apf_portal/docs/decisions/0006-persistence-postgresql-prisma.md

status, date, decision-makers, tags

status	date	decision-makers	tags
accepted	2026-04-29	R&D Lead	data backend

Persistence — PostgreSQL with Prisma

Context and Problem Statement

The BFF owns two persistent concerns: session state (covered by Redis in the future ADR-0009) and business data owned directly by the portal — the portal is not a pure proxy. Business data is relational, multi-audience (workforce vs. external customers), audit-sensitive, and must run on-prem.

Which database engine and which data-access layer do we adopt?

Decision Drivers

• On-prem deployment (no managed-service shortcut).
• Multi-audience data with strict isolation between workforce and customer scopes (Row-Level Security desirable).
• ACID transactions and a well-understood operational story (HA, backup, point-in-time recovery).
• TypeScript-first developer experience for the BFF.
• Long-term licensing safety — open source, no recent license-change risk (e.g. SSPL).

Considered Options

Database engine

• PostgreSQL (latest stable major). (Chosen.)
• MariaDB / MySQL.
• SQL Server (would imply Microsoft licensing terms).
• MongoDB / DocumentDB (rejected — not a fit for relational, audit-heavy business data; SSPL concerns).

Data-access layer

• Prisma (latest stable major). (Chosen.)
• Drizzle.
• Kysely.
• TypeORM.
• Plain pg driver + hand-written SQL.

Decision Outcome

Chosen options: PostgreSQL as the engine, Prisma as the data-access layer.

The engine is pinned to the latest stable major at workspace bootstrap, and tracks the upstream support window thereafter.

Prisma is wired into NestJS via the nestjs-prisma integration, with a dedicated PrismaService extending PrismaClient and exposed through DI. Migrations are managed by prisma migrate, committed to the repository, validated in CI, and applied through a controlled deployment step (covered by a future infrastructure ADR).

Consequences

• Good, because PostgreSQL is the de facto enterprise open-source RDBMS — mature HA, replication, RLS, JSONB, full-text search.
• Good, because RLS gives a strong substrate for the workforce/customer isolation called for by ADR-0007 (identity model, future).
• Good, because Prisma's schema-first model is a single source of truth, with type generation and a readable migration log.
• Good, because Prisma's NestJS integration is well-trodden territory.
• Bad, because Prisma adds a generated client and a query engine binary — operational surface to ship and version.
• Bad, because Prisma's escape hatch for advanced PostgreSQL features (advanced CTEs, window functions, RLS policies) is $queryRaw — usable, but a reminder that not everything goes through the ORM. We will not fight Prisma when raw SQL is the right tool.
• Neutral, because Drizzle is a credible "à la pointe" alternative; we accept the more conservative choice for now and re-evaluate at the next major architectural review.

Confirmation

• apps/portal-bff depends on prisma and @prisma/client.
• A single schema.prisma lives at apps/portal-bff/prisma/schema.prisma.
• nestjs-prisma's PrismaModule is imported globally in the BFF.
• CI runs prisma validate and prisma migrate diff against the staging schema.
• Database engine version is pinned in deployment manifests.

Pros and Cons of the Options

PostgreSQL (chosen)

• Good, because mature, open source under the PostgreSQL License (permissive), broad on-prem operational tooling.
• Good, because RLS is first-class — directly serves multi-audience isolation.
• Good, because JSONB + relational hybrid suits BFF data shapes.
• Bad, because operational complexity (HA, backups, tuning) must be owned — true of any on-prem RDBMS.

MariaDB / MySQL

• Good, because operationally well-known.
• Bad, because feature gap vs. PostgreSQL (RLS, JSON ergonomics, CTE maturity, generated columns).

MongoDB

• Good, because the document model fits some BFF caches.
• Bad, because business data here is relational; forcing a document store would push joins into the application — bricolage.
• Bad, because the SSPL license raises enterprise legal complications.

Prisma (chosen)

• Good, because schema-first declarative, strong DX, type generation, mainstream in 2026.
• Good, because a mature NestJS integration exists.
• Bad, because the query engine binary adds a deployment artifact.
• Bad, because some PostgreSQL features require $queryRaw.

Drizzle

• Good, because lighter, closer to SQL, faster, type-safe.
• Bad, because younger (2022) — smaller enterprise track record than Prisma. Kept on the watch list for re-evaluation.

Kysely

• Good, because elegant query-builder, no ORM magic, fully type-safe.
• Bad, because no schema/migrations of its own — must be paired with another tool, increasing surface area.

TypeORM

• Good, because long-standing default in NestJS history.
• Bad, because design issues (decorator hell, runtime metadata fragility, migration story) and slowing maintenance — community momentum has shifted to Prisma and Drizzle.

Plain pg driver + hand-written SQL

• Good, because zero abstraction; full control.
• Bad, because every cross-cutting concern (transactions, connection pooling, type mapping, migrations) is hand-built — the very bricolage we are forbidden.

More Information

• PostgreSQL: https://www.postgresql.org/
• Prisma: https://www.prisma.io/
• nestjs-prisma: https://nestjs-prisma.dev/
• Related ADRs: ADR-0005, ADR-0007 (identity model + RLS use, future), ADR-0009 (sessions in Redis, future), future infrastructure ADRs (Postgres HA, backup).