Overall, I don't think it's a good idea to add parallel listing for things like tables and namespaces as it just adds complexity for an incredibly narrow (and possibly poorly designed) use case.
I feel we should leave it up to the server to define whether it will provide consistency across paginated listing and avoid bleeding time-travel like concepts (like 'asOf') into the API. I really just don't see what practical value it provides as there are no explicit or consistently held guarantees around these operations. I'd agree with Micah's argument that if the server does provide stronger guarantees, it should manage those via the opaque token and respond with meaningful errors if it cannot satisfy the internal constraints it imposes (like timeouts). It would help to have articulable use cases to really invest in more complexity in this area and I feel like we're drifting a little into the speculative at this point. -Dan On Wed, Dec 20, 2023 at 3:27 PM Micah Kornfield <emkornfi...@gmail.com> wrote: > I agree that this is not quite useful for clients at this moment. But I'm >> thinking that maybe exposing this will help debugging or diagnosing, user >> just need to be aware of this potential expiration. > > > I think if servers provide a meaningful error message on expiration > hopefully, this would be a good first step in debugging. I think saying > tokens should generally support O(Minutes) at least should cover most > use-cases? > > On Tue, Dec 19, 2023 at 9:18 PM Renjie Liu <liurenjie2...@gmail.com> > wrote: > >> If we choose to manage state on the server side, I recommend not >>> revealing the expiration time to the client, at least not for now. We can >>> introduce it when there's a practical need. It wouldn't constitute a >>> breaking change, would it? >> >> >> I agree that this is not quite useful for clients at this moment. But I'm >> thinking that maybe exposing this will help debugging or diagnosing, user >> just need to be aware of this potential expiration. >> >> On Wed, Dec 20, 2023 at 11:09 AM Xuanwo <xua...@apache.org> wrote: >> >>> > For the continuation token, I think one missing part is about the >>> expiration time of this token, since this may affect the state cleaning >>> process of the server. >>> >>> Some storage services use a continuation token as a binary >>> representation of internal states. For example, they serialize a structure >>> into binary and then perform base64 encoding. Services don't need to >>> maintain state, eliminating the need for state cleaning. >>> >>> > Do servers need to expose the expiration time to clients? >>> >>> If we choose to manage state on the server side, I recommend not >>> revealing the expiration time to the client, at least not for now. We can >>> introduce it when there's a practical need. It wouldn't constitute a >>> breaking change, would it? >>> >>> On Wed, Dec 20, 2023, at 10:57, Renjie Liu wrote: >>> >>> For the continuation token, I think one missing part is about the >>> expiration time of this token, since this may affect the state >>> cleaning process of the server. There are several things to discuss: >>> >>> 1. Should we leave it to the server to decide it or allow the client to >>> config in api? >>> >>> Personally I think it would be enough for the server to determine it for >>> now, since I don't see any usage to allow clients to set the expiration >>> time in api. >>> >>> 2. Do servers need to expose the expiration time to clients? >>> >>> Personally I think it would be enough to expose this through the >>> getConfig api to let users know this. For now there is no requirement for >>> per request expiration time. >>> >>> On Wed, Dec 20, 2023 at 2:49 AM Micah Kornfield <emkornfi...@gmail.com> >>> wrote: >>> >>> IMO, parallelization needs to be a first class entity in the end >>> point/service design to allow for flexibility (I scanned through the >>> original proposal for the scan planning and it looked like it was on the >>> right track). Using offsets for parallelization is problematic from both a >>> consistency and scalability perspective if you want to allow for >>> flexibility in implementation. >>> >>> In particular, I think the server needs an APIs like: >>> >>> DoScan - returns a list of partitions (represented by an opaque >>> entity). The list of partitions should support pagination (in an ideal >>> world, it would be streaming). >>> GetTasksForPartition - Returns scan tasks for a partition (should also >>> be paginated/streaming, but this is up for debate). I think it is an >>> important consideration to allow for empty partitions. >>> >>> With this implementation you don't necessarily require separate server >>> side state (objects in GCS should be sufficient), I think as Ryan >>> suggested, one implementation could be to have each partition correspond to >>> a byte-range in a manifest file for returning the tasks. >>> >>> Thanks, >>> Micah >>> >>> >>> >>> On Tue, Dec 19, 2023 at 9:55 AM Walaa Eldin Moustafa < >>> wa.moust...@gmail.com> wrote: >>> >>> Not necessarily. That is more of a general statement. The pagination >>> discussion forked from server side scan planning. >>> >>> On Tue, Dec 19, 2023 at 9:52 AM Ryan Blue <b...@tabular.io> wrote: >>> >>> > With start/limit each client can query for own's chunk without >>> coordination. >>> >>> Okay, I understand now. Would you need to parallelize the client for >>> listing namespaces or tables? That seems odd to me. >>> >>> On Tue, Dec 19, 2023 at 9:48 AM Walaa Eldin Moustafa < >>> wa.moust...@gmail.com> wrote: >>> >>> > You can parallelize with opaque tokens by sending a starting point for >>> the next request. >>> >>> I meant we would have to wait for the server to return this starting >>> point from the past request? With start/limit each client can query for >>> own's chunk without coordination. >>> >>> >>> On Tue, Dec 19, 2023 at 9:44 AM Ryan Blue <b...@tabular.io> wrote: >>> >>> > I think start and offset has the advantage of being parallelizable (as >>> compared to continuation tokens). >>> >>> You can parallelize with opaque tokens by sending a starting point for >>> the next request. >>> >>> > On the other hand, using "asOf" can be complex to implement and may >>> be too powerful for the pagination use case >>> >>> I don't think that we want to add `asOf`. If the service chooses to do >>> this, it would send a continuation token that has the information embedded. >>> >>> On Tue, Dec 19, 2023 at 9:42 AM Walaa Eldin Moustafa < >>> wa.moust...@gmail.com> wrote: >>> >>> Can we assume it is the responsibility of the server to ensure >>> determinism (e.g., by caching the results along with query ID)? I think >>> start and offset has the advantage of being parallelizable (as compared to >>> continuation tokens). On the other hand, using "asOf" can be complex to >>> implement and may be too powerful for the pagination use case (because it >>> allows to query the warehouse as of any point of time, not just now). >>> >>> Thanks, >>> Walaa. >>> >>> On Tue, Dec 19, 2023 at 9:40 AM Ryan Blue <b...@tabular.io> wrote: >>> >>> I think you can solve the atomicity problem with a continuation token >>> and server-side state. In general, I don't think this is a problem we >>> should worry about a lot since pagination commonly has this problem. But >>> since we can build a system that allows you to solve it if you choose to, >>> we should go with that design. >>> >>> On Tue, Dec 19, 2023 at 9:13 AM Micah Kornfield <emkornfi...@gmail.com> >>> wrote: >>> >>> Hi Jack, >>> Some answers inline. >>> >>> >>> In addition to the start index approach, another potential simple way to >>> implement the continuation token is to use the last item name, when the >>> listing is guaranteed to be in lexicographic order. >>> >>> >>> I think this is one viable implementation, but the reason that the token >>> should be opaque is that it allows several different implementations >>> without client side changes. >>> >>> For example, if an element is added before the continuation token, then >>> all future listing calls with the token would always skip that element. >>> >>> >>> IMO, I think this is fine, for some of the REST APIs it is likely >>> important to put constraints on atomicity requirements, for others (e.g. >>> list namespaces) I think it is OK to have looser requirements. >>> >>> If we want to enforce that level of atomicity, we probably want to >>> introduce another time travel query parameter (e.g. asOf=1703003028000) to >>> ensure that we are listing results at a specific point of time of the >>> warehouse, so the complete result list is fixed. >>> >>> >>> Time travel might be useful in some cases but I think it is orthogonal >>> to services wishing to have guarantees around atomicity/consistency of >>> results. If a server wants to ensure that results are atomic/consistent as >>> of the start of the listing, it can embed the necessary timestamp in the >>> token it returns and parse it out when fetching the next result. >>> >>> I think this does raise a more general point around service definition >>> evolution in general. I think there likely need to be metadata endpoints >>> that expose either: >>> 1. A version of the REST API supported. >>> 2. Features the API supports (e.g. which query parameters are honored >>> for a specific endpoint). >>> >>> There are pros and cons to both approaches (apologies if I missed this >>> in the spec or if it has already been discussed). >>> >>> Cheers, >>> Micah >>> >>> >>> >>> >>> >>> >>> On Tue, Dec 19, 2023 at 8:25 AM Jack Ye <yezhao...@gmail.com> wrote: >>> >>> Yes I agree that it is better to not enforce the implementation to favor >>> any direction, and continuation token is probably better than enforcing a >>> numeric start index. >>> >>> In addition to the start index approach, another potential simple way to >>> implement the continuation token is to use the last item name, when the >>> listing is guaranteed to be in lexicographic order. Compared to the start >>> index approach, it does not need to worry about the change of start index >>> when something in the list is added or removed. >>> >>> However, the issue of concurrent modification could still exist even >>> with a continuation token. For example, if an element is added before the >>> continuation token, then all future listing calls with the token would >>> always skip that element. If we want to enforce that level of atomicity, we >>> probably want to introduce another time travel query parameter (e.g. >>> asOf=1703003028000) to ensure that we are listing results at a specific >>> point of time of the warehouse, so the complete result list is fixed. (This >>> is also the missing piece I forgot to mention in the start index approach >>> to ensure it works in distributed settings) >>> >>> -Jack >>> >>> On Tue, Dec 19, 2023, 9:51 AM Micah Kornfield <emkornfi...@gmail.com> >>> wrote: >>> >>> I tried to cover these in more details at: >>> https://docs.google.com/document/d/1bbfoLssY1szCO_Hm3_93ZcN0UAMpf7kjmpwHQngqQJ0/edit >>> >>> On Sun, Dec 17, 2023 at 6:07 PM Renjie Liu <liurenjie2...@gmail.com> >>> wrote: >>> >>> +1 for this approach. I agree that the streaming approach requires that >>> http client and servers have http 2 streaming support, which is not >>> compatible with old clients. >>> >>> I share the same concern with Micah that only start/limit may not be >>> enough in a distributed environment where modification happens during >>> iterations. For compatibility, we need to consider several cases: >>> >>> 1. Old client <-> New Server >>> 2. New client <-> Old server >>> >>> >>> >>> On Sat, Dec 16, 2023 at 6:51 AM Daniel Weeks <dwe...@apache.org> wrote: >>> >>> I agree that we want to include this feature and I raised similar >>> concerns to what Micah already presented in talking with Ryan. >>> >>> For backward compatibility, just adding a start and limit implies a >>> deterministic order, which is not a current requirement of the REST spec. >>> >>> Also, we need to consider whether the start/limit would need to be >>> respected by the server. If existing implementations simply return all the >>> results, will that be sufficient? There are a few edge cases that need to >>> be considered here. >>> >>> For the opaque key approach, I think adding a query param to >>> trigger/continue and introducing a continuation token in >>> the ListNamespacesResponse might allow for more backward compatibility. In >>> that scenario, pagination would only take place for clients who know how to >>> paginate and the ordering would not need to be deterministic. >>> >>> -Dan >>> >>> On Fri, Dec 15, 2023, 10:33 AM Micah Kornfield <emkornfi...@gmail.com> >>> wrote: >>> >>> Just to clarify and add a small suggestion: >>> >>> The behavior with no additional parameters requires the operations to >>> happen as they do today for backwards compatibility (i.e either all >>> responses are returned or a failure occurs). >>> >>> For new parameters, I'd suggest an opaque start token (instead of >>> specific numeric offset) that can be returned by the service and a limit >>> (as proposed above). If a start token is provided without a limit a >>> default limit can be chosen by the server. Servers might return less than >>> limit (i.e. clients are required to check for a next token to determine if >>> iteration is complete). This enables server side state if it is desired >>> but also makes deterministic listing much more feasible (deterministic >>> responses are essentially impossible in the face of changing data if only a >>> start offset is provided). >>> >>> In an ideal world, specifying a limit would result in streaming >>> responses being returned with the last part either containing a token if >>> continuation is necessary. Given conversation on the other thread of >>> streaming, I'd imagine this is quite hard to model in an Open API REST >>> service. >>> >>> Therefore it seems like using pagination with token and offset would be >>> preferred. If skipping someplace in the middle of the namespaces is >>> required then I would suggest modelling those as first class query >>> parameters (e.g. "startAfterNamespace") >>> >>> Cheers, >>> Micah >>> >>> >>> On Fri, Dec 15, 2023 at 10:08 AM Ryan Blue <b...@tabular.io> wrote: >>> >>> +1 for this approach >>> >>> I think it's good to use query params because it can be >>> backward-compatible with the current behavior. If you get more than the >>> limit back, then the service probably doesn't support pagination. And if a >>> client doesn't support pagination they get the same results that they would >>> today. A streaming approach with a continuation link like in the scan API >>> discussion wouldn't work because old clients don't know to make a second >>> request. >>> >>> On Thu, Dec 14, 2023 at 10:07 AM Jack Ye <yezhao...@gmail.com> wrote: >>> >>> Hi everyone, >>> >>> During the conversation of the Scan API for REST spec, we touched on the >>> topic of pagination when REST response is large or takes time to be >>> produced. >>> >>> I just want to discuss this separately, since we also see the issue for >>> ListNamespaces and ListTables/Views, when integrating with a large >>> organization that has over 100k namespaces, and also a lot of tables in >>> some namespaces. >>> >>> Pagination requires either keeping state, or the response to be >>> deterministic such that the client can request a range of the full >>> response. If we want to avoid keeping state, I think we need to allow some >>> query parameters like: >>> - *start*: the start index of the item in the response >>> - *limit*: the number of items to be returned in the response >>> >>> So we can send a request like: >>> >>> *GET /namespaces?start=300&limit=100* >>> >>> *GET /namespaces/ns/tables?start=300&limit=100* >>> >>> And the REST spec should enforce that the response returned for the >>> paginated GET should be deterministic. >>> >>> Any thoughts on this? >>> >>> Best, >>> Jack Ye >>> >>> >>> >>> -- >>> Ryan Blue >>> Tabular >>> >>> >>> >>> -- >>> Ryan Blue >>> Tabular >>> >>> >>> >>> -- >>> Ryan Blue >>> Tabular >>> >>> >>> >>> -- >>> Ryan Blue >>> Tabular >>> >>> >>> Xuanwo >>> >>>
