That's a very good point, I let Gabor and Shengkai follow up on that suggestion :)
Gyula On Fri, Mar 28, 2025 at 8:30 AM Timo Walther <twal...@apache.org> wrote: > Hi Gyula, > > if I understand the discussion correctly, you want to use a PTF without > table arguments to return a table (read from savepoint metadata)? If > this is the case, you don't need a PTF for it. A regular table function > can also do the job. IIRC we support TVF with constant args. > > Cheers, > Timo > > On 28.03.25 08:10, Gyula Fóra wrote: > > Hi Timo! > > > > Thanks for the answers. > > > > Just to give some context here is this thread: > > https://lists.apache.org/thread/08jwrocqyk1q82lnfdldhnyb79m496lp > > > > We were considering a PTF like state_metadata("checkpointpath") to > create a > > table with the available state metadata instead of creating a custom > > connector for reading the metadata. Our thinking was this could > completely > > replace the need for a new connector. > > > > But this would only make sense if state_metadata("checkpointpath") could > > work as a proper table, such as we can make batch operations on it as > well. > > > > Cheers, > > Gyula > > > > On Fri, Mar 28, 2025 at 7:39 AM Timo Walther <twal...@apache.org> wrote: > > > >> Hi Gabor, > >> > >> great that you already try out PTFs. I'm in the process of writing > >> documentation for it. Including a list of limitations. > >> > >> Please note that PTF won't be support in batch mode in the first phase. > >> For stateful PTFs we would need to use a batch state backend and also > >> other code paths around time need to be adjusted. > >> > >> Cheers, > >> Timo > >> > >> > >> On 28.03.25 03:31, Shengkai Fang wrote: > >>> I think it is by design. You can read the FLIP, it says: > >>> > >>> *Time Semantics*: > >>> > >>> - > >>> > >>> PTFs support event-time semantics only. > >>> - > >>> > >>> Processing-time doesn’t go well with batch mode and thus a unified > >> API > >>> should built on event-time. > >>> The proposed onWatermark timers allow for making processing > >> nevertheless > >>> and key-independent. An onWatermark should cover most processing > >> time use > >>> cases. > >>> > >>> > >>> But I think if the PTF doesn't implement the `onTime` method, it means > >> the > >>> function doesn't care about the time. In this case, we can just > >>> convert directly in batch mode. > >>> > >>> Best, > >>> Shengkai > >>> > >>> Gabor Somogyi <gabor.g.somo...@gmail.com> 于2025年3月28日周五 00:25写道： > >>> > >>>> Hi All, > >>>> > >>>> Seems like the process table function scan operation is not supported > in > >>>> batch mode. > >>>> Steps to repro [1] which gives the following exception: > >>>> > >>>> Caused by: org.apache.flink.table.api.TableException: Unsupported > >> function > >>>> for scan:PROCESS_TABLE > >>>> > >>>> Is this something which is planned? > >>>> > >>>> [1] > >>>> > >>>> > >> > https://github.com/gaborgsomogyi/flink/commit/494b297082de718eae16e4e555ed58cefa404676 > >>>> > >>>> BR, > >>>> G > >>>> > >>> > >> > >> > > > >
