFabric.
- Distribution. Replication per shard, distributed transactions, failover.
- Migration and the first hosted tenant. Copy-and-follow migration from PostgreSQL, backups and point-in-time restore, then 30 days on a real non-critical workload.
- Search, vectors and graphs, then retrieval with verified grounding, an AI runtime, a cost-based optimizer with post-quantum transport security, and research on quantum-assisted planning.
When hosted databases open
Hosted databases need phases 4 to 6: SQL access, replication, and backups that have been restored for real. The current estimate for a first hosted, non-critical workload is six to twelve months, and it will be revised at each gate. Early access teams will be the first to get databases, and the updates page records each step.
What will not ship early
- No hosted database before backups and restore have been rehearsed.
- No performance claims until they are measured on dedicated hardware.
- No feature that bypasses the verifier, including internal jobs and caches.