The only multiple locations table formats I'm currently aware of are Hive (partitions can live wherever) and Iceberg.
I think for Delta, Hudi, LanceDB, Paimon and File based tables they all have to live in the root location. I'm not sure of any other "file" based tables where this would be an issue but I'd love to know if someone else has ideas. I think with the rise in credential vending, splitting things amongst multiple prefixes is becoming less common. I don't oppose doing an array of locations but it may be enough to just leave this as an extension later. (Support location or locations) On Wed, May 7, 2025 at 8:52 PM yun zou <yunzou.colost...@gmail.com> wrote: > Hi Dmitri, > > If it's not "all" is it not strong enough for a spec, IMHO. If some tables > have multiple base locations how is Polaris going to deal with them? > > Sorry, when I say most of them, it was because I haven't tested all of them > (I only tested Delta and CSV before). > However, if Unity Catalog is only taking one location, I think that is a > strong enough proof that > one location is enough today. > > It is also more natural to start with one location, and if there are use > cases that > require support for multiple locations later, we can move on to V2 spec to > support multiple > tables locations. > > We're making a specification for Polaris. I do not think it is sufficient > to say we'll do the same as other (unspecified ATM) catalogs. > If we want to migrate users from other Catalog services to Polaris (through > federation), then Polaris will need to > provide corresponding capabilities. For example, Unity Catalog storage > location is a URI representation, when entity > are federated from Unity Catalog, we will need to be able to handle the URI > location. > If URI representation is a common standard that has been accepted by other > Catalog services like Unity Catalog, Gravitino, > Polaris should be compatible with that, otherwise it might cause problem > for users when they are migrating from one to > another. > > What will Polaris Server do with this location? > For generic tables, Polaris will provide credential vending for this > location in near future, I don't see we will provide > anything else in short or mid term, since we still want to promote > native support for Iceberg. > Or if you have anything special in your mind that you think we should > support? > > If Polaris has to define it in a spec, it will be hard to change in the > future. > Regardless of whether it is explicitly in the spec definition or as a > reserved property key, as long as they are explicitly > documented, they will be hard to change in the future. From that > perspective, those two approaches seem the same to me. > > Table location is critical information that is required by the engine side > to read and write the tables, which should > be explicitly defined to provide better sharing across engines. For > example, the delta table location is passed in the > table properties with a property key either "location" or "path" depends on > how the table is created. Now, if another > engine wants to read the delta table, it will need to understand those > keys, which are controlled by Spark today. If Spark > changes them one day, all sharing will stop working. > > As to whether we want to put it as an explicit field or a reserved key, I > think for a common field among various > table formats, it makes more sense to have it as an explicit field. For > properties that are specific to a particular table format, > it is more proper to just have a reserved key. > > If Polaris takes control of the location, I think we have to be more > careful > and at least try to make it future-proof. > > I don't think Polaris is taking control of the location, the location is > still controlled by the engine and users today like table names. > Polaris is a Catalog service, it records the generic table entity, and > returns the information back to the user on query. > It might be able to do some validation on the location (like check special > character), but it doesn't decide which location > the table will be used. I personally don't think it is a bad idea to let > the Catalog service also take control of generating > the table location, but I think that will require a lot of work. > > Best Regards, > Yun > > On Wed, May 7, 2025 at 5:22 PM Dmitri Bourlatchkov <di...@apache.org> > wrote: > > > No worries about the name. It is a possible alternative spelling :) > > > > On Wed, May 7, 2025 at 8:04 PM yun zou <yunzou.colost...@gmail.com> > wrote: > > > > > Hi Dmitri, > > > > > > Sorry, I accidentally typed your name wrong in the previous reply! > > > Apologize for this! > > > > > > For the S3 issue, I think we will need to deal with those regardless, > > > especially with the federation work going on, we will need to handle > all > > > those entities eventually coming from different Catalogs, and the URI > > > format seems the standard format used by various Catalog services. > > > > > > Best Regards, > > > Yun > > > > > > On Wed, May 7, 2025 at 4:55 PM yun zou <yunzou.colost...@gmail.com> > > wrote: > > > > > > > Hi Dimitri and Eric, > > > > > > > > Thanks a lot for the feedback! > > > > > > > > For the questions: > > > > - Is one value or many? > > > > It will be one value, similar to the location in Iceberg and the > > > > storage_location in unity catalog. > > > > > > > > Regarding to the point about having new data in new locations and > > keeping > > > > old data in old locations, do we support that for Iceberg > > > > today? > > > > For most of the Spark tables, it seems to only have one location. > > Also, I > > > > think it is better to start restricted first, and then extend it to > > > > allow multiple locations when the use case raises. > > > > > > > > Ref: > > > > Iceberg location: > > > > > > > > > > https://github.com/apache/iceberg/blob/main/open-api/rest-catalog-open-api.yaml#L3451 > > > > Storage location in Unity Catalog: > > > > > > > > > > https://github.com/apache/iceberg/blob/main/open-api/rest-catalog-open-api.yaml#L3451 > > > > > > > > - Is it a URI? > > > > Yes, it will be a URI, which seems the standard catalog > implementation. > > > > Regarding to the point about s3 v2 s3a, i assume that is a common > > > > problem that every catalog implementation needs to address, and we > will > > > > stay the same on this part. At least from the load table point of > view, > > > > Spark engine knows how to deal with such cases. > > > > > > > > - Does it point to any particular file? > > > > No, it doesn't point to a particular file. It is the base table > > location. > > > > > > > > - Is it a common prefix of all files within a table? > > > > It is supposed to be the base table location, which theoretically > > should > > > > be the common prefix of all files within a table I believe. > > > > > > > > - What happens when a value does not match these expectations? > > > > Whether it is one value or many is restricted by the spec already. > > > > For URI format, I think we can do a format check, and fail it. > > > > Other than that, we will not do any other special check, and we rely > on > > > > the client to put the correct value, otherwise, the other engine will > > > > not be able to successfully read the table. > > > > > > > > For the location keyword, as Eric has pointed out, we can potentially > > > have > > > > a reserved key for the properties. However, location is a common > > > > enough key among various table formats, which worths a dedicated key > to > > > > help store and load the information in a more straightforward > > > > way. For things that are specific to one or two formats, I think it > > > makes > > > > more sense to use a reserved property key. > > > > > > > > As a reference, in Iceberg, the CreateTable request and TableMetadata > > > does > > > > have an explicit location key in the spec. For write.data.path > > > > and write.metadata.path, they are passed as properties today. > > > > > > > > Best Regards, > > > > Yun > > > > > > > > > > > > On Wed, May 7, 2025 at 3:54 PM Dmitri Bourlatchkov <di...@apache.org > > > > > > wrote: > > > > > > > >> Another point: I'm pretty sure sooner or later users will want to > move > > > >> their data to some other location. As an option users may want to > > write > > > >> new > > > >> files into another location but keep old files in place. > > > >> > > > >> Also: if the location is a URI, how do we deal with s3 vs. s3a for > > > >> example? > > > >> > > > >> In Iceberg it is quite common for different engines to use different > > > >> access > > > >> tools, which often leads to different URI schemes. > > > >> > > > >> Cheers, > > > >> Dmitri. > > > >> > > > >> On Wed, May 7, 2025 at 6:46 PM Eric Maynard < > eric.w.mayn...@gmail.com > > > > > > >> wrote: > > > >> > > > >> > All good questions Dmitri — I’m especially interested in the first > > one > > > >> as > > > >> > from what I understand Iceberg tables can have metadata and data > at > > > two > > > >> > different paths that we need to vend credentials for. > > > >> > > > > >> > For iceberg tables, we just use special properties to track these > > > >> > locations. I wonder if we couldn’t do the same for generic tables. > > > >> > > > > >> > On Wed, May 7, 2025 at 3:42 PM Dmitri Bourlatchkov < > > di...@apache.org> > > > >> > wrote: > > > >> > > > > >> > > Hi Yun, > > > >> > > > > > >> > > Please clarify the meaning of the value of the new location > > > attribute. > > > >> > > > > > >> > > - Is is one value or many? > > > >> > > - Is it a URI? > > > >> > > - Does it point to any particular file? > > > >> > > - Is it a common prefix of all files within a table? > > > >> > > - What happens when a value does not match these expectation? > > > >> > > > > > >> > > Thanks, > > > >> > > Dmitri. > > > >> > > > > > >> > > On 2025/05/07 21:50:19 yun zou wrote: > > > >> > > > Hi folks, > > > >> > > > > > > >> > > > I would like to propose to add an optional `location` field to > > > >> > > > CreateGenricTable Request and LoadGenericTable response. > > > >> > > > > > > >> > > > The `location` is the location for the table, which is common > to > > > >> most > > > >> > > table > > > >> > > > formats including Iceberg, Delta, Hudi, csv, parquet etc. The > > > >> location > > > >> > > > information is critical for loading the table at engine side, > > > >> having a > > > >> > > > dedicated keyword could help improve the robustness for cross > > > engine > > > >> > > > sharing, instead of relying on the properties passed by the > > client > > > >> > side. > > > >> > > > > > > >> > > > Furthermore, this information is also required to provide > > > credential > > > >> > > > vending capabilities later. > > > >> > > > > > > >> > > > Here is the PR for adding the spec: > > > >> > > > https://github.com/apache/polaris/pull/1543 > > > >> > > > > > > >> > > > Looking forward to your reply and feedback! > > > >> > > > > > > >> > > > Best Regards, > > > >> > > > Yun > > > >> > > > > > > >> > > > > > >> > > > > >> > > > > > > > > > >
