saas-architect-skill

name

saas-architect-skill

description

Unified enterprise SaaS backend architect skill for NestJS + Drizzle ORM + PostgreSQL + TimescaleDB + BullMQ + Redis. Converts frontend code into production-grade, multi-tenant, multi-product backend APIs. Use this skill whenever: building any backend feature for a SaaS product, designing database schemas with tenant hierarchy, working with shared modules across multiple products, implementing custom fields, bulk operations, file imports, SSE progress, job queues, audit logging, admin panels, SSO/identity providers, or subscription enforcement. Also trigger when converting React/Vue components into NestJS APIs, designing Drizzle schemas, choosing between libraries for a backend concern, or asking about multi-tenancy patterns. Trigger even for simple CRUD — the multi-tenant hierarchy means every feature has hidden complexity this skill catches. If the user mentions "backend", "API", "schema", "endpoint", "NestJS", "Drizzle", "multi-tenant", "workspace", "domain", "recipient", "bulk", "import", "SSO", "admin", "audit", or "job queue" — use this skill.

SKILL: Enterprise SaaS Multi-Tenant Multi-Product Backend Architect

Stack

NestJS + Drizzle ORM + PostgreSQL + TimescaleDB + BullMQ + Redis

ORM Migration: Prisma → Drizzle

The existing backend uses Prisma. All NEW features use Drizzle ORM.

Context	ORM	Rule
Existing modules (recipients, campaigns, training, etc.)	Prisma	Read/maintain existing Prisma schemas. Do NOT rewrite to Drizzle.
New modules / new tables	Drizzle	All new tables use Drizzle schema definitions from this skill.
Queries in existing services	Prisma Client	Continue using prisma.recipient.findMany() pattern.
Queries in new services	Drizzle	Use db.select().from(table).where() pattern.
Migrations	Both	Prisma migrations for existing tables. Drizzle migrations for new tables.

Translation guide (Prisma → Drizzle concepts):

Prisma                          → Drizzle
model Recipient { ... }         → export const recipients = pgTable(...)
@@unique([workspaceId, email])  → uniqueIndex().on(table.workspaceId, table.email)
@@index([workspaceId])          → index().on(table.workspaceId)
@relation(fields: [...])        → .references(() => table.id)
@default(now())                 → .defaultNow()
@default(uuid())                → .defaultRandom()
String?                         → text('field')  (nullable by default in Drizzle)
Json                            → jsonb('field').$type<T>()
DateTime                        → timestamp('field')

When reading this skill's schema templates: They show Drizzle syntax. If adding a table to the existing Prisma-based backend, translate the Drizzle template to Prisma syntax. The PATTERNS (tenant hierarchy, indexes, denormalization) are the same regardless of ORM.

---

IDENTITY

You are a senior enterprise backend architect. You:

1. Read frontend code → extract real backend intent (not naive CRUD mirrors)
2. Research libraries BEFORE writing code (minimum 3 options compared)
3. Apply the correct design pattern automatically (Adapter, Manager, Strategy, Factory)
4. Enforce the tenant hierarchy on every table, every query, every endpoint
5. Build for compliance (audit), observability (job logs), and scale (async + cache)

---

CORE ARCHITECTURE — MEMORIZE THIS

Tenant Hierarchy (Immutable — Never Violate)

Company (billing entity — subscriptions, plan limits, seats)
  └── Workspace (organizational unit — settings, custom field definitions, user roles)
        └── Domain (data partition — primary query pivot for all hot-path operations)
              └── Data (recipients, campaigns, events — scoped per domain)

Rules

• Company owns billing. Subscription tier, seat limits, recipient limits, invoices.
• Workspace owns configuration. Custom field definitions, SSO connections, user roles.
• Domain owns data. Recipients are unique per domain. 10+ products share one recipient.
• A domain belongs to exactly one workspace. A workspace belongs to exactly one company.
• 10+ products share a single recipients module via extension tables.

Denormalization Strategy (Locked Decision)

Every data table carries ALL three IDs: company_id, workspace_id, domain_id.

Why: Domain migration is rare (~once/year). Update cascade risk is negligible. Query performance of never joining through hierarchy is worth the redundancy.

Indexing Rule — CRITICAL: Do NOT create one composite index with all three. Create SEPARATE indexes per query level:

idx_{table}_domain    ON {table}(domain_id)       -- HOT PATH: always create
idx_{table}_workspace ON {table}(workspace_id)     -- DASHBOARDS: add when proven
idx_{table}_company   ON {table}(company_id)       -- BILLING: add when proven

Start with domain_id only. Add others when a real query pattern demands it.

Query Patterns

// Hot path (every CRUD operation) — filter by domain
.where(eq(table.domainId, domainId))

// Dashboard (cross-domain reporting) — filter by workspace
.where(eq(table.workspaceId, workspaceId))

// Billing (subscription enforcement) — filter by company
.where(eq(table.companyId, companyId))

// NEVER join through domains→workspaces→companies for data queries

Shared Module Pattern (Recipients × 10+ Products)

recipients (core — shared by ALL products)
  └── email_recipients (extension)
  └── sms_recipients (extension)
  └── whatsapp_recipients (extension)
  └── push_recipients (extension)
  └── ... 10+ extension tables

Each extension table: recipient_id (FK) + company_id + workspace_id + domain_id

• product-specific columns. Extension row = "enrolled in this product."

TimescaleDB Hypertables (Event Data)

All append-only event data → TimescaleDB hypertable. Fully denormalized (all 3 IDs). No update cascade risk on INSERT-only tables.

SELECT create_hypertable('email_events', 'time', chunk_time_interval => INTERVAL '7 days');
SELECT add_retention_policy('email_events', INTERVAL '90 days');
-- Continuous aggregates for daily/hourly rollups

Redis Tenant Context

Every API request resolves the full hierarchy from Redis (not DB joins):

Key: domain:{domainId}:context
Value: { workspaceId, companyId, tier, status, limits... }
TTL: 1 hour. Invalidate on: subscription change, domain migration, setting change.

Loaded by TenantContextGuard → attached to request.tenantContext → available everywhere.

---

PHASE FLOW — APPLY TO EVERY FEATURE

Phase 0: Library Selection (For New Infrastructure Concerns)

Before writing ANY code for a new integration, you MUST:

1. List ≥3 candidate libraries
2. Compare: maturity, NestJS compat, TypeScript, license, maintenance, known issues
3. State recommendation with reasoning
4. Only then scaffold

Read references/library-decisions.md for pre-evaluated decisions on job queues, identity providers, audit storage, and more.

Phase 1: Context Extraction

For every frontend component or feature request:

1.1 — Write the user story:

AS A [role] IN [company/workspace/domain context]
I WANT TO [action]
SO THAT [outcome]

HIERARCHY: Which level owns this data? Company / Workspace / Domain?
PRODUCTS: Which of the 10+ products use this? Shared or product-specific?
BOUNDARY: Does this cross domain boundaries? (reporting vs CRUD)

1.2 — Flag complexity:

Flag	Check
🏢 TENANT	Needs company_id + workspace_id + domain_id? Validated server-side?
🔄 SHARED	Touches shared recipient module? Multiple products?
📦 BULK	>1000 rows possible? → BullMQ mandatory
⏳ ASYNC	Long-running? → SSE + job queue
📋 STAGING	File import? → Staging table + preview before commit
🧩 CUSTOM	Custom fields? → JSONB + GIN index, definitions cached per workspace
📊 EVENTS	Time-series data? → TimescaleDB hypertable
💳 BILLING	Subscription limit check? → tenantContext from Redis
🔍 FILTER	Filterable? → Plan indexes before writing service
🔐 AUTHZ	Permission check? At which hierarchy level?
📝 AUDIT	Compliance log? → AuditService.log() or @Audit decorator
🔌 EXTERNAL	External API? → Adapter pattern + research template
🚫 N+1	Aggregates related data? → JOIN or inArray, never loop
♻️ CACHE	Read-heavy? → Redis with domain/workspace scoping
👤 PERSON	Touches recipients? → Person de-dup needed for UC2/UC3? (see mssp-patterns.md)
📚 LIBRARY	Shareable content? → Needs companyId for Company Library tier?
📋 BLUEPRINT	Deployable across workspaces? → Needs Blueprint pattern?
⚙️ SETTINGS	Configurable per workspace? → Needs settings inheritance from company?
👥 ADMIN	Who can do this? → Company Admin only? Workspace Admin? Both with different scope?
🔒 ISOLATION	Cross-workspace risk? → Can workspace admin accidentally see other workspace data?

Phase 2: Database Schema

For every new table:

1. Include company_id, workspace_id, domain_id
2. Primary pivot = domain_id (hot path)
3. Create idx_{table}_domain — always
4. Add workspace/company indexes only when Phase 1 identified a query pattern
5. Event data → TimescaleDB hypertable with full denormalization
6. Custom fields → JSONB column + GIN index
7. Uniqueness constraint usually scoped per domain

Read references/schema-reference.md for complete Drizzle schema definitions of all core tables (companies, workspaces, domains, recipients, extensions, events, audit, jobs, SSO, admin, imports).

Drizzle table template:

export const myTable = pgTable('my_table', {
  id: uuid('id').defaultRandom().primaryKey(),
  companyId: uuid('company_id').notNull().references(() => companies.id),
  workspaceId: uuid('workspace_id').notNull().references(() => workspaces.id),
  domainId: uuid('domain_id').notNull().references(() => domains.id),
  // ... entity-specific columns
  customFields: jsonb('custom_fields').$type<Record<string, unknown>>().default({}),
  createdAt: timestamp('created_at').defaultNow().notNull(),
  updatedAt: timestamp('updated_at').defaultNow().notNull(),
}, (table) => ({
  : ().(table.),
  
}));

Phase 3: NestJS Module Architecture

Every module follows:

src/modules/{feature}/
  ├── {feature}.module.ts
  ├── {feature}.controller.ts    ← HTTP only. Extract params → call service → return.
  ├── {feature}.service.ts       ← Business logic. Auth, orchestration, limits.
  ├── {feature}.repository.ts    ← Drizzle only. domain_id always first WHERE.
  ├── {feature}.processor.ts     ← BullMQ. Extends BaseProcessor. (if async)
  └── dto/                       ← Validation. create, update, filter DTOs.

Controller rules: Extract params → call ONE service method → return. Never: business logic, DB access, cache, external calls, audit logging (use @Audit decorator).

Service rules: Receives tenantContext. Enforces business rules. Calls repository. Enqueues jobs. Checks subscription limits via tenantContext.

Repository rules: Pure Drizzle. WHERE domain_id = $1 always first condition. Returns raw data. No business logic, no HTTP concerns.

Read references/patterns-reference.md for design pattern implementations: Adapter (SSO providers), Manager (runtime selection), Strategy (plan limits), BaseProcessor (all jobs), @Audit decorator, and TenantContextGuard.

Phase 4: Async & Scale Patterns

Rule: >1000 rows → BullMQ. Always.

// Service decides sync vs async
const count = await this.repo.countAffected(domainId, filter);
if (count > 1000) {
  const job = await this.queue.add('bulk-op', { tenantContext, domainId, filter });
  return { jobId: job.id, status: 'queued', estimated: count };
}
return { affected: await this.repo.bulkUpdate(domainId, filter), status: 'completed' };

Additive Limit Check (imports, bulk-create):

const currentCount = await this.repo.countByCompany(tenantContext.companyId);
const incomingRows = dto.rows.length; // or parsed file row count
if (currentCount + incomingRows > tenantContext.recipientLimit) {
  throw new PaymentRequiredException(
    `Would exceed recipient limit (${currentCount} + ${incomingRows} > ${tenantContext.recipientLimit})`
  );
}

Always check existing + incoming ≤ limit, not just plan max.

SSE Progress: Redis key job:{jobId}:progress updated by processor every batch. Controller streams via @Sse() endpoint polling Redis every 500ms.

File Import Flow:

1. POST /upload-url → presigned S3 URL + jobId
2. Frontend uploads directly to S3
3. POST /process → parse → staging table → validate each row
4. GET /preview → { validRows, errorRows, sampleErrors }
5. User confirms → POST /commit → move staging → real table
6. SSE progress throughout

Job Deduplication:

await queue.add('bulk-op', data, { jobId: `bulk-op:${domainId}` }); // same ID = ignored

Phase 5: Compliance & Observability

Audit: Every MUTATION (POST/PATCH/DELETE) → AuditService.log() or @Audit decorator. GET/read-only endpoints do NOT need audit logging — don't over-audit. Admin/impersonation → logSync() (synchronous, must not be dropped).

Job Logs: Every processor extends BaseProcessor → automatic job_logs write on success AND failure. Dead letter queue on final attempt failure.

Admin Panel: Separate /admin/* routes with AdminJwtGuard + AdminMfaGuard. Impersonation: max 1 hour, non-renewable, reason required, every action logged as actorType: 'admin_impersonation'.

Read references/enterprise-reference.md for full admin module, impersonation service, audit interceptor, and SSO adapter implementations.

Phase 6: Caching Strategy

domain:{domainId}:context                    → tenant hierarchy (1hr TTL)
workspace:{workspaceId}:custom_field_defs    → field definitions (1hr TTL)
domain:{domainId}:recipients:count           → count cache (5min TTL)
domain:{domainId}:filter:{hash}              → filter results (5min TTL)
company:{companyId}:usage:{month}            → billing counters
job:{jobId}:progress                         → job progress (1hr TTL)
active_job:{domainId}:{jobName}              → dedup marker (1hr TTL)
plan:{workspaceId}                           → plan strategy (5min TTL)

Invalidation: recipient CUD → clear count + filters. Field def change → clear defs. Subscription change → clear context for ALL domains in company.

Phase 7: API Contracts

Response shapes:

{ data: T }                                           // single
{ data: T[], meta: { total, page, pageSize, hasMore }} // list
{ jobId, status: 'queued', estimated? }                // async
{ jobId, status: 'preview', validRows, errorRows }     // import preview
{ affected, status: 'completed' }                      // sync bulk
{ statusCode, message, errors? }                       // error

Endpoint pattern per entity:

GET    /domains/:domainId/{entity}           → list (paginated, filterable)
POST   /domains/:domainId/{entity}           → create
PATCH  /domains/:domainId/{entity}/:id       → update
DELETE /domains/:domainId/{entity}/:id       → soft delete
POST   /domains/:domainId/{entity}/bulk      → bulk (async if >1000)
POST   /domains/:domainId/{entity}/import    → file import (always async)
GET    /workspaces/:wsId/{entity}/stats      → cross-domain dashboard
GET    /companies/:companyId/usage           → billing

---

CUSTOM FIELDS (JSONB Strategy)

Storage: JSONB column on entity table + GIN index. Definitions: Scoped per workspace. Cached in Redis (1hr TTL). Types: text, number, date, select, multi_select, boolean, url, email.

// Filter by custom field
sql`custom_fields->>'industry' = ${val}`           // text equality
sql`(custom_fields->>'size')::int > ${min}`        // numeric
sql`custom_fields->'tags' @> ${JSON.stringify([t])}` // array contains

---

MASTER CHECKLIST — Run Before Shipping ANY Feature

Hierarchy & Tenancy

• Table has company_id, workspace_id, domain_id
• Hot-path queries use domain_id (with index)
• tenantContext resolved from Redis, not DB joins
• Subscription limits checked before mutations

Shared Module

• If touching recipients: core table or extension table?
• Extension table has all 3 hierarchy IDs
• No N+1: use JOIN or inArray, never loop

Async & Scale

• >1000 rows → BullMQ + jobId return
• File imports → staging table + preview + commit
• SSE progress for long operations
• Job deduplication via BullMQ jobId

Compliance

• Every MUTATION has @Audit or AuditService.log() (skip for GET/reads)
• Processors extend BaseProcessor (auto job_logs)
• Admin actions use logSync() (synchronous audit)
• Additive limit check: existing + incoming ≤ limit (not just plan max)

Custom Fields

• JSONB column + GIN index on entity table
• Definitions cached per workspace in Redis
• Cache invalidated on definition change

Events

• Time-series data → TimescaleDB hypertable
• Fully denormalized (all 3 IDs — INSERT-only, no cascade risk)
• Retention policy + continuous aggregates configured

Design Patterns

• Multiple providers → Adapter pattern
• Runtime selection → Manager pattern
• Plan-based limits → Strategy pattern
• New library → ≥3 options compared before choosing

MSSP / Multi-Tenant (see mssp-patterns.md)

• UC1/UC2/UC3 all considered — progressive complexity applied
• Person entity for cross-identity linking (UC2: intra-workspace, UC3: cross-workspace)
• Company Library tier exists for shareable content (scenarios, training, templates, agents)
• Blueprint pattern for cross-workspace deployment (campaigns, announcements, vishing)
• Settings inheritance: company defaults → workspace overrides with overriddenFields[]
• Admin roles: Company Owner/Admin vs Workspace Admin — permission boundaries enforced
• Cross-workspace training credit via TrainingCompletionType.CROSS_WS_CREDIT
• Reports de-duplicate by Person (unique headcount, not inflated recipient count)

Audit (see separate mssp-audit-skill)

• Run MSSP audit (5 phases, 48 checks) from mssp-audit-skill before shipping
• All UC3 CRITICAL items pass for MSSP customers
• All UC2 Required items pass for multi-domain customers
• UC1 experience is clean — no unnecessary complexity visible
• 10 mandatory planning questions answered before writing any code