Support force_path_style for self-hosted S3-compatible object storage

[FredPraca]

Is there an existing discussion that is already proposing this?

• I have searched the existing discussions

What part(s) of Encore does this feature request apply to?

• Encore.ts (TypeScript)
• Encore.go (Go)
• Encore CLI
• Local Development Dashboard
• Encore Cloud
• Other

Is your feature request related to a problem? Please describe it

This is related to the opened PR #2115

Describe the solution you'd like

Hi,

We're running Encore in a self-hosted Kubernetes environment and using SeaweedFS as an S3-compatible object storage backend.

Our infra-config.json contains an S3 endpoint similar to:

{
  "object_storage": [
    {
      "type": "s3",
      "region": "us-east-1",
      "endpoint": "http://seaweed-cluster.default.svc.cluster.local:8333",
      "access_key_id": "...",
      "secret_access_key": {
        "$env": "AWS_SECRET_ACCESS_KEY"
      },
      "buckets": {
        "attachments": {
          "name": "attachments"
        }
      }
    }
  ]
} 

The bucket is declared normally in the application:

export const attachmentsBucket = new Bucket("attachments", {
  versioned: false,
  public: false,
});

Problem

Encore's S3 client uses virtual-hosted-style addressing even when a custom S3 endpoint is configured.

With runtime tracing enabled, we can see the actual request being generated:

PUT http://attachments.seaweed-cluster.default.svc.cluster.local:8333/files/.../file.pdf?x-id=PutObject

This results in a DNS error:

DispatchFailure
ConnectorError
hyper::Error(Connect, ConnectError("dns error", ...))

The configured endpoint itself is reachable and resolves correctly from the same application pod:

seaweed-cluster.default.svc.cluster.local

The problem is that virtual-hosted-style addressing changes the hostname to:

attachments.seaweed-cluster.default.svc.cluster.local

which is not a Kubernetes Service DNS name.

For an S3-compatible service running inside Kubernetes, the expected request is instead path-style:

http://seaweed-cluster.default.svc.cluster.local:8333/attachments/files/.../file.pdf

Existing PR

I found #2115, which appears to implement exactly what is needed by adding force_path_style to the S3 infrastructure configuration and propagating it to the AWS Rust SDK:

if self.cfg.force_path_style {
    s3conf = s3conf.force_path_style(true);
}

with configuration along the lines of:

{
  "type": "s3",
  "endpoint": "http://seaweed-cluster.default.svc.cluster.local:8333",
  "force_path_style": true
}

However, it looks like #2115 was never merged.

Would you consider reviving/merging this functionality, or accepting a new PR based on it?

This option seems important for self-hosted deployments using S3-compatible implementations such as SeaweedFS, MinIO, Ceph, etc. Without it, the alternatives in Kubernetes involve DNS wildcard/rewrite workarounds solely to accommodate the S3 client's addressing mode.

Having an explicit force_path_style setting also seems preferable to automatically enabling path-style for custom endpoints, since it preserves the existing behavior for S3-compatible providers that support virtual-hosted-style addressing.

Thanks!

What is the motivation / use case for changing the behavior?

Allowing to use one of the standard way to address S3 resources.