<h1 align="center">🛰️ Universal Data Broker (UDB)</h1>
<p align="center"><em>One proto-driven gRPC control point in front of 18 databases — with built-in auth, multi-tenant RLS, migrations, CDC, and SDKs for 6 languages.</em></p>
<p align="center">
<a href="Cargo.toml"><img alt="Rust 2024" src="https://img.shields.io/badge/Rust-2024-b7410e?logo=rust&logoColor=white"></a>
<a href="proto/README.md"><img alt="gRPC + Protobuf" src="https://img.shields.io/badge/API-gRPC%20%2B%20Protobuf-244c5a?logo=grpc&logoColor=white"></a>
<a href="sdk/UDB_PROTOCOL_VERSION"><img alt="Protocol 1.0.0" src="https://img.shields.io/badge/protocol-1.0.0-2f855a"></a>
<a href="#-backend-matrix"><img alt="Backends" src="https://img.shields.io/badge/backends-18-4c51bf"></a>
<a href="#-supported-features"><img alt="DataBroker RPCs" src="https://img.shields.io/badge/DataBroker-73%20RPCs-0e7490"></a>
<a href="#native-control-plane"><img alt="Control plane" src="https://img.shields.io/badge/control%20plane-6%20services%20%C2%B7%2077%20RPCs-7c3aed"></a>
<a href="#-quickstart-per-language"><img alt="SDKs" src="https://img.shields.io/badge/SDKs-Go%20%C2%B7%20Python%20%C2%B7%20TS%20%C2%B7%20Java%20%C2%B7%20C%23%20%C2%B7%20PHP-1f6feb"></a>
<a href="LICENSE"><img alt="License MIT" src="https://img.shields.io/badge/license-MIT-555"></a>
</p>
UDB is a Rust implementation of a proto-driven data broker. It reads project-owned
`.proto` schemas, extracts storage annotations, builds a catalog manifest, generates
migration/bootstrap artifacts, and serves those schemas through a neutral gRPC
`DataBroker` API — fronted by a native auth/authz control plane.
```mermaid
flowchart LR
P["📦 project .proto<br/>(+ udb annotations)"] --> PA["🧩 parser →<br/>ProtoSchema AST"]
PA --> CM["📚 CatalogManifest<br/>+ checksum"]
CM --> GEN["🛠️ lint · drift ·<br/>migrations · SQL"]
CM --> RT["⚙️ DataBroker runtime"]
subgraph RT_PIPE["request pipeline"]
direction LR
AUTH["🔐 authn / authz"] --> ADM["🚦 channel admission"] --> IR["🔁 neutral IR"] --> EX["🔌 backend executor"]
end
RT --> RT_PIPE
EX --> DB[("🗄️ 18 backends<br/>SQL · vector · object ·<br/>cache · doc · graph · column")]
classDef accent fill:#1f6feb,stroke:#0b3d91,color:#fff;
class RT,RT_PIPE accent;
```
This repo is not only a parser and not only a gRPC server. It is a crate,
binary, runtime, protocol module, SDK workspace, backend plugin inventory,
operation IR, migration engine, and a set of operational runbooks. It also has
visible architectural history: early UDB was Postgres anchored, and the current
codebase is in a peer-to-peer transition where canonical stores are explicit
traits instead of implicit `PgPool` access.
## What This Project Is
UDB tries to solve a specific problem: many services want to read and write
business data, vectors, blobs, cache entries, CDC events, and admin/catalog
state, but every service talking directly to every database creates drift in
authorization, migrations, tenant isolation, observability, and retry behavior.
UDB centralizes those concerns:
- Project schemas stay in normal project-owned proto packages.
- UDB annotations describe relational tables, object fields, vector stores,
caches, document stores, graph stores, time-series/column stores, and security.
- The broker exposes one UDB-owned gRPC contract under `proto/udb/...`.
- Runtime requests carry tenant, purpose, scopes, service identity, project id,
and catalog version metadata.
- Backends are reached through a neutral logical IR and feature-gated plugin
modules instead of service code hand-writing each database dialect.
## ⚡ Supported Features
Data plane (`DataBroker`, 73 RPCs):
- **Relational** CRUD + batch (`Select`/`BatchSelect`/`Upsert`/`BatchUpsert`/`Delete`).
- **Vector** search / hybrid search / upsert (Qdrant, Weaviate, Pinecone, Elasticsearch knn).
- **Object/blob** put/get, presigned URLs, multipart (S3, MinIO, Azure Blob, GCS).
- **Cache** get/set/delete/scan (Redis, Memcached).
- **Document / graph / time-series / analytical** ops (MongoDB, Neo4j, ClickHouse, Cassandra).
- **Transactions**: per-request transactionality, real Postgres **2PC** and MySQL **XA** (`UDB_2PC_ENABLED`), **sagas** with recovery/compensation.
- **CDC → Kafka** via a transactional outbox relay, with DLQ, topic policy, and a CDC control plane.
- **Catalog & migrations**: staged/activate/rollback catalogs, proto-driven migration plan/apply with an audited op ledger.
- **Projections / materialized views**, **per-tenant RLS**, **field-level encryption** (AES-256-GCM-SIV), rate limiting / fair channels / backpressure, Prometheus metrics.
Control plane (`proto/udb/core/**`, isolated listener — see [Native Control Plane](#native-control-plane)):
- **Authn**: native JWT validation (JWKS/`kid`), **UDB-issued RS256 JWT signing + refresh tokens**, **Argon2id** passwords, **RFC 6238 TOTP MFA**, server-side sessions, **CSRF**, OTP, full user admin, mTLS + hybrid external identity (OIDC/Better Auth bridge).
- **Authz**: RBAC + ABAC + simple ReBAC over a Casbin enforcer, role/policy/relationship CRUD, audit decisions, **`GetNativeAccess`** (restricted role + scoped DSN + RLS session vars), **signed policy bundles** for offline SDK caches.
- **ApiKey**: hashed keys, scopes, rotation, revocation, usage stats.
- **Tenant / Notification / Analytics**: tenant + config management, notification logs/templates/preferences/delivery-stats (with Kafka emit), and pipeline/executor/reconciliation/throughput/SLA analytics.
All native control-plane CRUD is **proto-driven** (table + column shape resolved from the embedded `proto/udb/core/**` manifest via `NativeModel`) and **Postgres-backed, fail-closed** — no in-memory stores.
## Honest Status
The maintained docs now live under [`docs/README.md`](docs/README.md). The old
older notes and duplicate runbooks were consolidated so current status is easier
to verify. The source of truth for backend inventory is:
- [`src/backend/mod.rs`](src/backend/mod.rs): `BackendKind`, tier, role,
capability matrix, operation support.
- [`src/backend/plugins/mod.rs`](src/backend/plugins/mod.rs): compiled plugin
inventory.
- [`src/runtime/executors/`](src/runtime/executors): runtime executor modules.
- [`src/ir/compile/`](src/ir/compile): backend-specific IR compilers.
- [`Cargo.toml`](Cargo.toml): default and optional feature graph.
The current code recognizes 18 backend kinds and the default feature set enables
their plugin modules. A slim build can compile only the core pieces, for example
`--no-default-features --features postgres`.
Also note that the Docker files still show some historical path assumptions in
places. The Rust crate and CLI are the most reliable entry points while the repo
split/packaging work settles.
## Codebase Map
Approximate source shape at the time this README was written:
| [`src/runtime`](src/runtime) | 102 | Broker orchestration, service handlers, backend clients, CDC, catalog, system stores, security, metrics |
| [`src/ir`](src/ir) | 28 | Neutral logical operations and backend compilers |
| [`src/generation`](src/generation) | 17 | Manifest, SQL, DSN, drift, lint, and backend artifact generation |
| [`src/migration`](src/migration) | 7 | Diffing, plans, audited apply, phase runner, db_ops sync |
| [`src/control`](src/control) | 11 | Startup lifecycle, FSM, hooks, notifications, approval workflow |
| [`src/parser`](src/parser) | 10 | Hand-written proto lexer/parser and annotation extraction |
| [`src/backend`](src/backend) | 21 | Backend identity, capabilities, plugin contract, plugin inventory |
| [`src/cli`](src/cli) | 7 | `udb-proto-parser` command implementation |
| [`src/planning`](src/planning) | 4 | Request planning helpers for broker operations |
| [`src/schema`](src/schema) | 3 | Proto AST structs and deterministic checksums |
| [`crates/udb-portable`](crates/udb-portable) | 2 | WASM/edge-safe parser/checksum/schema-cache subset |
The public crate surface is collected in [`src/lib.rs`](src/lib.rs). The binary
entry point is tiny by design: [`src/main.rs`](src/main.rs) calls the CLI module.
## Request Flow
```mermaid
sequenceDiagram
autonumber
participant C as 📱 Client / SDK
participant S as ⚙️ DataBrokerService
participant A as 🔐 authz (v2)
participant Ch as 🚦 channels
participant IR as 🔁 IR + router
participant B as 🔌 Backend
C->>S: gRPC + metadata (tenant, purpose, scopes, …)
S->>S: SecurityContext · ensure_ready() · catalog compat
S->>A: authorize(identity, tenant, op, resource)
A-->>S: Decision (allow + decision_id / deny)
S->>Ch: acquire permit (limit · fairness · backpressure)
Ch-->>S: permit
S->>IR: lower to neutral IR · resolve target
IR->>B: execute (SET LOCAL app.current_* for RLS)
B-->>S: result
S-->>C: response + catalog/consistency headers (+ write receipt)
Note over S,B: side effects → metrics · audit · CDC · projection · saga · DLQ
```
For a normal gRPC call:
1. `DataBrokerService` receives the RPC in [`src/runtime/service/mod.rs`](src/runtime/service/mod.rs).
2. The handler extracts metadata into a `SecurityContext` and request context.
3. `ensure_ready()` checks the startup lifecycle FSM has reached `Completed`.
4. Catalog compatibility is checked against `x-udb-client-catalog-version`.
5. ABAC policies evaluate service identity, tenant, purpose, operation, scopes,
and message type.
6. A channel permit is acquired through [`src/runtime/channels.rs`](src/runtime/channels.rs);
this is where per-operation limits, fairness, and backpressure live.
7. The request is planned or lowered to neutral IR.
8. A backend target is resolved from project routing, target backend/instance,
circuit breaker state, and plugin registry.
9. The backend executor runs the operation.
10. Responses include catalog/consistency headers; mutations also include a
write receipt when possible.
11. Metrics, audit, CDC, projection, saga, or DLQ paths record side effects as
configured.
The `DataBroker` data-plane contract defines 73 RPCs in
[`proto/udb/services/v1/data_broker.proto`](proto/udb/services/v1/data_broker.proto).
They cover relational, vector, object, cache, document, graph, time-series,
analytical, transaction/2PC, CDC, resource admin, catalog, migration, DLQ, saga,
policy, project, health, and admin/audit surfaces.
Alongside the data plane, UDB now ships a **native control plane** under
`proto/udb/core/**` — six services (Authn, Authz, ApiKey, Tenant, Notification,
Analytics, 77 RPCs total) that run on a **separate, network-isolated listener**
(`UDB_AUTH_GRPC_ADDR`). See [Native Control Plane](#native-control-plane).
## Main Concepts
### Project Protos
Project/application protos are schema input. They do not need to import or
define the UDB `DataBroker` service. UDB parses annotations by suffix, so an
annotation may be canonical like `(udb.table)` or project-qualified like
`(acme.billing.v1.table)`.
The parser supports:
- table and column projections
- primary keys, indexes, foreign keys, checks
- RLS and tenant columns
- vector/cache/object/document/graph/time-series/column/model-registry stores
- proto3 reserved field names and ranges for drift safety
- language options propagated into the manifest
- annotation modes: compat, warn, strict
Key files:
- [`src/parser/mod.rs`](src/parser/mod.rs)
- [`src/parser/options.rs`](src/parser/options.rs)
- [`src/parser/db_parser.rs`](src/parser/db_parser.rs)
- [`src/schema/ast.rs`](src/schema/ast.rs)
- [`docs/annotations.md`](docs/annotations.md)
### Catalog Manifest
The catalog manifest is the broker's normalized view of parsed schemas. It is
where proto messages become tables, columns, stores, projections, security
metadata, language class names, checksums, warnings, and validation errors.
Key files:
- [`src/generation/manifest/mod.rs`](src/generation/manifest/mod.rs)
- [`src/generation/lint.rs`](src/generation/lint.rs)
- [`src/generation/sql/`](src/generation/sql)
- [`src/migration/diff.rs`](src/migration/diff.rs)
### Neutral IR
Data-plane operations lower into backend-neutral structs before compiler modules
turn them into SQL, JSON HTTP payloads, key/value operations, object operations,
or CQL/Cypher/etc.
The main IR operations are:
- `LogicalRead`
- `LogicalWrite`
- `LogicalDelete`
- `LogicalSearch`
- `LogicalAggregate`
- `LogicalResourceOp`
Key files:
- [`src/ir/operations.rs`](src/ir/operations.rs)
- [`src/ir/filter.rs`](src/ir/filter.rs)
- [`src/ir/compile/`](src/ir/compile)
- [`src/ir/cross_backend_tests.rs`](src/ir/cross_backend_tests.rs)
### 🗄️ Backend Matrix
UDB separates backend identity from runtime availability:
- `BackendKind` is the known backend enum.
- `BackendTier` groups SQL/cache/vector/object/document/graph/column stores.
- `BackendRole` says whether a backend can be canonical, projection-only, or
both.
- `BackendCapability` declares operation and consistency properties.
- `Backend` plugin structs register backend-specific setup, generation, and
conformance contracts.
The code declares **18 `BackendKind` variants** (`src/backend/mod.rs`), all enabled
in the default feature set. `role` is whether a backend can host UDB system tables
(canonical) or only serve reads/projections; `RLS` is how per-tenant context is
enforced (Postgres/MySQL/SQLite session GUCs, key-prefix for KV/object, filter
predicate for document/vector). Slim builds compile a subset, e.g.
`--no-default-features --features postgres`.
| Postgres | SQL | `postgres` (always on) | canonical | relational CRUD, tx | yes / **2PC** | session GUC |
| MySQL | SQL | `mysql` | canonical | relational CRUD, tx | yes / **XA+2PC** | session GUC |
| SQLite | SQL | `sqlite` | canonical | relational CRUD, tx | yes / — | context table |
| SQL Server | SQL | `mssql` | projection | relational CRUD, tx | yes / — | `SESSION_CONTEXT` |
| ClickHouse | column | `clickhouse` | both | analytical query, mutate | — | session setting |
| Redis | cache | `redis` | projection | cache get/set/del/scan | — | key prefix |
| Memcached | cache | `memcached` | projection | cache get/set | — | key prefix |
| Qdrant | vector | `qdrant` | projection | vector search/upsert | — | filter |
| Weaviate | vector | `weaviate` | projection | vector + hybrid search | — | filter |
| Pinecone | vector | `pinecone` | projection | vector + hybrid search | — | filter |
| MinIO | object | `s3` | projection | object put/get/presign | — | key prefix |
| S3 | object | `s3` | projection | object put/get/presign | — | key prefix |
| Azure Blob | object | `azureblob` | projection | object put/get | — | key prefix |
| Google Cloud Storage | object | `gcs` | projection | object put/get | — | key prefix |
| MongoDB | document | `mongodb` | canonical | document find/upsert, tx | yes / — | filter |
| Elasticsearch | search | `elasticsearch` | projection | search + hybrid | — | filter |
| Neo4j | graph | `neo4j` | both | graph query/mutate, tx | yes / — | Cypher param |
| Cassandra / ScyllaDB | column | `cassandra` | projection | wide-column query/mutate | LWT only | partition key |
Postgres is always compiled (never feature-gated); the other 17 are gated. MinIO
and S3 share the `s3` feature. `src/backend/mod.rs` is the source of truth for the
full `BackendCapability` matrix (transactions, XA/2PC, RLS, vector/hybrid search,
TTL, object-store, migration-ledger, consistency model).
### Canonical Stores
This is the most important architectural transition in the repo.
Older UDB paths assumed Postgres was the canonical store for system tables,
CDC, saga state, projection task state, migration audit, and consistency fences.
The newer peer-to-peer work introduces:
- `CanonicalStore`
- `DurabilityToken`
- `SystemStores`
- `CanonicalStoreRegistry`
- Postgres, MySQL, and SQLite implementations for system-store traits
Key files:
- [`src/runtime/canonical_store/mod.rs`](src/runtime/canonical_store/mod.rs)
- [`src/runtime/canonical_store/system_store.rs`](src/runtime/canonical_store/system_store.rs)
- [`src/runtime/canonical_store/postgres.rs`](src/runtime/canonical_store/postgres.rs)
- [`src/runtime/canonical_store/mysql.rs`](src/runtime/canonical_store/mysql.rs)
- [`src/runtime/canonical_store/sqlite.rs`](src/runtime/canonical_store/sqlite.rs)
- [`docs/architecture.md`](docs/architecture.md)
Do not read "universal DB layer" as "every backend has identical semantics."
The code tries to be explicit about what compiles, what is unsupported, and what
is eventually consistent or projection-only.
### Runtime System Tables
UDB owns internal catalog/system tables for:
- catalog versions and activation logs
- project catalog bindings
- migration runs and operation ledgers
- CDC event journal, offsets, lock log, control table, topic policy, DLQ
- saga coordinator
- projection tasks
- ABAC policies
- admin audit log
Preview the DDL:
```powershell
cargo run --bin udb-proto-parser -- system-ddl
```
Related files:
- [`src/runtime/system.rs`](src/runtime/system.rs)
- [`src/control/lifecycle.rs`](src/control/lifecycle.rs)
- [`src/runtime/core/catalog_sql.rs`](src/runtime/core/catalog_sql.rs)
- [`src/runtime/core/catalog_admin.rs`](src/runtime/core/catalog_admin.rs)
## Repository Layout
| [`src/lib.rs`](src/lib.rs) | Public library surface and compatibility re-exports |
| [`src/main.rs`](src/main.rs) | Binary entry point |
| [`src/cli`](src/cli) | CLI parsing and command handlers |
| [`src/parser`](src/parser) | Proto lexer/parser and annotation extraction |
| [`src/schema`](src/schema) | AST and checksum types |
| [`src/generation`](src/generation) | Manifest/SQL/DSN/drift/lint generation |
| [`src/ir`](src/ir) | Backend-neutral operation model and compilers |
| [`src/backend`](src/backend) | Backend inventory, plugin trait, capability matrix |
| [`src/runtime`](src/runtime) | Broker runtime, service handlers, backend executors, CDC, security, metrics |
| [`src/migration`](src/migration) | Migration diff/apply/sync/phase-runner |
| [`src/control`](src/control) | Startup lifecycle, FSM, approval, hooks, notifications |
| [`proto`](proto/README.md) | UDB-owned gRPC/protobuf contract |
| [`sdk`](sdk/README.md) | Generated/wrapped clients |
| [`examples`](examples) | Arbitrary project, multi-project, and toy plugin examples |
| [`configs`](configs) | YAML config examples |
| [`docs`](docs) | Operational docs, security, upgrade history, runbooks |
| [`crates/udb-portable`](crates/udb-portable) | WASM/edge parser/checksum/schema-cache subset |
## Quick Start For Developers
The fastest meaningful flow is to use the arbitrary project example, because
the UDB-owned protocol protos are service definitions, not domain schemas.
```powershell
cargo test --lib
```
```powershell
cargo run --bin udb-proto-parser -- lint examples/go_arbitary_project/proto --human
cargo run --bin udb-proto-parser -- catalog examples/go_arbitary_project/proto
cargo run --bin udb-proto-parser -- sql examples/go_arbitary_project/proto
cargo run --bin udb-proto-parser -- plan examples/go_arbitary_project/proto
```
Run a Postgres-backed broker locally:
```powershell
Copy-Item .env.example .env.local
$env:UDB_PG_DSN = "postgresql://udb:udb@localhost:5432/udb?sslmode=prefer"
$env:UDB_ABAC_DEFAULT_ALLOW = "true"
cargo run --bin udb-proto-parser -- serve examples/go_arbitary_project/proto "" 0.0.0.0:50051
```
Run local readiness checks:
```powershell
cargo run --bin udb-proto-parser -- doctor --human
cargo run --bin udb-proto-parser -- doctor --probe --human
```
## CLI
The binary is `udb-proto-parser`. Its name is older than its current scope; it
now drives parsing, generation, runtime serving, migration/admin checks, and the
local playground.
Schema and planning:
```powershell
cargo run --bin udb-proto-parser -- catalog <proto-root> [namespace]
cargo run --bin udb-proto-parser -- dsn <proto-root>
cargo run --bin udb-proto-parser -- sql <proto-root>
cargo run --bin udb-proto-parser -- plan <proto-root>
cargo run --bin udb-proto-parser -- lint <proto-root> --human
cargo run --bin udb-proto-parser -- drift <proto-root> --prior old_manifest.json
cargo run --bin udb-proto-parser -- explain <proto-root>
cargo run --bin udb-proto-parser -- manifest-export <proto-root>
cargo run --bin udb-proto-parser -- field-mask-preview <proto-root>
```
Runtime/admin:
```powershell
cargo run --bin udb-proto-parser -- serve <proto-root> "" 0.0.0.0:50051
cargo run --bin udb-proto-parser -- doctor --probe --human
cargo run --bin udb-proto-parser -- health-check
cargo run --bin udb-proto-parser -- system-ddl
cargo run --bin udb-proto-parser -- tracker-ddl
cargo run --bin udb-proto-parser -- admin dry-run <proto-root>
cargo run --bin udb-proto-parser -- admin force-sync <proto-root>
cargo run --bin udb-proto-parser -- admin verify-audit --limit 250
cargo run --bin udb-proto-parser -- admin release-lock
```
Policy and compatibility:
```powershell
$env:UDB_ABAC_POLICY_FILE = "docs/abac_seed.json"
cargo run --bin udb-proto-parser -- policy-lint
cargo run --bin udb-proto-parser -- policy-seed
cargo run --bin udb-proto-parser -- compat-matrix
cargo run --bin udb-proto-parser -- config-skeleton
```
Playground wrapper:
```powershell
cargo run --bin udb-proto-parser -- dev up
cargo run --bin udb-proto-parser -- dev status
cargo run --bin udb-proto-parser -- dev logs udb
cargo run --bin udb-proto-parser -- dev smoke
cargo run --bin udb-proto-parser -- dev down
```
## Configuration
Configuration is loaded as defaults plus optional file plus environment overlay.
The standard config path is `UDB_CONFIG_PATH`; the complete operator template is
[`.env.example`](.env.example). Env files are loaded in this order:
1. OS environment
2. `.env.<APP_ENV>`
3. `.env.local`
4. `.env.prod`
5. `.env`
Minimum required env for a normal Postgres-backed broker:
| `APP_ENV` | Selects `.env.<APP_ENV>` and labels the runtime environment |
| `UDB_ENV` | Security-mode switch; `production`/`prod` enables stricter defaults |
| `UDB_APP_NAME` | Broker/application identity |
| `UDB_PG_INSTANCES` | Named Postgres instances, usually `primary` |
| `UDB_PG_DSN_PRIMARY` | DSN for the named `primary` instance |
| `UDB_PG_DSN` or `DATABASE_URL` | Canonical primary Postgres DSN |
| `UDB_2PC_ENABLED` | Enables real Postgres prepared-transaction 2PC when `true` |
Common optional env variables:
| `UDB_CONFIG_PATH` | YAML/JSON/TOML runtime config path |
| `UDB_BACKEND_INSTANCES` | Named backend instance descriptor list |
| `UDB_REDIS_DSN` | Redis cache/rate-limit/idempotency |
| `UDB_QDRANT_URL` | Qdrant vector backend |
| `UDB_MINIO_ENDPOINT`, `UDB_MINIO_ACCESS_KEY`, `UDB_MINIO_SECRET_KEY` | MinIO/S3-compatible object storage |
| `UDB_NOSQL_DSN`, `UDB_NOSQL_API_URL` | MongoDB/Atlas Data API backend |
| `UDB_GRAPH_DSN`, `UDB_GRAPH_HTTP_URL` | Neo4j graph backend |
| `UDB_COLUMN_DSN`, `UDB_COLUMN_HTTP_URL` | ClickHouse column backend |
| `UDB_KAFKA_BROKERS` | Kafka brokers for CDC |
| `UDB_ABAC_DEFAULT_ALLOW` | Development-only relaxed authorization |
| `UDB_ALLOW_DEGRADED_BACKENDS` | Allow startup with optional backend failures |
| `UDB_METRICS_ADDR` | Prometheus scrape address, default `0.0.0.0:50052` |
| `UDB_GRPC_ADDR` | Default serve address when not supplied positionally |
| `UDB_TLS_*`, `UDB_MTLS_*` | Server TLS and client CA config |
See:
- [`.env.example`](.env.example)
- [`configs/database.yaml`](configs/database.yaml)
- [`configs/backends.yaml`](configs/backends.yaml)
- [`configs/services.yaml`](configs/services.yaml)
- [`src/runtime/config/mod.rs`](src/runtime/config/mod.rs)
## Security Model
UDB authorization is request-context based. Every non-health request should carry:
- `x-tenant-id`
- `x-user-id`
- `x-purpose`
- `x-correlation-id`
- `x-scopes`
- `x-service-identity`
- `x-udb-project-id`
- `x-udb-client-catalog-version`
The runtime supports:
- JWT service identity
- mTLS service identity
- dev-only header fallback
- ABAC policy evaluation
- PII masking
- field-level encryption
- tenant-aware request context injection
- audit logging
- admin audit hash-chain verification
- topic-policy enforcement for CDC
Start here:
- [`src/runtime/security.rs`](src/runtime/security.rs)
- [`src/runtime/service/mod.rs`](src/runtime/service/mod.rs)
- [`docs/security.md`](docs/security.md)
- [`docs/integration.md`](docs/integration.md)
## Native Control Plane
Beyond the data-plane `DataBroker`, UDB serves a UDB-owned **auth/admin control
plane** defined under `proto/udb/core/**`. These six services are **network-isolated
on a separate listener** (`UDB_AUTH_GRPC_ADDR`, default loopback `port+10`) because
they are a policy decision point that accepts the subject principal as input — they
must not sit on the public `DataBroker` port where any client could assert identity.
All of them are proto-driven (`NativeModel`) and Postgres-backed, failing closed
when no PG pool is configured; their tables are generated from the embedded
`proto/udb/core/**` manifest through the normal migration path.
```mermaid
flowchart TB
APP["📱 app / SDK<br/>(6 languages)"]
PEP["🛡️ trusted PEP / gateway"]
subgraph UDB["UDB process"]
direction TB
PUB["🌐 public listener<br/><b>DataBroker</b> · 73 RPCs"]
INT["🔒 internal listener · UDB_AUTH_GRPC_ADDR<br/><b>Authn · Authz · ApiKey · Tenant · Notification · Analytics</b>"]
end
BK[("🗄️ 18 backends")]
KAFKA["📨 Kafka (CDC / events)"]
APP -->|"data RPCs"| PUB
PEP -->|"auth / admin RPCs"| INT
INT -.->|"Authorize / GetNativeAccess"| PUB
PUB --> BK
INT --> BK
PUB -.->|"outbox relay"| KAFKA
INT -.->|"auth/notification events"| KAFKA
classDef pub fill:#0e7490,stroke:#083344,color:#fff;
classDef int fill:#7c3aed,stroke:#3b0764,color:#fff;
class PUB pub;
class INT int;
```
| `AuthnService` | `core/authn` | 23 | Authenticate (JWT / session / API key / external), login/logout, **RS256 JWT signing + refresh**, sessions, **TOTP MFA**, **CSRF**, OTP, user admin |
| `AuthzService` | `core/authz` | 23 | `Authorize`/`CheckAccess`/batch over RBAC+ABAC+ReBAC (Casbin), role/policy/relationship CRUD, audit decisions, **`GetNativeAccess`**, **`GetPolicyBundle`** |
| `ApiKeyService` | `core/apikey` | 7 | Create/get/list/update/revoke/validate API keys + usage stats |
| `TenantService` | `core/tenant` | 6 | Tenant + tenant-config CRUD |
| `NotificationService` | `core/notification` | 11 | Notifications, templates, preferences, delivery stats (emits `udb.notification.sent.v1` to Kafka) |
| `AnalyticsService` | `core/analytics` | 7 | Pipeline metrics, executor performance, reconciliation, throughput, SLA compliance |
Key capabilities:
- **Identity**: native JWT (static PEM or JWKS URL with `kid` rotation), UDB-issued
RS256 access tokens + refresh tokens (`UDB_JWT_PRIVATE_KEY`), Argon2id passwords
(legacy keyed-HMAC auto-upgraded on login), RFC 6238 TOTP MFA, server-side sessions
with idle/absolute TTL + revocation, mTLS SAN identity, and a hybrid external-identity
bridge (the external provider proves *who*; UDB authz still decides *what*).
- **Authorization**: one engine for RBAC (roles + bindings), ABAC (attribute
conditions), and simple ReBAC (relationship tuples) with tenant/project domains,
explicit-deny-wins, priority, and deterministic `decision_id` + audit records.
`UDB_AUTHZ_V2` (default **on**) routes broker enforcement through it.
- **Native fast path**: `GetNativeAccess` authorizes a request and, when allowed,
mints a short-lived restricted-role DSN plus the exact `app.current_*` session
variables to `SET LOCAL`, so an SDK can talk to Postgres directly while the
broker-generated RLS still applies.
- **Offline SDK authz**: `GetPolicyBundle` returns an HMAC-signed, time-boxed
snapshot the SDK caches to answer `can()` locally.
Source: [`src/runtime/authn/`](src/runtime/authn), [`src/runtime/authz/`](src/runtime/authz),
[`src/runtime/service/auth_service/`](src/runtime/service/auth_service),
[`docs/native-services.md`](docs/native-services.md).
## Protocol And SDKs
The UDB-owned broker contract is:
- [`proto/udb/entity/v1/types.proto`](proto/udb/entity/v1/types.proto)
- [`proto/udb/events/v1/udb_events.proto`](proto/udb/events/v1/udb_events.proto)
- [`proto/udb/services/v1/data_broker.proto`](proto/udb/services/v1/data_broker.proto)
The build script compiles those with `tonic-build` and writes a generated
`protocol.rs` include under Cargo's `OUT_DIR`.
Generate SDKs:
```powershell
.\scripts\gen_sdk.ps1
```
```bash
./scripts/gen_sdk.sh
```
SDK folders:
| Go | [`sdk/go`](sdk/go/README.md) |
| Python | [`sdk/python`](sdk/python/README.md) |
| TypeScript | [`sdk/typescript`](sdk/typescript/README.md) |
| C# | [`sdk/csharp`](sdk/csharp/README.md) |
| Java | [`sdk/java`](sdk/java/README.md) |
| PHP / Laravel | [`sdk/php`](sdk/php/README.md) |
Protocol version: [`sdk/UDB_PROTOCOL_VERSION`](sdk/UDB_PROTOCOL_VERSION).
## 🚀 Quickstart Per Language
Most SDKs ship generated stubs (in each SDK's `gen/` dir — no regen needed to
consume), a thin **broker client** that attaches the request metadata headers
(`x-tenant-id`, `x-user-id`, `x-purpose`, `x-correlation-id`, `x-scopes`,
`x-service-identity`, `x-udb-project-id`, `x-udb-client-catalog-version`), and an
**auth client** (`Authenticate` + `Authorize`/`can`). To regenerate after editing
protos: `buf generate` (or `scripts/gen_sdk.{ps1,sh}`).
> **TypeScript note:** the Node SDK (`@udb_plus/sdk`) loads the protos
> **dynamically** at runtime via `@grpc/proto-loader` (the `.proto` files are
> bundled into the package and resolved by `protoRoot.ts`) — you consume it
> through the package entry points (`@udb_plus/sdk`, `/client`, `/auth`), not by
> importing the `gen/` stubs. The committed `sdk/typescript/gen/**` tree is a
> buf drift-parity artifact (kept in lockstep with the protos by CI) and is
> intentionally excluded from the published package and the build; it requires
> `@bufbuild/protobuf` and is **not** part of the SDK's runtime. See
> [`sdk/typescript/gen/README.md`](sdk/typescript/gen/README.md).
<details open>
<summary><b>🐹 Go</b> — <code>go get github.com/fahara02/udb/sdk/go</code></summary>
```go
import (
entityv1 "github.com/fahara02/udb/sdk/go/gen/udb/entity/v1"
authzv1 "github.com/fahara02/udb/sdk/go/gen/udb/core/authz/services/v1"
"github.com/fahara02/udb/sdk/go/udbclient"
"google.golang.org/grpc"
"google.golang.org/grpc/credentials/insecure"
)
conn, _ := grpc.NewClient("localhost:50051", grpc.WithTransportCredentials(insecure.NewCredentials()))
meta := udbclient.Metadata{TenantID: "acme", UserID: "user-1", Purpose: "web.request",
Scopes: []string{"udb:read", "udb:write"}, ServiceIdentity: "billing.api",
ClientCatalogVersion: udbclient.ProtocolVersion}
udb := udbclient.New(conn, meta)
rs, _ := udb.Select(ctx, &entityv1.SelectRequest{MessageType: "acme.billing.v1.Invoice", Limit: 50})
auth := udbclient.NewAuthClient(conn, meta)
allowed, decision, _ := auth.Can(ctx, &authzv1.ResourceRef{MessageType: "acme.billing.v1.Invoice"}, "read", "")
// native fast path: grant, _ := auth.NativeAccess(ctx, res, "data.select", ""); udbclient.WithNativeTx(ctx, db, grant, fn)
```
</details>
<details>
<summary><b>🐍 Python</b> — <code>pip install udb-client</code></summary>
```python
from udb_client import Metadata, UdbClient, decode_records
from udb_client.auth import UdbAuthClient
from udb.core.authz.services.v1 import core_pb2 as authz
meta = Metadata(tenant_id="acme", user_id="user-1", purpose="billing.demo",
correlation_id="demo-001", scopes=("udb:read", "udb:write"))
with UdbClient("127.0.0.1:50051", meta) as udb:
udb.upsert(message_type="acme.billing.v1.Customer",
record={"customer_id": "cus_001", "tenant_id": "acme"},
conflict_fields=("customer_id",))
rs = udb.select(message_type="acme.billing.v1.Customer", limit=10)
print(decode_records(rs))
with UdbAuthClient("127.0.0.1:50051", meta) as auth:
allowed, decision = auth.can(authz.ResourceRef(message_type="acme.billing.v1.Customer"), "read")
```
</details>
<details>
<summary><b>🟦 TypeScript</b> — <code>npm i @grpc/grpc-js @grpc/proto-loader</code></summary>
```ts
import { dataBrokerClient, metadata, UdbMetadata } from "@udb_plus/sdk/client";
import { UdbAuthClient } from "@udb_plus/sdk/auth";
const meta: UdbMetadata = { tenantId: "acme", userId: "user-1", purpose: "web.request",
scopes: ["udb:read", "udb:write"], serviceIdentity: "billing.api" };
const broker = dataBrokerClient("localhost:50051");
broker.Select({ message_type: "acme.billing.v1.Invoice", limit: 50 }, metadata(meta),
(err: any, rs: any) => console.log(rs?.records));
const auth = new UdbAuthClient("localhost:50051", meta);
const [allowed, decision] = await auth.can({ message_type: "acme.billing.v1.Invoice" }, "read");
```
</details>
<details>
<summary><b>☕ Java</b> — Maven <code>dev.udb:udb-java-client</code> (Java 17)</summary>
```java
import dev.udb.client.*;
import com.udb.entity.v1.Types.*;
import com.udb.core.authz.services.v1.ResourceRef;
var meta = new UdbMetadata("acme", "web.request", "corr-123",
java.util.List.of("udb:read", "udb:write"), "billing.api", "user-1", "default", UdbClient.PROTOCOL_VERSION);
try (UdbClient udb = new UdbClient("localhost:50051", meta)) {
RecordSet rs = udb.select(SelectRequest.newBuilder()
.setMessageType("acme.billing.v1.Invoice").setLimit(50).build());
}
try (UdbAuthClient auth = new UdbAuthClient("localhost:50051", meta)) {
var d = auth.can(ResourceRef.newBuilder().setMessageType("acme.billing.v1.Invoice").build(), "read", "");
}
```
</details>
<details>
<summary><b>🟣 C#</b> — NuGet, target <code>net8.0</code></summary>
```csharp
using Udb.Client; using Udb.Entity.V1;
using AuthzV1 = udb.core.Authz.Services.V1;
await using var udb = new UdbClient("http://localhost:50051", new UdbMetadata(
TenantId: "acme", Purpose: "web.request", CorrelationId: "corr-123",
Scopes: new[] { "udb:read", "udb:write" }, ServiceIdentity: "billing.api", UserId: "user-1"));
RecordSet rs = await udb.SelectAsync(new SelectRequest { MessageType = "acme.billing.v1.Invoice", Limit = 50 });
await using var auth = new UdbAuthClient("http://localhost:50051", /* same meta */ default!);
var (allowed, decision) = await auth.CanAsync(new AuthzV1.ResourceRef { MessageType = "acme.billing.v1.Invoice" }, "read");
```
</details>
<details>
<summary><b>🐘 PHP / Laravel</b> — <code>composer require fahara02/udb-laravel</code> (needs <code>ext-grpc</code>)</summary>
```php
use Fahara02\UdbLaravel\Facades\Udb;
use Udb\Entity\V1\SelectRequest;
use Udb\Core\Authz\Services\V1\ResourceRef;
// request context auto-bound by middleware; pass UdbMetadata explicitly off-request
$rs = Udb::select((new SelectRequest())->setMessageType('acme.billing.v1.Invoice')->setLimit(50));
[$allowed, $decision] = app(\Fahara02\UdbLaravel\UdbAuthClient::class)
->can((new ResourceRef())->setMessageType('acme.billing.v1.Invoice'), 'read');
```
</details>
> Native fast-path transaction helpers (`WithNativeTx`/`native_transaction`/`withNativeTx`)
> and a local TTL authz cache (`AuthzCache`) ship in the Go, Python, and TypeScript
> SDKs; C#/Java/PHP expose `nativeAccess` + `getPolicyBundle` and apply the grant's
> `set_config` session vars manually. Full per-language detail: each SDK's README.
## Testing
Fast local tests:
```powershell
cargo test --lib
```
Backend feature sweeps:
```powershell
cargo test --all-features --lib
cargo test --no-default-features --features postgres --lib
cargo test --features clickhouse,mssql,cassandra --lib
```
Proto contract:
```powershell
buf lint
buf build
buf generate
```
Integration tests are opt-in:
```powershell
docker compose -f docker-compose.integration.yml up -d --wait
$env:UDB_INTEGRATION_TESTS = "1"
cargo test --test integration_tests -- --nocapture
docker compose -f docker-compose.integration.yml down -v --remove-orphans
```
The full default Rust suite is meant to run without external services. Live
Docker/infrastructure tests are guarded by env variables or `#[ignore]`.
See:
- [`TESTING.md`](TESTING.md)
- [`docs/testing.md`](docs/testing.md)
- [`tests/integration_tests.rs`](tests/integration_tests.rs)
- [`tests/parser_tests.rs`](tests/parser_tests.rs)
## Load, Soak, And Operations
Load profiles are scripted through `ghz`:
```powershell
$env:UDB_HOST = "localhost:50051"
$env:CONCURRENCY = "50"
$env:TOTAL_REQUESTS = "10000"
$env:PROFILE = "read-heavy"
.\scripts\load_test.ps1
```
Profiles include:
- `read-heavy`
- `write-heavy`
- `mixed-projection`
- `tenant-noisy-neighbor`
- `backend-outage`
- `reload-during-traffic`
- `multi-project-smoke`
Operational docs:
| Docs index | [`docs/README.md`](docs/README.md) |
| Architecture and backend inventory | [`docs/architecture.md`](docs/architecture.md) |
| Operations, topology, reload, backup, and load profiles | [`docs/operations.md`](docs/operations.md) |
| Security, audit, encryption, and supply chain | [`docs/security.md`](docs/security.md) |
| Testing and live acceptance | [`docs/testing.md`](docs/testing.md) |
## Examples
| [`examples/go_arbitary_project`](examples/go_arbitary_project/README.md) | A Go project namespace UDB does not own; shows table, cache, vector, object, PII, encryption end-to-end |
| [`examples/python_arbitary_project`](examples/python_arbitary_project/README.md) | The same arbitrary-project flow driven from the Python SDK |
| [`examples/php_arbitary_project`](examples/php_arbitary_project/README.md) | The same flow from the PHP/Laravel SDK |
| [`examples/native-services/go`](examples/native-services/go) | Using the native control plane (Authn/Authz/ApiKey/Tenant/Notification/Analytics) from Go |
| [`examples/multi_project`](examples/multi_project/README.md) | One broker serving unrelated projects with separate proto roots/catalogs |
| [`examples/toy_backend_plugin`](examples/toy_backend_plugin/README.md) | Minimal external backend plugin contract |
## Portable Crate
[`crates/udb-portable`](crates/udb-portable) is the browser/edge-safe subset.
It path-includes the same AST, checksum, lexer, and parser source files used by
the main crate. It deliberately excludes `tokio`, `sqlx`, `tonic`, cloud SDKs,
Kafka, Redis, and filesystem directory parsing.
Use it when a client or edge worker needs to parse proto source, compute the
same schema checksum as the server, or track catalog/schema compatibility
without embedding the whole broker.
## Kubernetes
[`deploy/kubernetes`](deploy/kubernetes/README.md) contains CRD contracts for:
- `UdbBroker`
- `UdbProjectCatalog`
- `UdbBackendInstance`
- `UdbMigrationRun`
- `UdbCdcStream`
- `UdbProjectionWorker`
Apply contracts:
```bash
kubectl apply -f deploy/kubernetes/crds/udb.io_crds.yaml
```
These are controller-neutral contracts. The repo contains CRDs, not a complete
operator implementation.
## Supply Chain
The intended gate is:
```powershell
cargo deny check advisories bans licenses sources
```
The policy denies unknown registries, git dependencies, and undocumented source
exceptions. See [`docs/security.md`](docs/security.md).
## Known Rough Edges
- Some newer backend plugins are still plugin-owned rather than fully covered by
one universal connection lifecycle.
- Disabled-feature reporting should be aligned for every backend plugin.
- Some Docker/package paths still reflect older monorepo layouts.
- The default build intentionally pulls many backend SDKs; use slim feature
builds to check dependency hygiene.
- Several live acceptance gates in the docs require real infrastructure and are
not satisfied by code-only tests.
- The crate currently warns on unused/dead code during build; the warnings are
tracked by the refactor history and are not treated as fatal yet.
## Where To Start When Changing Code
| Add or change proto annotation parsing | [`src/parser/options.rs`](src/parser/options.rs), [`src/parser/db_parser.rs`](src/parser/db_parser.rs), [`src/schema/ast.rs`](src/schema/ast.rs) |
| Add a backend operation | [`src/ir/operations.rs`](src/ir/operations.rs), [`src/ir/compile`](src/ir/compile), [`src/runtime/executors`](src/runtime/executors) |
| Add a backend plugin | [`src/backend/plugin.rs`](src/backend/plugin.rs), [`src/backend/plugins`](src/backend/plugins), [`examples/toy_backend_plugin`](examples/toy_backend_plugin/README.md) |
| Change gRPC behavior | [`proto/udb/services/v1/data_broker.proto`](proto/udb/services/v1/data_broker.proto), [`src/runtime/service`](src/runtime/service) |
| Change auth or metadata | [`src/runtime/security.rs`](src/runtime/security.rs), [`src/runtime/service/mod.rs`](src/runtime/service/mod.rs), [`src/embedded.rs`](src/embedded.rs) |
| Change catalog/migration behavior | [`src/generation/manifest`](src/generation/manifest), [`src/migration`](src/migration), [`src/control/lifecycle.rs`](src/control/lifecycle.rs) |
| Change system-store behavior | [`src/runtime/canonical_store`](src/runtime/canonical_store), [`src/runtime/system.rs`](src/runtime/system.rs) |
| Change config loading | [`src/runtime/config`](src/runtime/config), [`src/cli/env_setup.rs`](src/cli/env_setup.rs), [`build.rs`](build.rs) |