Hi Vignesh,

I agree that we have a difference here. It makes sense to align Azure and
GCS with S3.

I have a few comments:
- For S3, we distinguish endpoint (what clients see) from endpointInternal
(what Polaris uses), plus stsEndpoint. In private link or Kubernetes
deployments, the host Polaris reaches is often not the one clients use. I
think the Azure/GCS fields should follow the same logic, with the same
naming, so users won't be surprised.
- The security gate is the main point for me. For Azure and GCS, Polaris
sends bearer token to these hosts (the Entra token for the user delegation
key request, the source GCP token for IAM/STS). Unlike SigV4 on S3, a
bearer token can be replayed. As storage config can be set by catalog
admins, the new fields should be guarded by a realm feature flag, similar
to ALLOW_SETTING_S3_ENDPOINTS. Given the bearer token aspect, I would go
with disabled by default (but happy to discuss that :)). I think it's also
aligned with our thread model: configured endpoints must not redirect
credentials outside the intended trust relationship.
- The Entra authority and the GCP STS/IAM hosts related to the server
identity, not to a specific catalog. I would prefer them as server level
configuration rather than catalog storage config fields. For Entra, I think
DefaultAzureCredential already honors the AZURE_AUTHORITY_HOST env
variable, so it's worth to check whether that's enough before adding a new
Polaris config.
- In the Google Auth version we use (1.55.0), I checked and
DownscopedCredentials buids the STS URL from the domain, and the builder
has no endpoint setter. So a custom STS endpoint can't just be passed
through. We would need either a domain approach or our own token exchange
behind GcpCredentialOps. The IAM credentials side is easier, as
IamCredentialsSettings supports setEndpoint().
- AzureLocation currently rejects any host other than blob/
dfs.core.windows.net. My preference would be to keep locations unchanged
and have the endpoint override only change the network target. The way,
location validation and allowed locations keep working as they do today. On
the client side, we could use Iceberg adls.connection-string-xxx.

As this changes the Management API, I suggest to split the work into two
PRs: one for the spec, and the follow up.

Happy to review the PRs :)

Regards
JB

On Sat, Nov 7, 2026 at 4:46 PM vignesh a <[email protected]> wrote:

> Hi all,
>
> Polaris can already vend S3 credentials against a non-public
> endpoint. Catalog storage configuration includes endpoint and
> stsEndpoint, and the Floci S3 integration tests run
> vended-credentials on that path.
>
> Azure and GCS have no equivalent. Polaris mints Azure
> user-delegation SAS against
> https://{account}.dfs.core.windows.net (or blob), derived from
> the table location. GCS downscoping always uses the public IAM
> and STS hosts. Private-link and emulator deployments cannot
> direct that minting at a local or private endpoint.
>
> The in-tree Floci Azure and GCS integration tests show the
> gap. Iceberg clients receive static shared-key or OAuth tokens
> plus a client host extra, because Polaris cannot vend through
> Floci.
>
> Therefore I propose optional catalog endpoint fields for Azure
> and GCS, analogous to S3 endpoint and stsEndpoint. Polaris
> would use them when minting Azure SAS and GCS downscoped
> tokens, and would pass the storage endpoint to Iceberg clients
> as FileIO extras. Azure token fetch uses the default credential
> chain against public Entra, so an emulator that also issues
> tokens may need an authority override as well.
>
> The Floci Azure and GCS integration tests could then use
> vended-credentials instead of static tokens.
>
> If this direction looks reasonable, I'll put up a PR &
> adjust the shape based on feedback here..
>
> Thanks,
> Vignesh
>

Reply via email to