Hi Yufei,

Polaris will not vend credentials to clients if X-Iceberg-Access-Delegation
is not specified. However, it will vend if a client uses
X-Iceberg-Access-Delegation.

Polaris Servers cannot assume that all clients are well-behaving. I believe
it is valuable for Polaris to provide a configuration option for admin
users to have better control in this situation.

Secondary to that that, I think it should be possible to configure the
Polaris Server itself to use ambient credentials for its own storage access.

This follows from your earlier proposal (to which I agree) to
relegate SKIP_CREDENTIAL_SUBSCOPING_INDIRECTION to test-only use cases [1].

[1] https://lists.apache.org/thread/1trcnbm04zzqztpxkzrzs1pvrlyd49bg

Cheers,
Dmitri.

On Wed, Nov 4, 2026 at 1:53 PM Yufei Gu <[email protected]> wrote:

> If engines have ambient credentials, they don't need to do anything.
> For example, the default behavior of Spark[2] and Trino[1] does not
> send the header `X-Iceberg-Access-Delegation`. Polaris won't vend any
> credentials. This is a more flexible way for users, supporting mixed
> use cases where some engines need credential vending and others don't.
>
> 1. Trino, explicitly documented as disabled by default:
> - [REST catalog
> documentation](
> https://trino.io/docs/current/object-storage/metastores.html#rest-catalog
> ):
> `iceberg.rest-catalog.vended-credentials-enabled` “Defaults to
> `false`.”
> - [Header-setting
> code](
> https://github.com/trinodb/trino/blob/master/plugin/trino-iceberg/src/main/java/io/trino/plugin/iceberg/catalog/rest/IcebergRestCatalogPropertiesProvider.java#L61-L67
> ):
> adds `X-Iceberg-Access-Delegation: vended-credentials` only when
> enabled; custom headers can also be configured.
>
> 2. Spark with Iceberg REST
> - [SparkCatalog](
> https://github.com/apache/iceberg/blob/main/spark/v4.0/spark/src/main/java/org/apache/iceberg/spark/SparkCatalog.java#L156-L162
> ):
> forwards the configured catalog options to Iceberg.
> - [RESTSessionCatalog](
> https://github.com/apache/iceberg/blob/main/core/src/main/java/org/apache/iceberg/rest/RESTSessionCatalog.java#L181-L187
> ):
> constructs HTTP headers using `RESTUtil.configHeaders(config)`.
> - [RESTUtil.configHeaders](
> https://github.com/apache/iceberg/blob/main/core/src/main/java/org/apache/iceberg/rest/RESTUtil.java#L433-L435
> ):
> extracts explicitly configured `header.*` properties; it supplies no
> default delegation value.
>
> Yufei
>
> On Wed, Nov 4, 2026 at 10:20 AM Dmitri Bourlatchkov via dev
> <[email protected]> wrote:
> >
> > Hi All,
> >
> > From my POV the use case is clear and PR [5884] is ready to merge.
> >
> > If there are no further concerns, I propose merging it on Nov 6.
> >
> > Thanks,
> > Dmitri.
> >
> > [5884] https://github.com/apache/polaris/pull/5884
> >
> > On Mon, Nov 2, 2026 at 7:10 PM Dmitri Bourlatchkov <[email protected]>
> wrote:
> >
> > > Hi Yufei,
> > >
> > > > Could you share more details about the workload identity use case?
> > >
> > > Clients (Engines) may have ambient credentials in the execution env.,
> > > which allows them to access storage (e.g. GCS).
> > >
> > > In this case, Polaris does not have to vend credentials. However, with
> the
> > > current GCS storage integration, Polaris cannot avoid vending
> credentials.
> > >
> > > If the admin user elects to use workload identity for clients (ambient
> > > credentials), Polaris credential vending work is useless overhead.
> Moreover
> > > it may present a security gap by exposing different credentials to
> clients,
> > > which are supposed to authenticate via another (workload identity)
> > > channel. I believe it is important for the admin user to be able to
> > > minimize credential exposure via Polaris in this case.
> > >
> > > > I think we need a concrete use case to decide whether we need this
> [...]
> > >
> > > From my POV the use case I described above is self-sufficient.
> > >
> > > A similar practical use case relates to your reasonable and timely
> > > proposal to limit the use of SKIP_CREDENTIAL_SUBSCOPING_INDIRECTION
> [2].
> > >
> > > If we are to take the next step and adapt our Integration Test to act
> as
> > > examples of how to avoid relying on this flag, this results in certain
> > > coding difficulties, which are evident [1] in [5884].
> > >
> > > > [...] where the setting should go.
> > >
> > > The natural place for a new setting that controls access to derived
> > > credentials seems to be in the type-specific storage config [5884], as
> I
> > > proposed in my original email in this thread.
> > >
> > > Do you have an alternative suggestion for how to address this use case?
> > >
> > > Thanks,
> > > Dmitri.
> > >
> > > [1] https://github.com/apache/polaris/pull/5756#discussion_r4362688654
> > >
> > > [2] https://lists.apache.org/thread/1trcnbm04zzqztpxkzrzs1pvrlyd49bg
> > >
> > > [5756] https://github.com/apache/polaris/pull/5756
> > >
> > > [5884] https://github.com/apache/polaris/pull/5884
> > >
> > > On Mon, Nov 2, 2026 at 3:31 PM Yufei Gu <[email protected]> wrote:
> > >
> > >> Hi Dmitri,
> > >>
> > >> Thanks for clarifying. My earlier comment was about clients choosing
> > >> not to request credential vending through X-Iceberg-Access-Delegation.
> > >> I understand that this is a server-side admin decision, which differs
> > >> from clients opting out.
> > >>
> > >> Could you share more details about the workload identity use case? I’m
> > >> still not clear why admins would need to disable credential vending in
> > >> that scenario.
> > >>
> > >> stsUnavailable is for S3-compatible storage without STS support. If
> > >> there’s a similar limitation for Azure/GCS, these settings make sense
> > >> to me. But if this is an admin policy, I’m not sure storage
> > >> configuration is the right place, given that a storage config may
> > >> apply to multiple tables or namespaces in the future (Rishi's
> > >> proposal).
> > >>
> > >> I think we need a concrete use case to decide whether we need this and
> > >> where the setting should go. WDYT?
> > >>
> > >>
> > >> Yufei
> > >>
> > >> On Fri, Oct 30, 2026 at 3:17 PM Dmitri Bourlatchkov via dev
> > >> <[email protected]> wrote:
> > >> >
> > >> > Hi Yufei,
> > >> >
> > >> > > [...] we could use the header to achieve the same result.
> > >> >
> > >> > The proposed change allows admin users to configure Polaris to avoid
> > >> > vending credentials in Azure and GCS environments. It is a
> deliberate
> > >> > server-side admin decision.
> > >> >
> > >> > This was possible for S3 before. PR [5884] aligns Polaris
> capabilities
> > >> in
> > >> > Azure and GCS with existing Polaris capabilities in S3 environments.
> > >> >
> > >> > What alternative do you suggest? I'm not sure I can see that in
> JB's or
> > >> > Ayush's messages.
> > >> >
> > >> > Thanks,
> > >> > Dmitri.
> > >> >
> > >> > [5884] https://github.com/apache/polaris/pull/5884
> > >> >
> > >> > On Thu, Oct 29, 2026 at 7:30 PM Yufei Gu <[email protected]>
> wrote:
> > >> >
> > >> > > Apologies for the late reply.
> > >> > >
> > >> > > Thanks Dmitri for raising this, and Vignesh for working onit. As
> > >> Ayush and
> > >> > > JB mentioned, we could use the header to achieve the same result.
> Do
> > >> we
> > >> > > want to introduce another spec change, or am I missing some use
> cases?
> > >> > >
> > >> > > Yufei
> > >> > >
> > >> > >
> > >> > > On Thu, Oct 29, 2026 at 2:05 PM Dmitri Bourlatchkov <
> [email protected]
> > >> >
> > >> > > wrote:
> > >> > >
> > >> > > > Hi All,
> > >> > > >
> > >> > > > PR [5884] seems to be in a good shape. As far as I can tell, all
> > >> existing
> > >> > > > points have been addressed.
> > >> > > >
> > >> > > > If there are no objections, I propose merging it on Oct 30.
> > >> > > >
> > >> > > > Thanks,
> > >> > > > Dmitri.
> > >> > > >
> > >> > > > [5884] https://github.com/apache/polaris/pull/5884
> > >> > > >
> > >> > > > On Tue, Oct 27, 2026 at 5:46 PM Dmitri Bourlatchkov <
> > >> > > > [email protected]> wrote:
> > >> > > >
> > >> > > > > Hi JB, Ayush,
> > >> > > > >
> > >> > > > > Good points re: DefaultAccessDelegationModeResolver!
> > >> > > > >
> > >> > > > > Vignesh contributed PR [5884] to implement the new properties.
> > >> This PR
> > >> > > > > adjusts the logic in DefaultAccessDelegationModeResolver too.
> > >> Kudos to
> > >> > > > > Vugnesh!
> > >> > > > >
> > >> > > > > As far as I can tell, after [5884] if a client requests
> > >> > > > > "vended-credentials" but the server is configured
> > >> > > > > with skipCredentialDownscoping = true (for example), then the
> > >> client
> > >> > > will
> > >> > > > > receive a clear error message indicating that credential
> vending
> > >> is not
> > >> > > > > available.
> > >> > > > >
> > >> > > > > This is related to the Iceberg API clarification vote [1]. If
> > >> that spec
> > >> > > > > change passes in Iceberg, Polaris will be able to support
> "none"
> > >> as a
> > >> > > > > vending mode for a better client config experience, IMHO.
> > >> > > > >
> > >> > > > > In any case, please review [5884]. I think it still makes
> sense in
> > >> > > > Polaris
> > >> > > > > even without the Iceberg spec change.
> > >> > > > >
> > >> > > > > In Polaris code the main use case may be in tests, but some
> users
> > >> may
> > >> > > > want
> > >> > > > > to consider the new options for workload identity-based
> > >> deployments
> > >> > > too.
> > >> > > > >
> > >> > > > > [1]
> > >> https://lists.apache.org/thread/hc2764vozz7ygx5fjj64104hotrof0lk
> > >> > > > >
> > >> > > > > [5884] https://github.com/apache/polaris/pull/5884
> > >> > > > >
> > >> > > > > Cheers,
> > >> > > > > Dmitri.
> > >> > > > >
> > >> > > > > On Sun, Oct 25, 2026 at 5:53 AM Jean-Baptiste Onofré <
> > >> [email protected]>
> > >> > > > > wrote:
> > >> > > > >
> > >> > > > >> Hi all,
> > >> > > > >>
> > >> > > > >> +1 on the catalog-level
> > >> AzureStorageConfigInfo.skipSasTokenGeneration
> > >> > > > and
> > >> > > > >> GcpStorageConfigInfo.skipCredentialDownscoping properties. A
> > >> typed,
> > >> > > well
> > >> > > > >> defined setting is a better home for "don't vend" than the
> > >> > > > >> SKIP_CREDENTIAL_SUBSCOPING_INDIRECTION flag, which only
> applies
> > >> at the
> > >> > > > >> realm level. It also matches what stsUnavailable already does
> > >> for S3.
> > >> > > > >>
> > >> > > > >> I agree with Ayush on the access delegation mode.
> > >> > > > >>
> > >> > > > >> I think it would be great to clarify two points before
> working
> > >> on the
> > >> > > > >> implementation PR:
> > >> > > > >> - If a catalog opts out of vending and the client asks only
> for
> > >> > > > >> vend-credentials, we should reject the request up front with
> a
> > >> clear
> > >> > > > >> error.
> > >> > > > >> Failing later inside the vending path is harder to diagnose
> > >> imho. This
> > >> > > > >> should apply to all three providers, so the S3 stsUnavailable
> > >> case
> > >> > > > behaves
> > >> > > > >> the same way. If changing the S3 behavior is out of the scope
> > >> for the
> > >> > > > PR,
> > >> > > > >> we can track as a follow-up, but we need consistency here
> imho.
> > >> > > > >> - Both properties should default to false, so existing
> catalogs
> > >> keep
> > >> > > > >> vending credentials. They should be optional in the
> Management
> > >> API, so
> > >> > > > >> this
> > >> > > > >> is a backward compatible addition. We should also document
> them
> > >> next
> > >> > > to
> > >> > > > >> stsUnavailable (in site/content/in-dev/unreleased/), and say
> > >> there
> > >> > > that
> > >> > > > >> they are meant for non-cloud or emulator deployments.
> > >> > > > >>
> > >> > > > >> Happy to review the PRs ;)
> > >> > > > >>
> > >> > > > >> Regards
> > >> > > > >> JB
> > >> > > > >>
> > >> > > > >> On Fri, Oct 23, 2026 at 8:15 PM Dmitri Bourlatchkov <
> > >> [email protected]
> > >> > > >
> > >> > > > >> wrote:
> > >> > > > >>
> > >> > > > >> > Hi All,
> > >> > > > >> >
> > >> > > > >> > As discovered during the review of PR 5756 [1], now
> > >> > > > >> > that SKIP_CREDENTIAL_SUBSCOPING_INDIRECTION is not
> recommended
> > >> for
> > >> > > > >> > mainstream use [3], deployments (e.g. tests) using
> non-cloud
> > >> options
> > >> > > > for
> > >> > > > >> > ADLS and GS (GCS) storage need a mechanism for opting out
> of
> > >> minting
> > >> > > > >> > session tokens and resolving to local (server and/or
> client)
> > >> > > > credentials
> > >> > > > >> > for storage access.
> > >> > > > >> >
> > >> > > > >> > I propose adding the following properties at the
> Management API
> > >> > > level
> > >> > > > >> [1].
> > >> > > > >> >
> > >> > > > >> > * AzureStorageConfigInfo.skipSasTokenGeneration
> > >> > > > >> >
> > >> > > > >> > * GcpStorageConfigInfo.skipCredentialDownscoping
> > >> > > > >> >
> > >> > > > >> > Both names relate to the corresponding native APIs in each
> > >> > > technology.
> > >> > > > >> >
> > >> > > > >> > Setting these properties will effectively disable
> credential
> > >> > > vending.
> > >> > > > >> This
> > >> > > > >> > is similar to stsUnavailable in AwsStorageConfigInfo.
> > >> > > > >> >
> > >> > > > >> > The main purpose of the new properties is establishing a
> > >> > > catalog-level
> > >> > > > >> opt
> > >> > > > >> > in/out mechanism and providing well-defined user-level
> > >> settings.
> > >> > > > >> >
> > >> > > > >> > PolarisStorageIntegration implementations still retain
> value
> > >> without
> > >> > > > >> > credential vending because they potentially define
> > >> non-credential
> > >> > > > >> > properties in StorageAccessConfig. The latter case
> currently
> > >> applies
> > >> > > > >> only
> > >> > > > >> > to S3, but downstream customizations may utilize it too.
> > >> > > > >> >
> > >> > > > >> > [1]
> > >> > > >
> https://github.com/apache/polaris/pull/5756#discussion_r4362688654
> > >> > > > >> >
> > >> > > > >> > [2]
> > >> > > > >> >
> > >> > > > >> >
> > >> > > > >>
> > >> > > >
> > >> > >
> > >>
> https://github.com/apache/polaris/blob/5fafc67a832a8900fefb399abc00549d864470d6/spec/polaris-management-service.yml#L21
> > >> > > > >> >
> > >> > > > >> > [3]
> > >> > > https://lists.apache.org/thread/1trcnbm04zzqztpxkzrzs1pvrlyd49bg
> > >> > > > >> >
> > >> > > > >> > Thoughts?
> > >> > > > >> >
> > >> > > > >> > Thanks,
> > >> > > > >> > Dmitri.
> > >> > > > >> >
> > >> > > > >>
> > >> > > > >
> > >> > > > >
> > >> > > > > --
> > >> > > > > Dmitri Bourlatchkov
> > >> > > > > Senior Staff Software Engineer, Dremio
> > >> > > > > Dremio.com
> > >> > > > > <
> > >> > > >
> > >> > >
> > >>
> https://www.dremio.com/?utm_medium=email&utm_source=signature&utm_term=na&utm_content=email-signature&utm_campaign=email-signature
> > >> > > >
> > >> > > > /
> > >> > > > > Follow Us on LinkedIn <
> https://www.linkedin.com/company/dremio>
> > >> / Get
> > >> > > > > Started <https://www.dremio.com/get-started/>
> > >> > > > >
> > >> > > > >
> > >> > > > > The Agentic Lakehouse
> > >> > > > > The only lakehouse built for agents, managed by agents
> > >> > > > >
> > >> > > > >
> > >> > > > >
> > >> > > >
> > >> > >
> > >>
> > >
>

Reply via email to