Signals DPG Architecture

The Signals DPG is the network-aware backend at the heart of Blue Dots. It stores signals as schema-typed items and discovers/matches them across instances.

Components

• apps/api — Fastify + Zod (fastify-type-provider-zod) + Drizzle ORM + Redis. The network-aware API: items, actions, events, instance-local reads and inter-instance discovery. Entry point src/server.ts.
• apps/ui — React 19 + Vite. A schema-driven UI: forms and cards are rendered from the network and item schemas, so new domains need no UI code changes.
• packages/config — Zod env schemas, allowed-origin lists, network-config loader. All env vars are declared here, never parsed ad hoc.
• packages/database — Drizzle setup and partition-aware query helpers.
• packages/schemas — shared Zod schemas for request bodies, admin and the schema registry.
• packages/auth, packages/notification, packages/match_score — service clients and config.

It is a pnpm + Turborepo monorepo; the workspace alias @dpg/* maps to packages/*/src.

Network awareness

An instance serves one or more domains on a network. The two read layers are kept strictly separate:

• Instance-local (GET /api/v1/item/fetch) — reads an instance’s own items with a brief Redis cache.
• Inter-instance (GET /api/v1/network/item/fetch) — count-first discovery: ask peers how many relevant items they hold, select only relevant peers, fetch slices, merge and cache. Schema fetching and caching live in this layer.

This separation is the key scaling decision — see Read & Write Paths.

Data partitioning

Item tables are partitioned in PostgreSQL. Always use the partition-aware query helpers in @dpg/database so the planner can prune; an ad-hoc query across the parent table without a partition key will scan everything.

Engineering conventions

• ESM only, strict TypeScript, no any; import type for type-only imports.
• Files are snake_case. Route handler exports are snake_case (create_item); internal handler functions camelCase; Zod schemas PascalCase; DB columns snake_case.
• Routes never throw — they return reply.code(N).send({ error, message }) with a machine-readable error code, and handle Postgres 23505 (unique) / 23503 (FK) explicitly.
• DB schema lives in apps/api/db/postgres/schema/; migrations are generated, never hand-edited (pnpm db:generate:api).

See Tech Stack for the full toolchain and Data Model for the schema.