Hi All, I maintain my preference for the Admin Tool command in this case.
However, I'm fine to go with the community consensus to add a Management API endpoint for this. Please keep in mind that the number of table in a catalog can be very large. If the new endpoint provides a status per table, it will have to support pagination to avoid OOM attack vectors [1]. Another point is that scanning many table entities and cross-checking their locations (as the next logical stem in evolving this tool) can be a heavy operation to perform inside a live server handling ordinary loadTable requests from query engines. It can skew the load across Polaris nodes. [1] https://lists.apache.org/thread/2d3v9ob15w9fcpkb9qs65tmvhkmm1dcf Cheers, Dmitri. On Thu, Nov 19, 2026 at 1:24 PM Arnab Karmakar <[email protected]> wrote: > Hi all, > > Thanks, Dmitri, Yufei, JB and Vignesh, for the feedback. Here's how I read > the thread and what I think we should do next. > > Where to run the check > > Yufei, JB and Vignesh prefer a live-server check, so a catalog admin can > check the impact with the same credentials before flipping the flag. Dmitri > prefers the Admin Tool because it is standalone and easier to retire along > with the flag. I think the live check fits the main use case best, so I > propose: > > 1. Add a read-only Management API operation that checks one catalog, or > all catalogs in a realm. Validation runs on the server and reuses the > #5856 > logic. A realm-wide check requires the caller to be able to inspect > every > catalog, so the report is never partial. > 2. Expose it through the Polaris CLI. > 3. Treat an Admin Tool command as an optional follow-up for when no > server is available. > > Naming: I'll adopt Dmitri's check-entity-locations. It works whether the > flag is on or off. The exact CLI form will follow the existing CLI > conventions. > > Vignesh is right that the catalog property wins over RealmConfig. Without > an override, the report would evaluate catalogs as unstructured and hide > the tables that would start failing. So I'd keep an explicit "check as > structured" option, for example --as-structured, for the check you run > before flipping the flag. By default the check uses the effective setting, > and the output states which mode was used. An Admin Tool command wouldn't > need the option, since it reads RealmConfig. > > Output > > - The default output is a human-readable summary: the effective setting > per catalog, one line per incompatible table or view with its location > and > reason, then totals. > - --format json prints one JSON object per line, as Dmitri suggested. > - Each finding has a stable reason code. > - The exit code is 0 when nothing is found, 1 when incompatible entities > are found, and another non-zero code for errors. > > Thanks, > Arnab > > On Wed, Nov 18, 2026 at 3:07 PM vignesh a <[email protected]> wrote: > > > Hi all, > > > > +1 from me, I agree with Yufei and JB that the check > > should run against a live server. > > > > The original case is turning off > > polaris.config.allow.unstructured.table.location on a > > catalog. That update doesn't look at table > > locations at all. The check from #5856 only runs later, on load, > > commit, rename or credential vending. > > > > If a report reuses that helper against live catalogs, > > the catalog property wins over RealmConfig [1]. > > Until the property is flipped, those catalogs still > > evaluate as unstructured. The tables that would start > > returning 403 stay hidden. > > > > I'd add a "check as structured" option so a > > catalog admin can run it before flipping. > > > > Thanks, > > Vignesh > > > > [1] > > > > > https://github.com/apache/polaris/blob/ecae57895640839ae692649e1434d3570d4203cc/polaris-core/src/main/java/org/apache/polaris/core/config/PolarisConfiguration.java#L142-L166 > > >
