saas-multi-tenant（SAAS 多租户）

SaaS Multi-Tenant Architecture

When to Use This Skill

• The user is building a SaaS application where multiple customers share the same database
• The user asks about tenant isolation, row-level security, or data leakage prevention
• The user needs to scope every database query to a specific tenant without manual WHERE clauses
• The user asks about shared-schema vs schema-per-tenant vs database-per-tenant tradeoffs
• The user is implementing admin endpoints that must access data across tenants
• The user needs to add tenant_id columns to an existing single-tenant application
• The user asks about PostgreSQL RLS policies for tenant isolation
• The user is building tenant-aware middleware in Express, Fastify, or Next.js API routes

Do NOT use this skill when:

• The user is building a single-user application with no shared infrastructure
• The user asks about authentication only without tenant scoping (use an auth skill instead)
• The user needs general database schema design without multi-tenancy requirements

Core Workflow

1. Determine the tenancy model. Ask the user about their scale expectations and isolation requirements. For most SaaS apps under 1000 tenants, shared-schema with a tenant_id column on every table is the correct default. Schema-per-tenant adds operational overhead (migrations run N times). Database-per-tenant is only justified when tenants have regulatory data residency requirements.
2. Add tenant_id to every tenant-scoped table. The column must be NOT NULL, type UUID or TEXT, and included in every composite index. Never allow a tenant-scoped table to exist without this column — a missing tenant_id is a data leak waiting to happen.
3. Set up PostgreSQL Row-Level Security (RLS). Create a policy on each tenant-scoped table that filters rows by current_setting('app.current_tenant_id'). This acts as a database-level safety net — even if application code forgets a WHERE clause, RLS blocks cross-tenant reads.
4. Build tenant-aware middleware. At the start of every request, extract the tenant_id from the authenticated session or JWT claims. Set it on the database connection using SET LOCAL app.current_tenant_id = '...' inside a transaction. Every subsequent query in that request inherits the tenant scope automatically.
5. Scope all ORM queries by tenant. If using Prisma, apply a global middleware that injects where: {tenantId} into every findMany, findFirst, update, and delete call. If using Drizzle, create a base query builder that includes the tenant filter. Never rely on developers remembering to add the filter manually.
6. Handle tenant-aware migrations. Every new table migration must include tenant_id as a column. Write a linting rule or CI check that rejects any migration creating a table without tenant_id unless the table is explicitly marked as global (e.g., plans, feature_flags).
7. Build cross-tenant admin routes separately. Admin endpoints that aggregate data across tenants must bypass RLS explicitly using SET LOCAL role = 'admin_bypass' or a dedicated database role. These routes must be protected by a separate admin authentication flow — never reuse tenant user sessions for admin access.
8. Implement tenant provisioning. When a new customer signs up, create their tenant record, seed default data (roles, settings, onboarding state), and assign the founding user. Wrap this in a database transaction so partial provisioning never leaves orphan records.

Examples

Example 1: PostgreSQL RLS Policy for Tenant Isolation

-- Enable RLS on the table
ALTER TABLE projects ENABLE ROW LEVEL SECURITY;
ALTER TABLE projects FORCE ROW LEVEL SECURITY;

-- Policy: users can only see rows where tenant_id matches the session variable
CREATE POLICY tenant_isolation ON projects
  USING (tenant_id = current_setting('app.current_tenant_id')::uuid);

-- Policy for INSERT: new rows must match the current tenant
CREATE POLICY tenant_insert ON projects
  FOR INSERT
  WITH CHECK (tenant_id = current_setting('app.current_tenant_id')::uuid);

Example 2: Express Middleware That Sets Tenant Context per Request

import { Pool } from "pg";

const pool = new Pool({ connectionString: process.env.DATABASE_URL });

async function tenantMiddleware(req, res, next) {
  const tenantId = req.auth?.tenantId; // extracted from JWT during auth
  if (!tenantId) return res.status(403).json({ error: "No tenant context" });

  const client = await pool.connect();
  try {
    await client.query("BEGIN");
    // Use set_config — SET LOCAL does not accept bind placeholders ($1)
    await client.query("SELECT set_config('app.current_tenant_id', $1, true)", [tenantId]);
    req.db = client;
    req.tenantId = tenantId;

    // Cleanup on response finish — guarantees release even if handler skips next()
    res.on("finish", async () => {
      try { await client.query("COMMIT"); } catch { await client.query("ROLLBACK"); }
      client.release();
    });

    next();
  } catch (err) {
    await client.query("ROLLBACK").catch(() => {});
    client.release();
    next(err);
  }
}

Example 3: Prisma Middleware for Automatic Tenant Scoping

import { PrismaClient } from "@prisma/client";

// Tables that do NOT have tenant_id (global tables)
const GLOBAL_TABLES = new Set(["Plan", "FeatureFlag", "SystemConfig"]);

function createTenantPrisma(tenantId: string): PrismaClient {
  const prisma = new PrismaClient();

  prisma.$use(async (params, next) => {
    if (GLOBAL_TABLES.has(params.model ?? "")) return next(params);

    // Initialize args.where — Prisma passes undefined args for calls like findMany()
    params.args = params.args ?? {};
    params.args.where = params.args.where ?? {};

    // Inject tenant filter on reads (skip findUnique — it only accepts unique-field selectors)
    if (["findMany", "findFirst", "count", "aggregate"].includes(params.action)) {
      params.args.where = { ...params.args.where, tenantId };
    }

    // Inject tenant_id on creates
    if (["create", "createMany"].includes(params.action)) {
      params.args.data = params.args.data ?? {};
      if (params.action === "createMany") {
        params.args.data = params.args.data.map((d: any) => ({ ...d, tenantId }));
      } else {
        params.args.data = { ...params.args.data, tenantId };
      }
    }

    // Scope updates and deletes
    if (["update", "updateMany", "delete", "deleteMany"].includes(params.action)) {
      params.args.where = { ...params.args.where, tenantId };
    }

    return next(params);
  });

  return prisma;
}

Never Do This

1. Never query a tenant-scoped table without a tenant_id filter. Even if your ORM middleware handles it, raw SQL queries bypass middleware entirely. Every raw query must include WHERE tenant_id = $1 or rely on RLS. A single unscoped SELECT * FROM invoices leaks every customer's billing data.
2. Never store tenant_id only in the application session without enforcing it at the database level. Application-layer filtering is a suggestion. RLS is enforcement. If a bug in your middleware skips the tenant filter, only RLS prevents the data leak. Run both layers.
3. Never use auto-incrementing integer IDs for tenant-scoped resources. Sequential IDs (invoice #1042) let attackers enumerate other tenants' resources by incrementing the ID. Use UUIDs for all tenant-scoped primary keys. Reserve integer IDs for internal-only tables.
4. Never let tenant users access admin aggregation endpoints. A route like GET /admin/metrics that queries across all tenants must never be reachable with a regular tenant JWT. Use a separate authentication mechanism (API key, admin role claim with a different issuer) for cross-tenant routes.
5. Never run migrations with RLS enabled on the migration connection. The migration user needs to create tables, add columns, and modify policies. If RLS is active on the migration connection, ALTER TABLE commands may silently fail or affect only the "current tenant's" view. Use a dedicated superuser or bypassrls role for migrations.
6. Never share connection pools across tenants when using SET LOCAL. If you use SET LOCAL app.current_tenant_id inside a transaction, that setting is scoped to the transaction. But if a previous request's transaction was not properly committed or rolled back, the connection returns to the pool with stale tenant context. Always RESET app.current_tenant_id in the cleanup path.

Edge Cases

1. Tenant deletion and data retention. When a tenant cancels their subscription, you cannot simply DELETE FROM tenants WHERE id = $1. Foreign key cascades may time out on large datasets. Instead, soft-delete the tenant (set deleted_at), revoke all user sessions, then run a background job that deletes tenant data in batches over hours or days.
2. Tenant data export for GDPR/compliance. When a tenant requests a full data export, you need to query every tenant-scoped table for that tenant_id and package it. Build a registry of all tenant-scoped tables (parse your migration files or maintain a manifest) so the export job doesn't miss tables added after the export feature was built.
3. Shared resources between tenants. Some features require shared state — e.g., a marketplace where Tenant A's products are visible to Tenant B's users. These tables need a different RLS policy: read access is public (no tenant filter), but write access is still scoped to the owning tenant. Model these as owner_tenant_id instead of tenant_id.
4. Tenant-aware background jobs. When a cron job or queue worker processes tasks, there is no HTTP request to extract tenant_id from. The job payload must include tenant_id, and the worker must set the database session variable before processing. Never run background jobs without tenant context — they will either fail on RLS or bypass it entirely.
5. Connection pool exhaustion with schema-per-tenant. If you use one PostgreSQL schema per tenant and each schema requires its own connection pool, 500 tenants means 500 pools. This exhausts max_connections fast. Use a connection pooler like PgBouncer in transaction mode, or switch to shared-schema before hitting this wall.

Best Practices

1. Create a tenants table as the single source of truth. Every tenant_id foreign key in every table points back to tenants.id. Include columns for name, slug (for subdomain routing), plan_id, created_at, and deleted_at. This table is the root of your entire data model.
2. Index tenant_id as the first column in every composite index. PostgreSQL uses leftmost prefix matching for composite indexes. An index on (tenant_id, created_at) serves both "all items for tenant X" and "items for tenant X sorted by date." An index on (created_at, tenant_id) only helps date-range queries across all tenants.
3. Use subdomains or path prefixes for tenant routing. acme.yourapp.com or yourapp.com/org/acme — both work. Map the subdomain or path to a tenant_id lookup at the edge (middleware or reverse proxy). This lookup should be cached (Redis or in-memory with 60s TTL) since it runs on every single request.
4. Separate tenant-scoped tables from global tables explicitly. Maintain a list (code constant or database table) of which tables are global (no tenant_id) and which are tenant-scoped. Use this list in your ORM middleware, your migration linter, and your data export job. If a table isn't in either list, the CI check should fail.
5. Test with at least 3 tenants in your seed data. A single tenant in development hides every multi-tenancy bug. Two tenants hides bugs where the first tenant's data leaks to the second but not vice versa. Three tenants catches ordering and filtering bugs that only appear with multiple peers.
6. Rate-limit and quota per tenant, not globally. A global rate limit of 1000 requests/minute means one noisy tenant can exhaust the quota for everyone. Implement per-tenant rate limiting using a Redis key pattern like ratelimit:{tenant_id}:{endpoint} with a sliding window counter.

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.