The server-side retry idea is interesting, but it won't eliminate the problem completely, will it?
If not, I'd suggest pursuing the server-side retry idea as a separate effort. We still need to settle on what status code to return for TARGET_ENTITY_CONCURRENTLY_MODIFIED: - 429 (without the Retry-After header) - 503 I still think 503 is slightly preferable, but won't fight for it either. Thanks, Alex On Thu, Jun 11, 2026 at 10:24 AM Nándor Kollár <[email protected]> wrote: > > Thanks, Dmitri, for the explanation. It now makes sense to me to > handle the retries on the server side. If there's still a conflict > after a couple of retry attempts, then a 429 response code seems > reasonable to me. > > Thanks, > Nandor > > Dmitri Bourlatchkov <[email protected]> ezt írta (időpont: 2026. jún. > 11., Cs, 0:06): > > > > Hi Nandor, > > > > Rename is fundamentally different from other table operations in that it > > only affect a catalog-owned piece of data, which is the name. > > > > Table metadata or properties are not affected. > > > > I imagine the likely conflict is between a rename and a metadata update, > > which is conceptually not a client-side conflict. The server should be able > > to handle it locally. > > > > However, metadata update conflicts have to bounce to the client > > because Polaris cannot resolve them in most cases. The only exception AFAIK > > is the compact/update conflict [1285]. > > > > A rename/rename conflict will bounce to the client as a 404 on the first > > retry. > > > > [1285] https://github.com/apache/polaris/pull/1285 > > > > Cheers, > > Dmitri. > > > > On Wed, Jun 10, 2026 at 3:47 PM Nándor Kollár <[email protected]> wrote: > > > > > Hi All, > > > > > > I'm not against a server-side retry, but I think in that case we > > > should do the same for other table update operations no? That sounds > > > like a more consistent approach. > > > > > > Thanks, > > > Nandor > > > > > > Dmitri Bourlatchkov <[email protected]> ezt írta (időpont: 2026. jún. > > > 10., Sze, 15:41): > > > > > > > > Hi All, > > > > > > > > How about we make Polaris retry the rename a few times on the server > > > side? > > > > If it gets TARGET_ENTITY_CONCURRENTLY_MODIFIED all the times, we > > > eventually > > > > fail with 429. > > > > > > > > Prolonged optimistic lock failures probably mean that there are, indeed, > > > > too many requests. Ideally we should respond with 429 on all requests > > > > clashing on the entity in question (not just the rename), but I guess it > > > is > > > > not technically feasible ATM. > > > > > > > > WDYT? > > > > > > > > Thanks, > > > > Dmitri. > > > > > > > > On Wed, Jun 10, 2026 at 9:10 AM Alexandre Dutra <[email protected]> > > > wrote: > > > > > > > > > Hi all, > > > > > > > > > > I also reviewed the PR and left some comments. Just to summarize my > > > > > thoughts: > > > > > > > > > > - ENTITY_CANNOT_BE_RESOLVED and CATALOG_PATH_CANNOT_BE_RESOLVED should > > > > > be mapped to 404 -> NoSuchNamespaceException. The comments for them in > > > > > BaseResult are imho inaccurate (they are non-retriable). > > > > > > > > > > - TARGET_ENTITY_CONCURRENTLY_MODIFIED: unfortunately 409 is precluded > > > > > because of the Iceberg spec. I am not a big fan of 429 because it > > > > > could force clients to throttle. So, I think the current proposal of > > > > > 503 -> ServiceUnavailableException is the least worst choice (as it's > > > > > retriable). > > > > > > > > > > Thanks, > > > > > Alex > > > > > > > > > > On Wed, Jun 10, 2026 at 11:18 AM Nándor Kollár < > > > [email protected]> > > > > > wrote: > > > > > > > > > > > > Hi All, > > > > > > > > > > > > The Iceberg REST spec doesn't appear to define 429 as a valid > > > response > > > > > > status for rename operations, and I don't think it's an ideal choice > > > > > > either, since it typically indicates rate-limiting issues rather > > > > > > than > > > > > > conflicting updates. > > > > > > > > > > > > In my opinion, 409 would be the most appropriate status code, but > > > > > > the > > > > > > REST spec reserves it for a different purpose. Perhaps 428 > > > > > > Precondition Required could be used to signal a conflict, but that > > > > > > status is generally intended for GET-then-PUT concurrency scenarios, > > > > > > which doesn't seem to match this case. > > > > > > > > > > > > I think we'll be diverging from the Iceberg spec either way, since > > > > > > it > > > > > > doesn't define a response code for conflicting rename operations. > > > > > > Given that, it's probably better to use a status code that isn't > > > > > > defined by the spec at all (such as 429) than to reuse one that the > > > > > > spec already assigns a different meaning to. Considering this, I > > > > > > vote > > > > > > for 429 as the least worst option. > > > > > > > > > > > > As of ENTITY_CANNOT_BE_RESOLVED, it sounds like a 404 for me too. > > > > > > However, the comment suggests that it may be used for conflict > > > > > > scenarios as well, and client should retry: > > > > > > > > > > > > // the specified entity (and its path) cannot be resolved. There is > > > > > > a > > > > > > possibility that by the > > > > > > // time a call is made by the client to the persistent storage, > > > > > > something has changed due to > > > > > > // concurrent modification(s). The client should retry in that case. > > > > > > ENTITY_CANNOT_BE_RESOLVED(4), > > > > > > > > > > > > Thanks, > > > > > > Nandor > > > > > > > > > > > > Dmitri Bourlatchkov <[email protected]> ezt írta (időpont: 2026. jún. > > > > > > 10., Sze, 0:43): > > > > > > > > > > > > > > Hi All, > > > > > > > > > > > > > > I reviewed PR [4646] (but did not leave any comments in GH, > > > replying > > > > > here) > > > > > > > and the current 500 error is most certainly not correct for this > > > > > failure > > > > > > > mode. 503 is not ideal either, as I commented earlier. > > > > > > > > > > > > > > From the PR I gather that people are generally uncomfortable > > > returning > > > > > a > > > > > > > 409 response because it has a narrow meaning in the Iceberg REST > > > API > > > > > spec. > > > > > > > It is a fair point. > > > > > > > > > > > > > > Re: the TARGET_ENTITY_CONCURRENTLY_MODIFIED case. How about 429 > > > (Too > > > > > Many > > > > > > > Requests)? > > > > > > > > > > > > > > 429 is clearly retryable and does not carry any implications about > > > the > > > > > > > state of the system after handling the request. > > > > > > > > > > > > > > The message could say "Unable to rename entity due to overlapping > > > > > > > concurrent modifications". We do not have to set the Retry-After > > > > > header. > > > > > > > > > > > > > > Re: ENTITY_CANNOT_BE_RESOLVED. I believe this is a solid 404 case. > > > > > > > > > > > > > > WDYT? > > > > > > > > > > > > > > [4646] https://github.com/apache/polaris/pull/4646 > > > > > > > > > > > > > > Cheers, > > > > > > > Dmitri. > > > > > > > > > > > > > > > > > > > > > On Tue, Jun 9, 2026 at 12:02 AM Dmitri Bourlatchkov < > > > [email protected]> > > > > > > > wrote: > > > > > > > > > > > > > > > Hi Nandor, > > > > > > > > > > > > > > > > Good question :) > > > > > > > > > > > > > > > > I did not read the PR yet, but my gut feel is towards the 409 > > > error > > > > > code > > > > > > > > because 5xx generally means a fundamental issue with the service > > > that > > > > > > > > goes beyond the scope of client requests. > > > > > > > > > > > > > > > > In a more general perspective, traditional HTTP status codes are > > > > > often too > > > > > > > > narrow to express all the API minute error details. My personal > > > view > > > > > is > > > > > > > > that a rich payload object in the response can be useful in such > > > > > cases... > > > > > > > > but again that will require a spec change. > > > > > > > > > > > > > > > > That said, if the request does not require additional client > > > input > > > > > for a > > > > > > > > retry, Polaris should retry. I assume we can refactor the code > > > > > > > > to > > > > > clearly > > > > > > > > distinguish retryable and non-retryable failures on the server > > > side. > > > > > That > > > > > > > > part should not require spec changes. > > > > > > > > > > > > > > > > Cheers, > > > > > > > > Dmitri. > > > > > > > > > > > > > > > > On Mon, Jun 8, 2026 at 9:48 AM Nándor Kollár < > > > > > [email protected]> > > > > > > > > wrote: > > > > > > > > > > > > > > > >> Hi all, > > > > > > > >> > > > > > > > >> I'd like to ask for the community's opinion on the appropriate > > > > > > > >> response status code for table/view rename operations when > > > there is > > > > > a > > > > > > > >> conflicting operation in progress. > > > > > > > >> > > > > > > > >> A PR was recently raised [1], which I believe highlighted the > > > > > question > > > > > > > >> of what the correct status code should be in such conflict > > > > > scenarios. > > > > > > > >> To me, the Iceberg REST Catalog specification does not clearly > > > > > address > > > > > > > >> this case. Neither 409 Conflict nor 503 Service Unavailable > > > seems > > > > > > > >> entirely appropriate for indicating to the client that the > > > operation > > > > > > > >> could not be completed due to a conflict and that retrying the > > > > > > > >> operation may succeed. > > > > > > > >> > > > > > > > >> I think 409 Conflict might be the better choice, but that would > > > > > > > >> require a change to the specification. It would also end up > > > serving > > > > > > > >> two different purposes: a non-retriable scenario, where the > > > target > > > > > > > >> name is already reserved, and a retriable scenario, where the > > > > > > > >> operation failed due to a temporary conflict. What do you > > > > > > > >> think? > > > > > > > >> > > > > > > > >> [1] https://github.com/apache/polaris/pull/4646 > > > > > > > >> > > > > > > > >> Thanks, > > > > > > > >> Nandor > > > > > > > >> > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > > > Kollár Nándor > > > > > > > >
