Short answers, sourced.

Not as a service. The storage engine, the plan layer with its verifier, columnar collections and a first SQL subset over the PostgreSQL protocol are built and tested on a single node. There is no public build or hosted endpoint yet. Hosted databases open to early-access teams after replication, backups and point-in-time restore are in place.

After the SQL front end, distribution, and backups that have been restored for real. The current estimate for a first hosted, non-critical workload is six to twelve months from October 2026, revised at every phase gate. The roadmap is written as gates, not dates.

A hosted database that speaks the PostgreSQL protocol, for a non-critical workload; backups and point-in-time restore from the first day; the benchmark and end-to-end test scripts; and a direct line to the engineers. It is free during development.

Self-hosting on your own hardware is the first deployment model and runs with no runtime call to any cloud: a single static binary per node, the gateway and tooling. It will be offered alongside the managed service, and its licence terms are published with the first release.

Yes, after replication and backups: clusters that are operated for you, with replication, backups, point-in-time restore and upgrades. Usage is priced on nodes and storage, not on queries. Tiers are announced with the service, not before.

Linux. Measurements so far are from a single shared development host; the dedicated benchmark host planned for published targets is 8 vCPU, 32 GB and local NVMe. Multi-node deployments are specified but not built.

It speaks the PostgreSQL wire protocol through a gateway, ufgw: psql and the Go driver pgx connect unchanged with a token as the password, over the simple protocol and over prepared statements with $1 parameters in text and binary formats. SQL is parsed with PostgreSQL's own parser. The SQL subset is deliberately small today.

CREATE TABLE with a primary key (row or columnar storage); INSERT, UPDATE and DELETE, each applied atomically; SELECT with WHERE, inner and left joins, GROUP BY with count, sum, avg, min and max, ORDER BY, LIMIT and OFFSET; comparisons, AND/OR/NOT, IN, BETWEEN, IS NULL, arithmetic and casts; the types smallint, int, bigint, real, double precision, numeric(p,s), text, boolean, uuid, date, timestamp and bytea. Everything else fails at once with SQLSTATE 0A000 and names the missing feature; nothing is silently approximated.

Not yet. Drivers connect, because the simple and extended protocols are built, but ORMs and database GUIs read the pg_catalog and information_schema tables when they connect, and they need transactions. Both are the next gateway items, together with RETURNING and ON CONFLICT; the gate for that step is one real application's test suite passing. Python, Node.js, Java and Rust drivers speak the protocols that are built but are not in the test suite yet.

Every query reads only your tenant's rows; there is no tenant column to select, filter on or write. Primary keys are unique per tenant. The schema belongs to the application and is managed with a separate operator credential that can never read or write tenant data, and an ordinary tenant token can never manage the schema.

Licence and service terms are published with the first release of each deployment model, not before. Nothing on this site grants a licence to the software. The repository is private during development; specifications, documentation and benchmark reports are published here, and early-access teams receive the build and the test scripts.

Early access is free, in exchange for feedback on real workloads. Self-hosting will be priced per node, with an optional support subscription; the managed service per node and storage. Exact tiers are announced with the service.

Every query becomes a plan, and the data node runs a verifier on it before executing it. A plan the verifier accepts returns only data the requester may see, and its result does not change when other tenants' data changes. The guarantee holds whoever produced the plan: optimizer, plan cache, internal job or a hand-written client. The core theorem is proved in Lean 4 and the behaviour is tested on generated databases that differ only in hidden data.

No. UltraFabric holds no SOC 2, ISO 27001 or similar certification yet, and the security page will say so when that changes, not before. What exists is a written security model, a declared-leakage register for side channels, and a responsible-disclosure process at [email protected].

Row-level policies (today tenant isolation is the access boundary); the vector, text and graph index probes, which are specified but not built; timing and resource-use side channels, which are entered in the declared-leakage register with mitigations; and multi-node distribution, which is single node today.

From dated benchmark reports in the repository, each naming its commit, host, tool and raw result file. All runs so far are on a shared development host with other services running, one node and data that fits in memory, so every figure is labelled indicative. No target is published until it is measured on dedicated hardware against a named baseline at a stated version.

In parts. On the key-value runs of 19 October 2026, point reads reached 412 447 operations per second against RocksDB's 274 508 in the same run, with memory at or below RocksDB's. On TPC-H at scale factor 0.1, columnar collections are 11 to 46 times slower than single-threaded DuckDB with every answer equal; the project's own 3× gate is not met and the executor is the next work. A primary-key lookup through the gateway takes 0.14 ms on the development host.

You do. Early-access databases are provided free, as is, for non-critical workloads. The team does not look at your data except to diagnose a problem you report, and deletes it when access ends. A database may be withdrawn or reset with notice.

In order: transactions, the catalog tables ORMs read and parallel reads (the rest of the SQL front end); distribution with replication per shard, distributed transactions and failover, tested Jepsen-style; migration from PostgreSQL, backups and the first hosted tenant; then search, vectors and graphs behind the same verifier, each proven by a public benchmark against a named leader.

No. Internal jobs, caches and any future learned optimizer are plan authors like any client; there is no privileged path to a row. That rule, no hosted database before restore is rehearsed, and no performance claim before dedicated hardware, are the three things the roadmap says will not ship early.