Hi Yufei, I missed your analysis on the #5246 review [1] in the opening mail. +1 on constructing the defaults from that client-facing set, and I'll follow it for GET /v1/config.
Polaris already puts prefix and namespace-separator in overrides, and advertises endpoints and idempotency-key-lifetime as separate fields, so I'll leave those as they are. For defaults, I'll send rest-page-size and rest-metrics-reporting-enabled when they are present on the catalog. Polaris leaves those keys unset on the catalog entity today, so they only appear if an operator has stored them. polaris.config.* stays out, stripped by prefix as Prithvi suggested, and I'll also drop the registered catalog config keys, including the legacy unprefixed names JB listed. Those remain available on the Management GET catalog, along with default-base-location, which Polaris uses to derive table locations. table-default.* and any leftover custom keys stay off the Iceberg defaults until a later thread names an extension for them. I'll wait a bit for other thoughts on the thread, then open a dedicated PR with a CHANGELOG entry and GetConfigTest coverage of the defaults. Thanks, Vignesh [1] https://github.com/apache/polaris/pull/5246#discussion_r4420704567 On Tue, 10 Nov 2026 at 04:38, Yufei Gu <[email protected]> wrote: > Vignesh, thanks for starting this thread. My earlier analysis[1] seems > to have been missed. I raised this two weeks ago in my review of PR > 5246, including details on which properties should be sent to > clients(what IRC clients expect, what Polaris sends now). It'd be > great to take that original analysis into account. > > 1. https://github.com/apache/polaris/pull/5246#discussion_r4420704567 > > Yufei > > On Sun, Nov 8, 2026 at 10:48 PM Jean-Baptiste Onofré <[email protected]> > wrote: > > > > Hi Vignesh > > > > +1 to define the contract for GET /v1/config, and also +1 to strip by > > prefix as Privith suggested. > > > > However, I think the prefix alone is not enough. A few catalog configs > stil > > have a legacy key without the polaris.config prefix (registered with > > legacyCatalogConfiig() in FeatureConfiguration). > > I'm thinking about: > > - allow.overlapping.table.location > > - allow.unstructured.table.location > > - etc > > > > These are server side configurations, just like the polaris.config.* > ones. > > They are still resolved from catalog properties, so catalogs created with > > the old names would keep sending them in defaults. > > I propose we strip both: > > - any key starting with polaris.config > > - any registered catalog config key > > (PolarisConfiguration.getAllCatalogConfigs() already returns the full > set, > > legacy keys included) > > > > New overrides are then excluded by default, and the legacy keys are > covered > > too. > > > > I checked the Spark plugin: from the merged config it only reads prefix > and > > rest page size, so I don't see any client depending on these keys. > > > > About the next steps, we have to be careful because this changes what > > existing deployments return. I would do it in a dedicated PR with a > > CHANGELOG entry, plus a test in GetConfigTest that checks the content of > > defaults so the contract doesn't drift. The PR #5684 (changing > > ENFORCE_CATALOG_CONFIG_AUTHORIZATION) can follow as a next step. > > > > Regards > > JB > > > > On Sat, Nov 7, 2026 at 12:32 PM vignesh a <[email protected]> > wrote: > > > > > Hi all, > > > > > > Iceberg REST GET /v1/config is how clients bootstrap a > > > catalog [1]. After #5246 [2], a 200 still returns the full > > > catalog properties bag; the endpoint gate is off by default. > > > The CHANGELOG says that default will not flip until > > > client-facing catalog configuration is defined [3]. I'd like > > > to define it. > > > > > > Today a 200 puts CatalogEntity.getPropertiesAsMap() into > > > ConfigResponse.defaults. Overrides already carry prefix and the > > > namespace separator; endpoints and the optional > > > idempotency-key-lifetime are separate fields. Storage config, > > > connection config and secrets live in internal properties and > > > are already omitted. > > > > > > What still goes out in defaults is the whole public properties > > > map: default-base-location, operator custom keys, Iceberg > > > catalog properties, and polaris.config.* catalog overrides > > > (drop-with-purge, unstructured locations, list pagination, > > > credential vending, and so on). > > > > > > Management GET catalog is a separate API, gated by > > > CATALOG_READ_PROPERTIES. Iceberg GET /v1/config exists for REST > > > clients to configure the catalog and HTTP client [1]. I read > > > polaris.config.* as server/operator policy, not Iceberg client > > > configuration. Stripping keys Spark or other clients actually > > > consume would be a breaking change, and I'd rather not add a > > > second privilege (GET_CATALOG_CONFIG_PROPERTIES). One authorized > > > 200, with a defined bag. > > > > > > The plan I have in mind: > > > > > > Keep default-base-location, custom keys and Iceberg catalog > > > properties in /v1/config defaults.. > > > Drop polaris.config.* keys from the Iceberg response; they > > > stay available via Management GET catalog. > > > Leave overrides, endpoints and idempotency-key-lifetime as is. > > > Document the contract, then consider flipping > > > ENFORCE_CATALOG_CONFIG_AUTHORIZATION [3].. > > > > > > If any polaris.config.* key is consumed by an Iceberg client > > > today, it should stay in defaults; please name it here. > > > > > > Thoughts? > > > > > > Thanks, > > > *Vignesh* > > > > > > [1] Iceberg REST spec, GET /v1/config (defaults, then client > > > config, then overrides) > > > [2] https://github.com/apache/polaris/pull/5246 > > > [3] https://github.com/apache/polaris/issues/5684 > > > >
