Hi Yufei,
When the private link still uses the public storage
hostname, DNS and routing already send
{account}.dfs.core.windows.net (or blob) and the
public GCS hosts to the private IP. Polaris already
mints against those names. For that case, DNS is
enough for the storage path.
The catalog endpoint and endpointInternal fields are
for when the HTTP host is a different name, or when
Polaris and clients reach storage through different
hosts. That's the S3 pattern JB described for private
link and Kubernetes. Those are the catalog fields
for that case.
I'll keep Entra and GCP STS/IAM as server identity.
I hope that answers the private-link question.
Thanks,
Vignesh
On Wed, 11 Nov 2026 at 05:10, Yufei Gu <[email protected]> wrote:
> > On private links, if the setup keeps the public storage hostname, it's
> really a DNS and routing concern, and Polaris already mints against
> {account}.dfs.core.windows.net (or blob) and the public GCS IAM and STS
> hosts. The endpoint fields only come into play when Polaris and clients
> reach storage through different hosts, or when the host isn't the public
> cloud name at all. I think that's the split for the private-link use case
> Yufei mentioned.
>
> Vignesh, I'm not sure I understand your reply. Could you please
> clarify what this means in the context of our discussion?
>
> Yufei
>
> On Tue, Nov 10, 2026 at 12:02 PM vignesh a <[email protected]> wrote:
> >
> > Hi all,
> >
> > +1 on aligning Azure and GCS with the S3 storage-config
> > pattern.
> >
> > On private links, if the setup keeps the public storage
> > hostname, it's really a DNS and routing concern, and
> > Polaris already mints against {account}.dfs.core.windows.net
> > (or blob) and the public GCS IAM and STS hosts. The
> > endpoint fields only come into play when Polaris and
> > clients reach storage through different hosts, or when the
> > host isn't the public cloud name at all. I think that's
> > the split for the private-link use case Yufei mentioned.
> >
> > For naming, I'll follow JB's suggestion and use the S3
> > names: endpoint for clients and endpointInternal for
> > Polaris. Clients would get the storage endpoint as FileIO
> > extras (adls.connection-string for Azure, gcs.service.host
> > for GCS), and AzureLocation and the allowed-location checks
> > stay untouched. The new fields go behind a realm flag, off
> > by default. I'll keep stsEndpoint out of this spec for both
> > Azure and GCS, since the Entra authority and GCP STS/IAM
> > hosts belong with server identity, and DefaultAzureCredential
> > should already pick up AZURE_AUTHORITY_HOST.
> >
> > As Dmitri mentioned, these can be first-class fields on the
> > Azure and GCS Storage Configuration, with the named
> > credentials work [5951] [5952] going on in parallel. On
> > universeDomain, DownscopedCredentials builds the STS URL as
> > https://sts.{universe}/v1/token, so a local or custom STS
> > URL would still need handling in GcpCredentialOps.
> >
> > I'll start with the Management API spec PR and follow up
> > with the implementation..
> >
> > Thanks,
> > Vignesh
> >
> > [5951] https://github.com/apache/polaris/pull/5951
> > [5952] https://github.com/apache/polaris/pull/5952
>