Hi, Weijie. Thanks for your answer! > No, Introducing and declaring new state > at runtime is something we want to explicitly disallow.
I just thinked about how some operators define their useState() when their real used states may be changed at runtime, e.g. different state types for different state sizes. So these operators only define all states they may use which could be explained by the caller, right ? On Mon, Mar 11, 2024 at 10:57 PM Maximilian Michels <m...@apache.org> wrote: > The FLIP mentions: "The contents described in this FLIP are all new > APIs and do not involve compatibility issues." > > In this thread it looks like the plan is to remove the old state > declaration API. I think we should consider keeping the old APIs to > avoid breaking too many jobs. The new APIs will still be beneficial > for new jobs, e.g. for SQL jobs. > > -Max > > On Fri, Mar 8, 2024 at 4:39 AM Zakelly Lan <zakelly....@gmail.com> wrote: > > > > Hi Weijie, > > > > Thanks for your answer! Well I get your point. Since partitions are > > first-class citizens, and redistribution means how states migrate when > > partitions change, I'd be fine with deemphasizing the concept of > > keyed/operator state if we highlight the definition of partition in the > > document. Keeping `RedistributionMode` under `StateDeclaration` is also > > fine with me, as I guess it is only for internal usage. > > But still, from a user's point of view, state can be characterized along > > two relatively independent dimensions, how states redistribute and the > data > > structure. Thus I still suggest a chained-like configuration API that > > configures one aspect on each call, such as: > > ``` > > # Keyed stream, no redistribution mode specified, the state will go with > > partition (no redistribution). ---- Keyed state > > StateDeclaration a = States.declare(name).listState(type); > > > > # Keyed stream, redistribution strategy specified, the state follows the > > specified redistribute strategy. ---- Operator state > > StateDeclaration b = > > States.declare(name).listState(type).redistributeBy(strategy); > > > > # Non-keyed stream, redistribution strategy *must be* specified. > > StateDeclaration c = > > States.declare(name).listState(type).redistributeBy(strategy); > > > > # Broadcast stream and state > > StateDeclaration d = States.declare(name).mapState(typeK, > > typeV).broadcast(); > > ``` > > It can drive users to think about redistribution issues when needed. And > it > > also provides more flexibility to add more combinations such as > > broadcasting list state, or chain more configurable aspects such as > adding > > `withTtl()` in future. WDYT? > > > > > > Best, > > Zakelly > > > > On Thu, Mar 7, 2024 at 6:04 PM weijie guo <guoweijieres...@gmail.com> > wrote: > > > > > Hi Jinzhong, > > > > > > Thanks for the reply! > > > > > > > Overall, I think that the “Eager State Declaration” is a good > proposal, > > > which can enhance Flink's state management capabilities and provide > > > possibilities for subsequent state optimizations. > > > > > > It's nice to see that people who are familiar with the state stuff like > > > this proposal. :) > > > > > > > When the user attempts to access an undeclared state at runtime, it > is > > > more reasonable to throw an exception rather than returning > Option#empty, > > > as Gyula mentioned above. > > > > > > Yes, I agree that this is better then a confused empty, and I have > modified > > > the corresponding part of this FLIP. > > > > > > > In addition, I'm not quite sure whether all of the existing usage in > > > which states are registered at runtime dynamically can be migrated to > the > > > "Eager State Declaration" style with minimal cost? > > > > > > I think for most user functions, this is fairly straightforward to > migrate. > > > But states whose declarations depend on runtime information(e.g. > > > RuntimeContext) are, in principle, not supported in the new API. > Anyway, > > > the old and new apis are completely incompatible, so rewriting jobs is > > > inevitable. User can think about how to write a good process function > that > > > conforms to the eager declaration style. > > > > > > > For state TTL, should StateDeclaration also provide interfaces for > users > > > to declare state ttl? > > > > > > Of course, We can and we need to provide this one. But whether or not > it's > > > in this FLIP isn't very important for me, because we're mainly talking > > > about the general principles and ways of declaring and accessing state > in > > > this FLIP. I promise we won't leave it out in the end D). > > > > > > > > > > > > Best regards, > > > > > > Weijie > > > > > > > > > Jinzhong Li <lijinzhong2...@gmail.com> 于2024年3月7日周四 17:34写道： > > > > > > > Hi Weijie, > > > > > > > > Thanks for driving this! > > > > > > > > 1. Overall, I think that the “Eager State Declaration” is a good > > > proposal, > > > > which can enhance Flink's state management capabilities and provide > > > > possibilities for subsequent state optimizations. > > > > > > > > 2. When the user attempts to access an undeclared state at runtime, > it is > > > > more reasonable to throw an exception rather than returning > Option#empty, > > > > as Gyula mentioned above. > > > > In addition, I'm not quite sure whether all of the existing usage in > > > which > > > > states are registered at runtime dynamically can be migrated to the > > > "Eager > > > > State Declaration" style with minimal cost? > > > > > > > > 3. For state TTL, should StateDeclaration also provide interfaces for > > > users > > > > to declare state ttl? > > > > > > > > Best, > > > > Jinzhong Li > > > > > > > > > > > > On Thu, Mar 7, 2024 at 5:08 PM weijie guo <guoweijieres...@gmail.com > > > > > > wrote: > > > > > > > > > Hi Hangxiang, > > > > > > > > > > Thanks for your reply! > > > > > > > > > > > We have also discussed in FLIP-359/FLINK-32658 about limiting the > > > user > > > > > operation to avoid creating state when processElement. Could > current > > > > > interfaces also help this? > > > > > > > > > > I think so. It is illegal to create state at runtime in our > proposal. > > > > > > > > > > > > > > > > Could you provide more examples about how useStates() works ? > Since > > > > some > > > > > operations may change their used states at runtime, the value this > > > method > > > > > returns will be modified at runtime, right? > > > > > > > > > > > > > > > No, Introducing and declaring new state > > > > > at runtime is something we want to explicitly disallow. You can > simply > > > > > assume that useState is only called when the JobGraph is generated, > > > > > and any future changes to it are invalid and illegal. > > > > > > > > > > > > > > > > IIUC, RedistributionMode/Strategy should not be used by users, > right > > > ? > > > > > If so, I'm +1 to move them to inner interfaces which seems a bit > > > > confusing > > > > > to users. > > > > > > > > > > > > > > > As for this question, I think my answer to Zakelly should be > helpful. > > > > > > > > > > > > > > > > > > > > Best regards, > > > > > > > > > > Weijie > > > > > > > > > > > > > > > weijie guo <guoweijieres...@gmail.com> 于2024年3月7日周四 16:58写道： > > > > > > > > > > > Hi Zakelly, > > > > > > > > > > > > Thanks for your reply! > > > > > > > > > > > > > My advice would be to conceal RedistributionMode/Strategy from > the > > > > > > standard user interface, particularly within the helper class > > > 'State'. > > > > > But > > > > > > I'm OK to keep it in `StateDeclaration` since its interfaces are > > > > > basically > > > > > > used by the framework. > > > > > > > > > > > > I'm sorry, I didn't mention some of the details/concepts > introduced > > > by > > > > > the Umbrella FLIP and FLIP-409 in this FLIP. This might make it > hard to > > > > > understand the motivation behind > > > > > > RedistributionMode, I'll add more context in FLIP then. > > > > > > > > > > > > > > > > > > > > > > > > Briefly, in V2, we explicitly define the concept of partition. > In the > > > > > case of KeyedPartitionStream, one key corresponds to one > partition.For > > > > > NonKeyedPartitionStream one parallelism/subtask corresponds to one > > > > > partition. All states are considered to be confined within the > > > partition. > > > > > On this basis, an obvious question is whether and how the state > should > > > be > > > > > redistribution when the partition changes? So we divide the state > into > > > > > three categories: > > > > > > > > > > > > - Don't need to redistribute states when the partition > changes. > > > > > > - Has to decide how to distribute states when the partition > > > changes. > > > > > > - Always has the same state across different partitions. > > > > > > > > > > > > After introducing the concept of partition, the redistribution > > > > > pattern/mode of state is the more essential difference between > states. > > > > For > > > > > this reason, we don't want to emphasize keyed/operator state in > the V2 > > > > API > > > > > > any > > > > > > more. Keep in mind, partition are first-class citizens. And, > even in > > > > V1, > > > > > we have to let the user know that split/union are two different > > > > strategies > > > > > for list state. > > > > > > > > > > > > > > > > > > > > > > > > As for whether or not to expose RedistributionMode to users, I > have > > > an > > > > > open mind. But as I said just now, we still can't avoid this > problem in > > > > the > > > > > splitRedistributionListState and unionRedistributionListState. IMO, > > > it's > > > > > better to explain it in the API level instead of avoiding it. WDTY? > > > > > > > > > > > > Best regards, > > > > > > > > > > > > Weijie > > > > > > > > > > > > > > > > > > weijie guo <guoweijieres...@gmail.com> 于2024年3月7日周四 16:39写道： > > > > > > > > > > > >> Hi Gyula, > > > > > >> > > > > > >> > > > > > >> Thanks for your reply! > > > > > >> > > > > > >> > > > > > >> Let me answer these questions: > > > > > >> > > > > > >> > > > > > >> > What is the semantics of the usesStates method? When is it > called? > > > > Can > > > > > >> the used state change dynamically at runtime? Can the logic > depend > > > on > > > > > something computed in open(..) for example? > > > > > >> > > > > > >> > > > > > >> > > > > > >> useStates is used to predefine all the states that the process > > > > function > > > > > needs to access. In other words, we want to avoid declaring the > state > > > > > dynamically at runtime and this allows the SQL planner and JM to > > > optimize > > > > > the job better. As a result, this logic must be fully available at > > > > compile > > > > > time (when the JobGraph is generated), so it can't rely on > computations > > > > > that are executed after deploy to TM. > > > > > >> > > > > > >> > > > > > >> > > > > > > >> Currently state access is pretty dynamic in Flink and I would > assume > > > > > many jobs create states on the fly based on some required logic. > Are we > > > > > planning to address these use-cases? > > > > > >> > > > > > >> > > > > > >> It depends on what type of context we need. If the type and > number > > > of > > > > > states depend on runtime context, that's something we want to > avoid. If > > > > it > > > > > only depended on information available at compile time, I think we > > > could > > > > > support > > > > > >> it. > > > > > >> > > > > > >> > > > > > >> > > > > > > >> Are we planning to support deleting/dropping states that are not > > > > > required anymore? > > > > > >> > > > > > >> > > > > > >> > > > > > >> We really don't want the user to be able to dynamically > > > declare/delete > > > > > a state at runtime, but if you just want to clear/clean the value > of > > > > state, > > > > > the new API works the same as the old API. > > > > > >> > > > > > >> > > > > > >> > I think if a state is not declared or otherwise cannot be > > > accessed, > > > > an > > > > > >> exceptions must be thrown. We cannot confuse empty value with > > > > something > > > > > >> inaccessible. > > > > > >> > > > > > >> > > > > > >> After thinking about it a bit more, I think you have a point! > > > > > >> It's important to make a clear distinction between an empty > state > > > and > > > > > illegal access, especially since flink currently discourage > setting a > > > > > non-null default value for the state. > > > > > >> I will modify the proposal as you suggested then :) > > > > > >> > > > > > >> > > > > > >> > The RedistributionMode enum sounds a bit strange to me, as it > > > > doesn't > > > > > >> actually specify a mode of redistribution. It feels more like a > > > flag. > > > > > Can > > > > > >> we simply have an Optional<RedistributionStrategy> instead? > > > > > >> > > > > > >> > > > > > >> We actually define three types RedistributionMode instead of two > > > > because > > > > > >> we don't want to think of IDENTICAL as a redistribution > strategy, > > > it's > > > > > just > > > > > >> an invariant: the State of that type is always the same across > > > > > partitions. > > > > > >> If it only has None and REDISTRIBUTABLE, I think your proposal > is > > > > > >> feasible then. But we don't want to confuse these three > > > > semantics/modes. > > > > > >> > > > > > >> > > > > > >> > BroadcastStates are currently very limited by only Map-like > > > states, > > > > > and > > > > > >> the new interface also enforces that. Can we remove this > limitation? > > > > If > > > > > >> not, should broadcastState declaration extend mapstate > declaration? > > > > > >> > > > > > >> > > > > > >> > > > > > >> Personally, I don't want to make this restriction. This is also > why > > > > the > > > > > method in StateManager to get BroadcastState has the parameter of > > > > > BroadcastStateDeclaration instead of MapStateDeclaration. In the > > > future, > > > > if > > > > > the state backend supports other types of broadcast state, we can > add a > > > > > corresponding method to the States utility class to get the > > > > > BroadcastSateDeclaration. > > > > > >> > > > > > >> > > > > > >> > > > > > >> Best regards, > > > > > >> > > > > > >> Weijie > > > > > >> > > > > > >> > > > > > >> Hangxiang Yu <master...@gmail.com> 于2024年3月7日周四 11:55写道： > > > > > >> > > > > > >>> Hi, Weijie. > > > > > >>> Thanks for your proposal. > > > > > >>> I'd like to start the discussion with some questions: > > > > > >>> 1. We have also discussed in FLIP-359/FLINK-32658 about > limiting > > > the > > > > > user > > > > > >>> operation to avoid creating state when processElement. Could > > > current > > > > > >>> interfaces also help this? > > > > > >>> > > > > > >>> 2. Could you provide more examples about how useStates() works > ? > > > > Since > > > > > >>> some > > > > > >>> operations may change their used states at runtime, the value > this > > > > > method > > > > > >>> returns will be modified at runtime, right ? > > > > > >>> If so, I'm thinking if we could get some deterministic State > > > > > Declaration > > > > > >>> Set before running which could help a lot for some state > operations > > > > > e.g. > > > > > >>> pre-check schema compatibility, queryable schema. > > > > > >>> > > > > > >>> 3. IIUC, RedistributionMode/Strategy should not be used by > users, > > > > > right ? > > > > > >>> If so, I'm +1 to move them to inner interfaces which seems a > bit > > > > > >>> confusing > > > > > >>> to users. > > > > > >>> > > > > > >>> On Thu, Mar 7, 2024 at 11:39 AM Zakelly Lan < > zakelly....@gmail.com > > > > > > > > > >>> wrote: > > > > > >>> > > > > > >>> > Hi Weijie, > > > > > >>> > > > > > > >>> > Thanks for proposing this! > > > > > >>> > > > > > > >>> > Unifying and optimizing state definitions is a very good > thing. I > > > > > like > > > > > >>> the > > > > > >>> > idea of 'definition goes before using', so overall +1 for > this > > > > > >>> proposal. > > > > > >>> > > > > > > >>> > However, I think the current definition is somewhat unclear. > > > From a > > > > > >>> user's > > > > > >>> > point of view, I believe that state can be characterized > along > > > two > > > > > >>> > relatively independent axes: the scenario (keyed, non-keyed, > or > > > > > >>> broadcast) > > > > > >>> > and the data structure (single value, list, map). I recommend > > > that > > > > we > > > > > >>> fully > > > > > >>> > decouple these aspects, rather than linking the nature of the > > > > > >>> definition to > > > > > >>> > specific assumptions, such as equating broadcast states with > > > maps, > > > > or > > > > > >>> > considering list states could be non-keyed. > > > > > >>> > Furthermore, the concept of 'Redistribution' may impose a > > > cognitive > > > > > >>> burden > > > > > >>> > on general users. My advice would be to conceal > > > > > >>> RedistributionMode/Strategy > > > > > >>> > from the standard user interface, particularly within the > helper > > > > > class > > > > > >>> > 'State'. But I'm OK to keep it in `StateDeclaration` since > its > > > > > >>> interfaces > > > > > >>> > are basically used by the framework. My preferred syntax > would > > > be: > > > > > >>> > ``` > > > > > >>> > StateDeclaration a = > State.declare(name).keyed().listState(type); > > > > > >>> > StateDeclaration b = > > > > State.declare(name).broadcast().mapState(typeK, > > > > > >>> > typeV); > > > > > >>> > StateDeclaration c = > > > > > State.declare(name).keyed().aggregatingState(type, > > > > > >>> > function); > > > > > >>> > ``` > > > > > >>> > WDYT? > > > > > >>> > > > > > > >>> > > > > > > >>> > Best, > > > > > >>> > Zakelly > > > > > >>> > > > > > > >>> > On Wed, Mar 6, 2024 at 11:04 PM Gyula Fóra < > gyula.f...@gmail.com > > > > > > > > > >>> wrote: > > > > > >>> > > > > > > >>> > > Hi Weijie! > > > > > >>> > > > > > > > >>> > > Thank you for the proposal. > > > > > >>> > > > > > > > >>> > > I have some initial questions to start the discussion: > > > > > >>> > > > > > > > >>> > > 1. What is the semantics of the usesStates method? When is > it > > > > > >>> called? Can > > > > > >>> > > the used state change dynamically at runtime? Can the logic > > > > depend > > > > > on > > > > > >>> > > something computed in open(..) for example? > > > > > >>> > > > > > > > >>> > > Currently state access is pretty dynamic in Flink and I > would > > > > > assume > > > > > >>> many > > > > > >>> > > jobs create states on the fly based on some required > logic. Are > > > > we > > > > > >>> > planning > > > > > >>> > > to address these use-cases? > > > > > >>> > > > > > > > >>> > > Are we planning to support deleting/dropping states that > are > > > not > > > > > >>> required > > > > > >>> > > anymore? > > > > > >>> > > > > > > > >>> > > 2. Get state now returns an optional, but you mention that: > > > > > >>> > > " If you want to get a state that is not declared or has no > > > > access, > > > > > >>> > > Option#empty is returned." > > > > > >>> > > > > > > > >>> > > I think if a state is not declared or otherwise cannot be > > > > accessed, > > > > > >>> an > > > > > >>> > > exceptions must be thrown. We cannot confuse empty value > with > > > > > >>> something > > > > > >>> > > inaccessible. > > > > > >>> > > > > > > > >>> > > 3. The RedistributionMode enum sounds a bit strange to me, > as > > > it > > > > > >>> doesn't > > > > > >>> > > actually specify a mode of redistribution. It feels more > like a > > > > > >>> flag. Can > > > > > >>> > > we simply have an Optional<RedistributionStrategy> instead? > > > > > >>> > > > > > > > >>> > > 4. BroadcastStates are currently very limited by only > Map-like > > > > > >>> states, > > > > > >>> > and > > > > > >>> > > the new interface also enforces that. > > > > > >>> > > Can we remove this limitation? If not, should > broadcastState > > > > > >>> declaration > > > > > >>> > > extend mapstate declaration? > > > > > >>> > > > > > > > >>> > > Cheers, > > > > > >>> > > Gyula > > > > > >>> > > > > > > > >>> > > Cheers > > > > > >>> > > Gyuka > > > > > >>> > > > > > > > >>> > > On Wed, Mar 6, 2024 at 11:18 AM weijie guo < > > > > > >>> guoweijieres...@gmail.com> > > > > > >>> > > wrote: > > > > > >>> > > > > > > > >>> > > > Hi devs, > > > > > >>> > > > > > > > > >>> > > > I'd like to start a discussion about FLIP-433: State > Access > > > on > > > > > >>> > > > DataStream API V2 > > > > > >>> > > > [1]. This is the third sub-FLIP of DataStream API V2. > > > > > >>> > > > > > > > > >>> > > > > > > > > >>> > > > After FLIP-410 [2], we can already write a simple > stateless > > > job > > > > > >>> using > > > > > >>> > the > > > > > >>> > > > DataStream V2 API. But as we all know, stateful > computing is > > > > > >>> Flink's > > > > > >>> > > trump > > > > > >>> > > > card. In this FLIP, we will discuss how to declare and > access > > > > > >>> state on > > > > > >>> > > > DataStream API V2 and we manage to avoid some of the > > > > shortcomings > > > > > >>> of V1 > > > > > >>> > > in > > > > > >>> > > > this regard. > > > > > >>> > > > > > > > > >>> > > > You can find more details in this FLIP. Its relationship > with > > > > > other > > > > > >>> > > > sub-FLIPs can be found in the umbrella FLIP > > > > > >>> > > > [3]. Looking forward to hearing from you, thanks! > > > > > >>> > > > > > > > > >>> > > > > > > > > >>> > > > Best regards, > > > > > >>> > > > > > > > > >>> > > > Weijie > > > > > >>> > > > > > > > > >>> > > > > > > > > >>> > > > [1] > > > > > >>> > > > > > > > > >>> > > > > > > > > >>> > > > > > > > >>> > > > > > > >>> > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-433%3A+State+Access+on+DataStream+API+V2 > > > > > >>> > > > > > > > > >>> > > > [2] > > > > > >>> > > > > > > > > >>> > > > > > > > > >>> > > > > > > > >>> > > > > > > >>> > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-410%3A++Config%2C+Context+and+Processing+Timer+Service+of+DataStream+API+V2 > > > > > >>> > > > > > > > > >>> > > > [3] > > > > > >>> > > > > > > > > >>> > > > > > > > > >>> > > > > > > > >>> > > > > > > >>> > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-408%3A+%5BUmbrella%5D+Introduce+DataStream+API+V2 > > > > > >>> > > > > > > > > >>> > > > > > > > >>> > > > > > > >>> > > > > > >>> > > > > > >>> -- > > > > > >>> Best, > > > > > >>> Hangxiang. > > > > > >>> > > > > > >> > > > > > > > > > > > > > -- Best, Hangxiang.
