Architectures with Private subnets are a relic of on-premises Three tier architecture.
The assumption with this legacy (non-Cloud) architecture and specifically its data layer, is that data required network isolation in a private subnet. Why? Because historically SQL databases have simple clear text communication protocols.
To meet security shortcomings in the data layer with legacy SQL server protocols, databases are deployed in private non-Internet addressable network zones, to limit data breach risks.
Serverless isn’t only about thinking about “servers” less, it’s also about thinking about your data layer less.
DynamoDB and Amazon S3 - Cloud Object Storage are both disruptive and now mature managed data layer solutions that are importantly publicly addressable on the Internet. Cloud data layers have the security design and controls to not require network isolation!
Sidenote: AWS RDS / Elasticache / Elasticsearch are not fully managed IAM integrated services, therefore not considered serverless or Cloud native.
Yes, you can design a “private” Architecture with serverless technologies, which have an:
- Public facing load balancer - ALB
- Lambda(s) in private subnets
- S3/DynamoDB connected via a “VPC endpoint”
This private design offers little benefit in exchange for a lot of complexity and cost to your Infrastructure.
Lets play devil’s advocate.
Arguments to why you need a Private subnet in your serverless Architecture
Reduce data costs
In a private subnet when you communicate with S3/DynamoDB you don’t incur bandwidth costs, if you’re onboarded with a S3/Dynamodb gateway setup (not easy). If your app needs to communicate on the Internet, you would also need to setup a NAT gateway and that incurs another set of costs.
Furthermore if you communicate across private subnets you can incur costs. So this argument is a pretty poor benefit for the complexity of Private architecture.
Architect for a DNS firewall
To add a DNS firewall, you need a private hosted zone.
No need to access the public internet
When your application doesn’t need access to the Internet, you might argue it should be in a private subnet. However it’s far easier to provision a lambda without any (API gateway) input events and not worry about lower level network boundaries.
Lambdas in a private subnet would need to be explicitly onboarded for every service, which impedes innovation.
There is no real benefit to provision even a lambda that “doesn’t use the Internet” on a private subnet. AWS identity controls are capable!
A private subnet with VPC endpoints should be lower latency, though might be only relevant when downloading / uploading large files.
Compliance / Audit
SOC 2 and PCI DSS might require network isolation, depending on your interpretation.
If you require to communicate with other AWS accounts without exposing data to the Internet, it makes sense to setup a PrivateLink.
The performance benefit is moot, as VPC peering is faster and cheaper.
AWS mandate secure TLS protocols for communicating between accounts. Private Links often result in opaque URLs like https://eu2r4q0h43.execute-api.ap-southeast-1.amazonaws.com/stg/store because having named URL hostnames becomes much harder in a private subnet and complex routing rules.
AWS identity based security protocols like AWSv4 signing and API resource policies, should be used instead of network based Private Links.
 resource policies can be used to whitelist aws:SourceVpce which is network based, but it should be more account based.
Ultimately don’t forget that most likely the other account will be using AWS Dynamodb and S3 which runs on a public subnet. Was the Private Link worth it?
Reduce attack surface area
The surface area to address and attack the application as well as the data layer should be the same in both public (assuming API gateway / lambda / S3 or DynamoDB) / private designs.
One off cost to setting up Private subnets
Some will argue that setting up the private subnets is a one off cost and would be captured as “Instructure as Code”, like CDK.
In many more cases complex VPCs are setup with Clickops making it very difficult to reproduce in a new account.
There is no such thing as a one off cost of complexity. Every lambda needs private subnets defined and its very easy to make mistakes!
Security in depth
Private architectures can be viewed as a layer of protection that’s well understood, widely adopted in practice, and almost completely manageable with automation.
Reddit: Does your Serverless Architecture need Private Subnets?
Unfortunately this is not true in practice. Issues arise from complex network routing. And the benefits are neglibile if you have conducted a threat model.
Why stop using a Private subnet?
Using a managed data layer like AWS S3/DynamoDB, means your data already exists in a public subnet.
A cloud service connected on the Internet doesn’t mean it will communicate on the Internet. Major fault situations aside, connectivity between services in public subnets never use the public Internet.
Why create a complex public-private-public Architecture when a simpler one can suffice?
Unless you need DNS firewall, in some legacy compliance environment or don’t trust the public Internet for communicating between accounts, you probably do not need a Private subnet architecture.
Cloud native security is based on identity controls IAM, not “network zones” like the past.
The complexity of supporting a Private architecture will hurt your Availability / reliability as the Confidentiality, Integrity and Availability Triad will attest!
Sidenote: Zerotrust / Beyondcorp appears to be where we are headed by government mandate in any case!