Ali Fouad KwikaCTO · Dubai
All insights · 5 min read · Ali Fouad Kwika

Multi-Tenant SaaS Architecture: Silo, Pool or Bridge?

Multi-Tenant SaaS Architecture: Silo, Pool or Bridge?

Start with a pooled model, where all customers share the same application and database, and isolate tenants in the data layer. Move a customer to dedicated resources only when a contract, a regulator or a performance problem requires it. That is the short answer for most B2B SaaS products at an early stage. The mixed result, some parts shared and some dedicated, is what AWS calls the bridge model, and it is a reasonable place for a growing product to end up.

The rest of this article explains the three models, the questions that decide between them, and the mistakes that are expensive to undo.

The three models

The AWS Well-Architected SaaS Lens uses three terms that have become common vocabulary:

  • Silo: each tenant has dedicated resources, such as its own stack or database, while identity, onboarding and operations are still shared across tenants.
  • Pool: tenants share resources. This is the classic multi-tenant model and the source of SaaS economies of scale.
  • Bridge: a mix. Some services are siloed and others pooled, depending on needs such as regulation, noisy-neighbour behaviour, cost and agility.

Microsoft's Azure Architecture Center describes the same ground with different names: automated single-tenant deployments, fully multitenant deployments, and vertically or horizontally partitioned deployments in between. A horizontally partitioned design, for example, shares the application tier and gives each tenant its own database.

What each model costs you

Microsoft's guidance is direct about the trade-offs, and they match what I look for in an architecture review:

  • Fully shared gives the lowest cost and a single deployment to maintain. The risks are data leaking between tenants, one tenant's load slowing the others, and a bad release reaching every customer at once.
  • Dedicated per tenant gives the strongest isolation and fewer noisy-neighbour effects. Cost grows in line with the number of tenants, and the model only works with heavy automation of deployment and operations.
  • Partitioned designs let you place trial customers on shared infrastructure and premium customers on dedicated infrastructure. The price is a codebase that must support both, and migrations between them.

Five questions that decide the model

  1. What do your contracts promise? Enterprise and government customers may ask for a dedicated database, a named region or their own encryption keys. Read the largest three contracts before drawing the diagram.
  2. What data do you hold? Personal, financial and health data raise the cost of a leak. In the UAE, check which data protection regime applies to you; onshore that is Federal Decree-Law No. 45 of 2021.
  3. How uneven is the load? If one customer can be a hundred times larger than the median, plan limits and a path to dedicated capacity for that customer.
  4. How many tenants will you have? Ten large customers and ten thousand small ones lead to different answers. Per-tenant resources that are comfortable at ten are an operations problem at ten thousand.
  5. How strong is your automation? Every dedicated component must be created, upgraded, monitored and backed up without manual work. If you cannot do that yet, stay pooled.

Isolation in the data layer

In a pooled database, every row belongs to a tenant and every query must respect that. Relying on developers to remember a filter in each query is the weakest form of isolation. Stronger options are a schema or database per tenant, or enforcement by the database itself.

PostgreSQL supports row-level security: you enable it on a table and define policies that decide which rows a role can see or change. With it enabled and no policy in place, access is denied by default. The documentation also lists the limits. Superusers and roles with the BYPASSRLS attribute skip the policies, and so do table owners unless row security is forced on the table. So the application must connect with a role that is none of these, and that rule should be tested, not assumed.

Whatever the mechanism, add automated tests that try to read one tenant's data as another tenant. Run them on every release.

Noisy neighbours and tiers

A noisy neighbour is a tenant whose usage degrades the service for others. The defences are plain: rate limits per tenant, quotas on heavy operations, queues for background work, and per-tenant metrics so you can see who is consuming what. Per-tenant metrics also tell you what each customer costs to serve, which pricing decisions need.

Tiers follow from this. A standard tier runs pooled. A premium tier can buy dedicated capacity or a dedicated database. Design the tenant identifier and the routing layer so that moving a tenant between tiers is an operation, not a project.

Mistakes that are expensive to undo

  • No tenant identifier in the data model from the first day.
  • Tenant context taken from the request instead of from the authenticated session.
  • Shared caches, search indexes or file storage without tenant separation, while the database is carefully isolated.
  • Dedicated environments created by hand for the first big customer, then copied for the next five.
  • One customer's custom code merged into the main product behind flags nobody dares to remove.

Frequently asked questions

Is a database per tenant safer?

It reduces the chance that a query error exposes another tenant's data, and it makes per-customer backup and deletion simpler. It also multiplies migrations, monitoring and cost. It is a good fit for a small number of high-value tenants.

Can I move from pool to silo later?

Yes, if tenant identity is clean in the data model and the routing layer knows where each tenant lives. Without those two things the move is a rewrite.

Does data residency force a silo model?

Not always. A regional deployment that is pooled within the region can meet a residency requirement. Check the wording of the contract or regulation first.

Which model is best for a startup?

Pooled, with isolation enforced in the data layer and tested, and with the tenant identifier everywhere. Add dedicated options when a paying customer needs them.

Where I can help

I am Ali Fouad Kwika, a Dubai-based CTO and founder who has worked in technology since 2011, and the Founder and CTO of SkyLara. I review SaaS and platform architecture for founders and boards, as a fractional CTO or as part of technical due diligence. See also CTO and executive leadership, or contact me.

Technical references in this article were checked on 15 November 2026 against the AWS, Microsoft and PostgreSQL documentation and the UAE legislation portal linked above.

Have a product to build or a team to scale?

Tell me where you are and where you want to be. Let’s find the shortest path to get there.