Hey Shengkai! Thanks for the questions, you hit on some very good practical points. Let me provide my answers below, in the meantime I have already updated the FLIP to include some of your suggestions :)
1. How would schema inference work for RowDataSerializer? That's a good observation, I did not notice this. Probably the simplest solution would be to introduce a new version in the RowDataSerializerSnapshot and include the names for this use-case. This would not really impact checkpointing times/performance but would allow a straightforward mapping for sql states. If we feel that this is too much internal change, then we could also keep it as is for now using simply f0, f1... 2. Is one keyed-state table per operator the right abstraction? This is a very good point and something that has bothered me as well from a design perspective. There is no single good abstraction here because there are completely different use cases. When you just want to look into the state for a single / multiple keys and you mostly have simple value list states, the single table representation is superior from both query syntax and performance perspective. It avoids JOINS and maps to the simple mental model that for a certain key you have state x,y,z. Due to this straightforward mental model I still think this is the good default representation. With projection pushdown, it's easy to select one/several specific columns without reading / touching any other. The big issue is however with large collection states that simply cannot be represented within a single row. This happens very often and is one of the main reasons someone would even use a list state (if they understand how they work internally, but not all users do...). Large map, window, list states won't work in the simple row model and are completely impractical. Based on this, my recommendation would be to keep all keyed states in a single table as per the original proposal (one column per keyed state) but also add an extra table per list / map state with the flattened schema. So if the operator has a value and list state, then there would be 2 tables. One with both states as columns (as per original design) + 1 flattened list state table (key, index, value) or for map states (key, map_key, value). This way we cover both use cases naturally. I am also open to making this configurable on the catalog level. I have added this to the FLIP 3. Could you clarify the assumption of "reading state without user classes"? Turns out from an implementation perspective it's not too bad and pojo/avro state schemas can be inferred quite naturally for most cases. However if the user indeed provides the user jar on the classpath then the whole schema resolution will become even simpler because then we do not need any custom inference. For our own use-cases and in general I would not like to assume that user classes will be easily available or that a catalog will represent mostly a single application. On the contrary the way We intend to use this, is definitely mostly without userjars and to represent multiple applications at the same time. 4. Could StateCatalog expose more fine-grained metadata? I think this is a very good idea. I have updated the FLIP to include an operator level metadata table as well (one for each operator). I would love to include everything that you suggested, I think the practical limit is what kind of information is part of the checkpoint and what isn't . This also ties to some questions Roman had about more detailed metadata. Makes sense Cheers Gyula On Tue, Jul 7, 2026 at 4:59 AM Shengkai Fang <[email protected]> wrote: > Hi Gyula, > > Thanks for the FLIP. I like the direction of making savepoint/checkpoint > state discoverable and queryable from SQL. I have a few questions and > concerns about the proposed abstraction. > > 1. How would schema inference work for RowDataSerializer? > > From the current RowDataSerializer snapshot, it looks like the snapshot > persists the LogicalType[] and nested serializer snapshots, so the field > types can be restored. However, the top-level RowDataSerializer constructed > from a RowType seems to store only the child LogicalTypes, not the RowType > field names. Would StateCatalog expose generated names such as f0/f1, or is > there another source for recovering the original field names? > > 2. Is one keyed-state table per operator the right abstraction? > > I wonder whether one table per named keyed state would be a better base > abstraction, with an optional operator-level wide view on top. In > particular, MapState can contain an unbounded or highly variable number of > entries per state key. Exposing it as a MAP<K,V> column may require fully > deserializing the map for a key into heap. A normalized table shape such > as: > > (state_key, map_key, map_value) > > seems more scalable and SQL-friendly for MapState. Similarly, > ValueState/ListState/MapState have different natural table shapes, so tying > the physical table boundary to the operator may be too coarse. > > 3. Could you clarify the assumption of "reading state without user > classes"? > > This is a very attractive goal, but it also seems to introduce substantial > complexity for POJOs, Avro SpecificRecord, subclasses, and custom > serializers. If StateCatalog is positioned as a job-level/application-level > catalog, would requiring the job jar or user artifacts be acceptable as a > first step? That might simplify the design while still covering many > operational/debugging use cases. > > 4. Could StateCatalog expose more fine-grained metadata? > > For debugging state, it would be useful to expose state-level metadata such > as state name, state type, serializer snapshot/serializer class, TTL > configuration, namespace/window information where applicable, backend state > type, and possibly whether a state can be read lazily/streamingly. > > Best, > Shengkai > > Roman Khachatryan <[email protected]> 于2026年7月7日周二 08:44写道： > > > Hi Gyula, > > > > Thanks for the proposal, this looks very useful! A few questions and > > comments: > > > > 1. Following up on Han's question about checkpoint retention: I > understand > > external coordination is out of scope for now, but could the catalog at > > least detect that a checkpoint was subsumed/deleted mid-query and fail > with > > a clear error, rather than a low-level file-not-found? And do you see > > ownership/pinning as a possible follow-up FLIP once checkpoint reading > > picks up adoption? > > 2. Does the proposal allow querying file-level metadata (file size, > > creation date, etc.)? This would be useful for debugging > compaction-related > > issues. > > 3. If yes, could data and metadata queries have separate access modes? In > > many environments access to data is much stricter than access to > metadata, > > so being able to grant metadata-only access to the catalog would broaden > > where it can be deployed. > > 4. Just to confirm: incremental checkpoints are expected to work through > > the regular restore mechanisms, given sufficient retention? > > 5. +1 on bringing non-keyed state into scope — a concrete use case: > > inspecting Kafka transaction state (currently stored in non-keyed > operator > > state) would be very valuable for debugging EOS issues. > > 6. Could you explain why timers are not supported? They live in keyed > state > > and the state processor API can read registered timers, so I'm wondering > > whether this is a fundamental limitation or just table-mapping scope. > > 7. Does the proposal allow querying checkpoint metadata (such as > > SharingFilesStrategy, isSavepoint, etc.)? This could be useful for > > debugging CLAIM mode issues. > > > > > > Regards, > > Roman > > > > > > On Mon, Jul 6, 2026 at 1:02 PM Gyula Fóra <[email protected]> wrote: > > > > > Hi Zakelly! > > > > > > That's a good point and we have to ensure that it works. In theory SQL > > > related states are relatively easy to cover and represent. The RowData > > > state would be mapped directly to ROW<...> similar to other pojo key > > > states. > > > > > > Cheers > > > Gyula > > > > > > On Sun, Jul 5, 2026 at 4:03 PM Zakelly Lan <[email protected]> > > wrote: > > > > > > > Hi Gyula, > > > > > > > > Thanks for driving this, it's a nice addition and I fully support it. > > One > > > > thing to make sure: > > > > > > > > For the state generated by some Flink SQL jobs, does the StateCatalog > > > infer > > > > this internal `RowData` structure and expose it as a SQL `ROW<...>` > > type? > > > > For example, a regular streaming join side may be stored as a state > > such > > > as > > > > `left-records` / `right-records`, whose value or map key/value > > contains a > > > > `RowData` for the original input row. > > > > > > > > > > > > Best, > > > > Zakelly > > > > > > > > On Fri, Jul 3, 2026 at 4:13 PM Dennis-Mircea Ciupitu < > > > > [email protected]> wrote: > > > > > > > > > Hi Gyula, > > > > > > > > > > Thanks for the detailed answers. This addresses my questions well > and > > > the > > > > > direction sounds great. > > > > > > > > > > +1 (non-binding) from my side. > > > > > > > > > > Best regards, > > > > > Dennis > > > > > > > > > > > > > > > On Thu, Jul 2, 2026 at 3:26 PM Gyula Fóra <[email protected]> > > > wrote: > > > > > > > > > > > Hi Dennis! > > > > > > > > > > > > Thank you for the questions. Much recent work in the state > > connector > > > > api > > > > > > has been done basically towards this type of nice cataloging and > > > > flexible > > > > > > access. There are a few holes and things that have to be changed, > > not > > > > > > everything is enumerated in the FLIP but we have to have an open > > mind > > > > and > > > > > > make all necessary changes as you said to make this truly nice > and > > > > > > comprehensive as much as possible. Most state processor apis are > > > marked > > > > > > experimental so we can be flexible within reason :) > > > > > > > > > > > > Now to the concrete questions: > > > > > > > > > > > > 1. Non-keyed state support / scope > > > > > > I think non-keyed states should definitely be in the scope of the > > > FLIP > > > > in > > > > > > terms of design , and my intention was not to exclude them I just > > > > focused > > > > > > on the keyed state as that is readily available in our prototype > > > > > > implementation (without much changes to the existing > connectors). I > > > > will > > > > > > try to update the FLIP to include non-keyed states more in detail > > > but I > > > > > > think the case is pretty straightforward. From a table > > representation > > > > > > perspective, they can follow a similar pattern such as: > > > > > > uid_opUID_statename_broadcast , uid_opUID_statename_list . A > > > > > corresponding > > > > > > SQL connector can easily be added to support these based on the > > > > existing > > > > > > datastream connector. I will make sure to add separate tickets > for > > > > these > > > > > > types of states once the FLIP is accepted and this work can very > > > easily > > > > > be > > > > > > parallelized across different state types within the existing > > catalog > > > > > > frameworks. This way keyed/non-keyed states will live directly > > > together > > > > > in > > > > > > a single catalog/db. > > > > > > > > > > > > In the future we can even go a step further and include connector > > > > > specific > > > > > > state views such as kafka offsets etc with custom connector > > specific > > > > > > plugins > > > > > > > > > > > > 2/3. Serializer transparency and robustness > > > > > > From a practical standpoint both generated (synthetic) > serializers > > > and > > > > > > custom classes / kryo and pluggable logic could work but the > whole > > > > > catalog > > > > > > concepts requires a certain behaviour to be useful. The catalog > > would > > > > > point > > > > > > to savepoint directories and discover all state in it > (potentially > > > from > > > > > > multiple jobs). Configuration has to be done in a generic way, I > > > don't > > > > > see > > > > > > a problem with introducing configs for specifying custom > > > > > > serializers/factories either generically for certain specific > > > classes. > > > > In > > > > > > most cases however this won't be necessary as the state snapshot > > > itself > > > > > > usually has a reference (classname) of the original user classes. > > If > > > > the > > > > > > catalog process has access to those classes it will use that > > > directly, > > > > or > > > > > > other confugred serializers, and only if not available fall back > to > > > > > > generating serializers for POJO/TUPLE types. There is obviously a > > > limit > > > > > to > > > > > > what is possible here initially, Kryo being one exception where > you > > > > > either > > > > > > have the class or not. > > > > > > > > > > > > I would like to however point out that we do not have to support > > > > > everything > > > > > > initially, we can start with what is currently available, use the > > > > > classpath > > > > > > / generated serializers and as we develop we will find the limits > > of > > > > this > > > > > > approach and then can extend with configuration as it feels > natural > > > > > instead > > > > > > of trying to create a super complex initial solution. But I > > > definitely > > > > > > agree that we should support custom serializer already specified > in > > > the > > > > > > config that is otherwise used by flink for the jobs (but I think > > this > > > > > > should more or less work out of the box). > > > > > > > > > > > > 4. The metadata view is currently reused based on the existing > > table > > > > > valued > > > > > > function. Let's take this as a followup under this umbrella to > > > improve > > > > / > > > > > > extend the metadata view. I don't think we need a separate FLIP > but > > > it > > > > > also > > > > > > feels out of scope here. > > > > > > > > > > > > Cheers > > > > > > Gyula > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > On Thu, Jul 2, 2026 at 1:02 PM Dennis-Mircea Ciupitu < > > > > > > [email protected]> wrote: > > > > > > > > > > > > > Hi all, > > > > > > > > > > > > > > Thank you for driving this. Being able to discover > > > > > savepoints/checkpoints > > > > > > > and query their state as SQL tables without shipping the > original > > > > user > > > > > > > classes is a genuinely valuable addition, and it's nice that it > > > > builds > > > > > on > > > > > > > the existing state-table connector and savepoint_metadata work > > > rather > > > > > > than > > > > > > > starting from scratch. > > > > > > > > > > > > > > A few points and questions, mostly around scope and the > > serializer > > > > > story: > > > > > > > > > > > > > > 1. Non-keyed state and the DataStream path. > > > > > > > - The FLIP scopes out BroadcastState, operator ListState > > and > > > > > > > UnionState because "no readily available Table API > > connectors > > > > > exist > > > > > > > for > > > > > > > these state types." That's a fair characterization of the > > > Table > > > > > > > layer, but > > > > > > > the state-processor DataStream API already reads all > three > > > > today > > > > > > > (SavepointReader#readBroadcastState / #readUnionState / > > > > > > > #readListState). So > > > > > > > the limitation is really in the keyed-only SQL mapping > > > > > > > (KeyedStateReader > > > > > > > runs inside a keyed backend), not in the snapshots > > > themselves. > > > > > > > - Is the keyed-only scope a deliberate UX/table-mapping > > > > decision, > > > > > > or > > > > > > > would a DataStream-backed reader be considered so the > > catalog > > > > > isn't > > > > > > > strictly less capable than the API it extends? Even if > > > > non-keyed > > > > > > > contents > > > > > > > stay out of scope initially, it would be good to frame > this > > > > > > > explicitly as a > > > > > > > Table-mapping constraint rather than a general one. > > > > > > > 2. Serializer transparency - the "no user classes" premise > vs. > > > > > custom > > > > > > > serializers. > > > > > > > - The design relies on Flink's transparent serializer > > formats > > > > to > > > > > > > decode state without user dependencies, which is great > for > > > > > > > POJO/Avro/basic > > > > > > > types. But two serialization efforts point the other way: > > > > > FLIP-398 > > > > > > > [1] > > > > > > > (released) already lets users configure serializers per > > type > > > > via > > > > > > > pipeline.serialization-config, and FLIP-538 [2] (in > > > discussion) > > > > > > adds > > > > > > > pluggable custom generic-type serializers (e.g. Apache > > Fory) > > > > and > > > > > > > promotes > > > > > > > TypeSerializer/TypeSerializerSnapshot to @Public. As > > FLIP-538 > > > > > > > itself notes, > > > > > > > state written with a custom serializer becomes dependent > on > > > > that > > > > > > > serializer > > > > > > > to decode - external tooling without it cannot read those > > > > bytes. > > > > > > > - Could we make the deserialization side pluggable and > > > > > > config-driven, > > > > > > > mirroring FLIP-398's serialization-config, with a > graceful > > > > > fallback > > > > > > > (e.g. > > > > > > > expose the raw bytes / skip the column) when a format > isn't > > > > > > > transparently > > > > > > > decodable? There already seems to be a seam for this > > > > > > > (SavepointTypeInformationFactory), and making it a > > > first-class, > > > > > > > config-selectable option would keep the catalog > > > > > forward-compatible > > > > > > as > > > > > > > serialization support grows. > > > > > > > 3. Robustness of the transparent decoding path. > > > > > > > - Related to (2): reconstructing values by mirroring the > > > binary > > > > > > > layout (PojoToRowDataDeserializer) is the most powerful > but > > > > also > > > > > > the > > > > > > > most > > > > > > > fragile part of the design. How is it expected to behave > > > across > > > > > > > serializer > > > > > > > schema evolution / state migration (a serializer snapshot > > > that > > > > > > > differs from > > > > > > > the writer's), Kryo-fallback fields, nested/generic > types, > > > and > > > > > > > nullability? > > > > > > > - It would help to spell out the supported matrix and the > > > > failure > > > > > > > mode (hard error vs. degrade to raw bytes) up front, > since > > > this > > > > > > > is exactly > > > > > > > where "read without the user classes" is most likely to > > break > > > > in > > > > > > > practice. > > > > > > > 4. Observability / summary reporting. > > > > > > > - The metadata view is a great start. Two small asks: > > > > > > > - per-subtask (or per-key-group) size granularity in > > > > addition > > > > > to > > > > > > > per-operator, since skew is usually what you are > chasing > > > on > > > > > > > large state; > > > > > > > - optionally rounding out the size breakdown with > > > > managed/raw > > > > > > > operator state and channel state sizes for a full > > picture > > > > > > (noting > > > > > > > the > > > > > > > latter are in-flight / unaligned-checkpoint buffers > > rather > > > > > > > than user state). > > > > > > > - A prominent upfront summary of the largest operators / > > > state > > > > is > > > > > > > often what users want before drilling in. > > > > > > > > > > > > > > > > > > > > > Best regards, > > > > > > > Dennis > > > > > > > > > > > > > > [1] > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/spaces/FLINK/pages/282102217/FLIP-398+Improve+Serialization+Configuration+And+Usage+In+Flink > > > > > > > [2] > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/spaces/FLINK/pages/373886828/FLIP-538+Support+Custom+Generic+Type+Serializer > > > > > > > > > > > > > > On Mon, Jun 29, 2026 at 12:53 PM Gyula Fóra <[email protected] > > > > > > wrote: > > > > > > > > > > > > > > > Hi Flink Devs! > > > > > > > > > > > > > > > > I would like to start the discussion about FLIP-599: State > > > Catalog > > > > > [1] > > > > > > > > > > > > > > > > State and stateful processing has always been one of the most > > > > > > fundamental > > > > > > > > features of Flink and a major contributor to its success and > > > global > > > > > > > > adoption. > > > > > > > > > > > > > > > > Over the years several apis and methods have been developed > to > > > > > address > > > > > > > the > > > > > > > > need for external access and analytics such as the state > > > processor > > > > > > > > datastream / java apis, the since deprecated queryable state > > > > > > abstractions > > > > > > > > and more recently a number of table / SQL api connectors to > > > access > > > > > > state > > > > > > > > metadata and keyed states in a somewhat limited way. > > > > > > > > > > > > > > > > Extending the current capabilities of the state-process-api, > > this > > > > > FLIP > > > > > > > aims > > > > > > > > to lift state processing, analytics and observability to a > new > > > > level > > > > > > by > > > > > > > > introducing the State Catalog. > > > > > > > > > > > > > > > > State Catalog is a Flink SQL Catalog implementation that > allows > > > > > > > discovering > > > > > > > > savepoints/checkpoints and mapping their state automatically > to > > > SQL > > > > > > > tables. > > > > > > > > The tables are derived for the different operators and their > > > keyed > > > > > > states > > > > > > > > with schema matching the state structure. Most importantly it > > > > > supports > > > > > > > > reading POJO / Avro and other structured and basic type > states > > > > > without > > > > > > > the > > > > > > > > original user classes (dependencies) by relying on Flink's > > > > > transparent > > > > > > > and > > > > > > > > efficiently structured serializer formats. > > > > > > > > > > > > > > > > We have a fully functional prototype implementation developed > > > with > > > > > > Gabor > > > > > > > > Somogyi that we will be happy to share if the community > accepts > > > the > > > > > > > > proposal! > > > > > > > > > > > > > > > > Looking forward to your feedback and suggestions! > > > > > > > > > > > > > > > > Gyula > > > > > > > > > > > > > > > > [1] > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/spaces/FLINK/pages/438009922/FLIP-599+State+Catalog > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > >
