Thanks Jane for driving this FLIP. The FLIP is quite interesting. Since the execution plan has finer granularity than the plain SQL script, Hints at SQL level might not be able to touch specific operators, which turns out that the idea of leveraging the compiled execution plan is brilliant.
However, there are some concerns that might need to be considered. - One thing I didn't fully understand. I might be wrong. Could those ttl configs be survived when SQL jobs are restarted? Does that mean that, once I modified the compiled sql plan, the json file will become the sql job? I have to always call the EXECUTE PLAN every time when the job needs to be restarted? In case that the original SQL script has been changed, we need to compile a version2 sql plan and copy the ttl configs from version1 sql plan to version2 and drop version1. This means we have to keep the compiled json file and create a link with the original SQL script. I am not sure if I understood it correctly, it seems like a lot of maintenance effort. - If I am not mistaken, the compiled sql plan introduced by FLIP-190 is only used for SQL job migration/update. Common stages that Flink uses to produce the execution plan from SQL does not contain the compiling step. This makes one tool do two different jobs[1], upgrade + ttl tuning. and tighten the dependency on compiling sql plans. Flink SQL users have to deal with a compiled sql plan for performance optimization that is not designed for it. - The regular working process for Flink SQL users is changed, from only dealing with SQL like scripts to moving between SQL like scripts and file modifications back and forth. This is a big change for user behaviours. One option could be that we upgrade/extend the COMPILE PLAN to allow users update ttl for operators at the script level. But I am not sure if it is possible to point out specific operators at this level. Another option is to print out the result of COMPILE PLAN and enable EXECUTE PLAN 'json plan as string'. Third option is to leverage a data platform to virtualize the compiled sql plan and provide related interactions for updating ttl and submit(execute) the modified compiled sql plan. On the other side, there is one additional benefit with this proposal: we could fine tune SQL jobs while we migrate/upgrade them. That is nice! Best regards, Jing [1] https://en.wikipedia.org/wiki/Single-responsibility_principle On Fri, Mar 24, 2023 at 4:02 PM Leonard Xu <xbjt...@gmail.com> wrote: > Thanks Jane for the proposal. > > TTL of state is an execution phase configuration, serialized json graph > file is the graph for execution phase, supporting the operator level state > TTL in the execution json file makes sense to me. > > From the user's perspective, I have two concerns: > 1. By modifying the execution graph node configuration, this raises the > cost for users to understand, especially for SQL users. > 2. Submitting a SQL job through `exec plan json file` is not so intuitive > as users cannot see the SQL detail of the job > > Best, > Leonard > > On Fri, Mar 24, 2023 at 5:07 PM Shengkai Fang <fskm...@gmail.com> wrote: > > > Hi, Jane. > > > > Thanks for driving this FLIP and this feature are very useful to many > > users. But I have two problems about the FLIP: > > > > 1. How the Gateway users use this feature? As far as I know, the EXEUCTE > > PLAN only supports local file right now. Is it possible to extend this > > syntax to allow for reading plan files from remote file systems? > > > > 2. I would like to inquire if there are any limitations on this feature? > I > > have encountered several instances where the data did not expire in the > > upstream operator, but it expired in the downstream operator, resulting > in > > abnormal calculation results or direct exceptions thrown by the operator > > (e.g. rank operator). Can we limit that the expiration time of downstream > > operator data should be greater than or equal to the expiration time of > > upstream operator data? > > > > Best, > > Shengkai > > > > Yun Tang <myas...@live.com> 于2023年3月24日周五 14:50写道： > > > > > Hi, > > > > > > From my point of view, I am a bit against using SQL hint to set state > TTL > > > as FlinkSQL could be translated to several stateful operators. If we > want > > > to let different state could have different TTL configs within one > > > operator, the SQL hint solution could not work. A better way is to > allow > > a > > > graphical IDE to display the stateful operators and let users configure > > > them. And the IDE submits the json plan to Flink to run jobs. > > > > > > For the details of the structure of ExecNodes, since the state name is > > > unique in the underlying state layer, shall we introduce the "index" > tag > > to > > > identify the state config? > > > What will happen with the conditions below: > > > 1st run: > > > { > > > "index": 0, > > > "ttl": "259200000 ms", > > > "name": "join-lef-state" > > > }, > > > { > > > "index": 1, > > > "ttl": "86400000 ms", > > > "name": "join-right-state" > > > } > > > > > > 2nd run: > > > { > > > "index": 0, > > > "ttl": "86400000 ms", > > > "name": "join-right-state" > > > }, > > > { > > > "index": 1, > > > "ttl": "259200000 ms", > > > "name": "join-lef-state" > > > } > > > > > > Best > > > Yun Tang > > > ________________________________ > > > From: Jane Chan <qingyue....@gmail.com> > > > Sent: Friday, March 24, 2023 11:57 > > > To: dev@flink.apache.org <dev@flink.apache.org> > > > Subject: Re: [DISCUSS] FLIP-292: Support configuring state TTL at > > operator > > > level for Table API & SQL programs > > > > > > Hi Shammon and Shuo, > > > > > > Thanks for your valuable comments! > > > > > > Some thoughts: > > > > > > @Shuo > > > > I think it's more properly to say that hint does not affect the > > > equivalenceof execution plans (hash agg vs sort agg), not the > equivalence > > > of execution > > > results, e.g., users can set 'scan.startup.mode' for kafka connector by > > > dynamic table option, which > > > also "intervene in the calculation of data results". > > > > > > IMO, the statement that "hint should not interfere with the calculation > > > results", means it should not interfere with internal computation. On > the > > > other hand, 'scan.startup.mode' interferes with the ingestion of the > > data. > > > I think these two concepts are different, but of course, this is just > my > > > opinion and welcome other views. > > > > > > > I think the final shape of state ttl configuring may like the that, > > > userscan define operator state ttl using SQL HINT (assumption...), but > it > > > may > > > affects more than one stateful operators inside the same query block, > > then > > > users can further configure a specific one by modifying the compiled > json > > > plan... > > > > > > Setting aside the issue of semantics, setting TTL from a higher level > > seems > > > to be attractive. This means that users only need to configure > > > 'table.exec.state.ttl' through the existing hint syntax to achieve the > > > effect. Everything is a familiar formula. But is it really the case? > > Hints > > > apply to a very broad range. Let me give an example. > > > > > > Suppose a user wants to set different TTLs for the two streams in a > > stream > > > join query. Where should the hints be written? > > > > > > -- the original query before configuring state TTL > > > create temporary view view1 as select .... from my_table_1; > > > create temporary view view2 as select .... from my_table_2; > > > create temporary view joined_view as > > > select view1.*, view2.* from my_view_1 a join my_view_2 b on > a.join_key = > > > b.join_key; > > > > > > Option 1: declaring hints at the very beginning of the table scan > > > > > > -- should he or she write hints when declaring the first temporary > view? > > > create temporary view view1 as select .... from my_table_1 > > > /*+(OPTIONS('table.exec.state.ttl' > > > = 'foo'))*/; > > > create temporary view view2 as select .... from my_table_2 > > > /*+(OPTIONS('table.exec.state.ttl' > > > = 'bar'))*/; > > > create temporary view joined_view as > > > select view1.*, view2.* from my_view_1 a join my_view_2 b on > a.join_key = > > > b.join_key; > > > > > > Option 2: declaring hints when performing the join > > > > > > -- or should he or she write hints when declaring the join temporary > > view? > > > create temporary view view1 as select .... from my_table_1; > > > create temporary view view2 as select .... from my_table_2; > > > create temporary view joined_view as > > > select view1.*, view2.* from my_view_1 > > /*+(OPTIONS('table.exec.state.ttl' = > > > 'foo'))*/ a join my_view_2 /*+(OPTIONS('table.exec.state.ttl' = > > 'bar'))*/ b > > > on a.join_key = b.join_key; > > > > > > From the user's point of view, does he or she needs to care about the > > > difference between these two kinds of style? Users might think the two > > may > > > be equivalent; but in reality, as developers, how do we define the > range > > in > > > which hint starts and ends to take effect? > > > > > > Consider the following two assumptions > > > > > > 1. Assuming the hint takes effect from the moment it is declared and > > > applies to any subsequent stateful operators until it is overridden by > a > > > new hint. > > > If this is the assumption, it's clear that Option 1 and Option 2 are > > > different because a ChangelogNormalize node can appear between scan and > > > join. Meanwhile, which stream's TTL to apply to the following query > after > > > the stream join? It is unclear if the user does not explicitly set it. > > > Should the engine make a random decision? > > > > > > 2. Assuming that the scope of the hint only applies to the current > query > > > block and does not extend to the next operator. > > > In this case, the first way of setting the hint will not work because > it > > > cannot be brought to the join operator. Users must choose the second > way > > to > > > configure. Are users willing to remember this strange constraint on SQL > > > writing style? Does this indicate a new learning cost? > > > > > > The example above is used to illustrate that while this approach may > seem > > > simple and direct, it actually has many limitations and may produce > > > unexpected behavior. Will users still find it attractive? IMO *hints > only > > > work for a very limited situation where the query is very simple, and > its > > > scope is more coarse and not operator-level*. Maybe it deserves another > > > FLIP to discuss whether we need a multiple-level state TTL > configuration > > > mechanism and how to properly implement it. > > > > > > @Shammon > > > > Generally, Flink jobs support two types > > > of submission: SQL and jar. If users want to use `TTL on Operator` for > > SQL > > > jobs, they need to edit the json file which is not supported by general > > job > > > submission systems such as flink sql-client, apache kyuubi, apache > > > streampark and .etc. Users need to download the file and edit it > > manually, > > > but they may not have the permissions to the storage system such as > HDFS > > in > > > a real production environment. From this perspective, I think it is > > > necessary to provide a way similar to > > > hits that users can configure the `TTL on Operator` in their sqls which > > > help users to use it conveniently. > > > > > > IIUC, SQL client supports the statement "EXECUTE PLAN > > > 'file:/foo/bar/example.json'". While I think there is not much evidence > > to > > > say we should choose to use hints, just because users cannot touch > their > > > development environment. As a reply to @Shuo, the TTL set through hint > > way > > > is not at the operator level. And whether it is really "convenient" > needs > > > more discussion. > > > > > > > I agree with @Shuo's idea that for complex cases, users can combine > > hits > > > and `json plan` to configure `TTL on Operator` better. > > > > > > Suppose users can configure TTL through > > > <1> SET 'table.exec.state.ttl' = 'foo'; > > > <2> Modify the compiled JSON plan; > > > <3> Use hints (personally I'm strongly against this way, but let's take > > it > > > into consideration). > > > IMO if the user can configure the same parameter in so many ways, then > > the > > > complex case only makes things worse. Who has higher priority and who > > > overrides who? > > > > > > Best, > > > Jane > > > > > > > > > On Fri, Mar 24, 2023 at 11:00 AM Shammon FY <zjur...@gmail.com> wrote: > > > > > > > Hi jane > > > > > > > > Thanks for initializing this discussion. Configure TTL per operator > can > > > > help users manage state more effectively. > > > > > > > > I think the `compiled json plan` proposal may need to consider the > > impact > > > > on the user's submission workflow. Generally, Flink jobs support two > > > types > > > > of submission: SQL and jar. If users want to use `TTL on Operator` > for > > > SQL > > > > jobs, they need to edit the json file which is not supported by > general > > > job > > > > submission systems such as flink sql-client, apache kyuubi, apache > > > > streampark and .etc. Users need to download the file and edit it > > > manually, > > > > but they may not have the permissions to the storage system such as > > HDFS > > > in > > > > a real production environment. > > > > > > > > From this perspective, I think it is necessary to provide a way > similar > > > to > > > > hits that users can configure the `TTL on Operator` in their sqls > which > > > > help users to use it conveniently. At the same time, I agree with > > @Shuo's > > > > idea that for complex cases, users can combine hits and `json plan` > to > > > > configure `TTL on Operator` better. What do you think? Thanks > > > > > > > > > > > > Best, > > > > Shammon FY > > > > > > > > > > > > On Thu, Mar 23, 2023 at 9:58 PM Shuo Cheng <njucs...@gmail.com> > wrote: > > > > > > > > > Correction: “users can set 'scan.startup.mode' for kafka connector” > > -> > > > > > “users > > > > > can set 'scan.startup.mode' for kafka connector by dynamic table > > > option” > > > > > > > > > > Shuo Cheng <njucs...@gmail.com>于2023年3月23日 周四21:50写道： > > > > > > > > > > > Hi Jane, > > > > > > Thanks for driving this, operator level state ttl is absolutely a > > > > desired > > > > > > feature. I would share my opinion as following: > > > > > > > > > > > > If the scope of this proposal is limited as an enhancement for > > > compiled > > > > > > json plan, it makes sense. I think it does not conflict with > > > > configuring > > > > > > state ttl > > > > > > in other ways, e.g., SQL HINT or something else, because they > just > > > work > > > > > in > > > > > > different level, SQL Hint works in the exact entrance of SQL API, > > > while > > > > > > compiled json plan is the intermediate results for SQL. > > > > > > I think the final shape of state ttl configuring may like the > that, > > > > users > > > > > > can define operator state ttl using SQL HINT (assumption...), but > > it > > > > may > > > > > > affects more than one stateful operators inside the same query > > block, > > > > > then > > > > > > users can further configure a specific one by modifying the > > compiled > > > > json > > > > > > plan... > > > > > > > > > > > > In a word, this proposal is in good shape as an enhancement for > > > > compiled > > > > > > json plan, and it's orthogonal with other ways like SQL Hint > which > > > > works > > > > > in > > > > > > a higher level. > > > > > > > > > > > > > > > > > > Nips: > > > > > > > > > > > > > "From the SQL semantic perspective, hints cannot intervene in > the > > > > > > calculation of data results." > > > > > > I think it's more properly to say that hint does not affect the > > > > > > equivalence of execution plans (hash agg vs sort agg), not the > > > > > equivalence > > > > > > of execution results, e.g., users can set 'scan.startup.mode' for > > > kafka > > > > > > connector, which also "intervene in the calculation of data > > results". > > > > > > > > > > > > Sincerely, > > > > > > Shuo > > > > > > > > > > > > On Tue, Mar 21, 2023 at 7:52 PM Jane Chan <qingyue....@gmail.com > > > > > > wrote: > > > > > > > > > > > >> Hi devs, > > > > > >> > > > > > >> I'd like to start a discussion on FLIP-292: Support configuring > > > state > > > > > TTL > > > > > >> at operator level for Table API & SQL programs [1]. > > > > > >> > > > > > >> Currently, we only support job-level state TTL configuration via > > > > > >> 'table.exec.state.ttl'. However, users may expect a fine-grained > > > state > > > > > TTL > > > > > >> control to optimize state usage. Hence we propose to > > > > > serialize/deserialize > > > > > >> the state TTL as metadata of the operator's state to/from the > > > compiled > > > > > >> JSON > > > > > >> plan, to achieve the goal that specifying different state TTL > when > > > > > >> transforming the exec node to stateful operators. > > > > > >> > > > > > >> Look forward to your opinions! > > > > > >> > > > > > >> [1] > > > > > >> > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=240883951 > > > > > >> > > > > > >> Best Regards, > > > > > >> Jane Chan > > > > > >> > > > > > > > > > > > > > > > > > > > > >
