Hi Danny, thanks for the proposal. +1 for adding table hints, it is really a necessary feature for flink sql to integrate with a catalog.
For error handling, I think it would be more natural to throw an exception when error table hint provided, because the properties in hint will be merged and used to find the table factory which would cause an exception when error properties provided, right? On the other hand, unlike other hints which just affect the way to execute the query, the property table hint actually affects the result of the query, we should never ignore the given property hints. For the format of property hints, currently, in sql client, we accept properties in format of string only in DDL: 'connector.type'='kafka', I think the format of properties in hint should be the same as the format we defined in ddl. What do you think? Bests, Wenlong Lyu On Tue, 10 Mar 2020 at 14:22, Danny Chan <yuzhao....@gmail.com> wrote: > To Weike: About the Error Handing > > To be consistent with other SQL vendors, the default is to log warnings > and if there is any error (invalid hint name or options), the hint is just > ignored. I have already addressed in the wiki. > > To Timo: About the PROPERTIES Table Hint > > • The properties hints is also optional, user can pass in an option to > override the table properties but this does not mean it is required. > • They should not include semantics: does the properties belong to > semantic ? I don't think so, the plan does not change right ? The result > set may be affected, but there are already some hints do so, for example, > MS-SQL MAXRECURSION and SNAPSHOT hint [1] > • `SELECT * FROM t(k=v, k=v)`: this grammar breaks the SQL standard > compared to the hints way(which is included in comments) > • I actually didn't found any vendors to support such grammar, and there > is no way to override table level properties dynamically. For normal RDBMS, > I think there are no requests for such dynamic parameters because all the > table have the same storage and computation and they are almost all batch > tables. > • While Flink as a computation engine has many connectors, especially for > some message queue like Kafka, we would have a start_offset which is > different each time we start the query, such parameters can not be > persisted to catalog, because it’s not static, this is actually the > background we propose the table hints to indicate such properties > dynamically. > > > To Jark and Jinsong: I have removed the query hints part and change the > title. > > [1] > https://docs.microsoft.com/en-us/sql/t-sql/queries/hints-transact-sql-query?view=sql-server-ver15 > > Best, > Danny Chan > 在 2020年3月9日 +0800 PM5:46，Timo Walther <twal...@apache.org>，写道： > > Hi Danny, > > > > thanks for the proposal. I agree with Jark and Jingsong. Planner hints > > and table hints are orthogonal topics that should be discussed > separately. > > > > I share Jingsong's opinion that we should not use planner hints for > > passing connector properties. Planner hints should be optional at any > > time. They should not include semantics but only affect execution time. > > Connector properties are an important part of the query itself. > > > > Have you thought about options such as `SELECT * FROM t(k=v, k=v)`? How > > are other vendors deal with this problem? > > > > Regards, > > Timo > > > > > > On 09.03.20 10:37, Jingsong Li wrote: > > > Hi Danny, +1 for table hints, thanks for driving. > > > > > > I took a look to FLIP, most of content are talking about query hints. > It is > > > hard to discussion and voting. So +1 to split it as Jark said. > > > > > > Another thing is configuration that suitable to config with table > hints: > > > "connector.path" and "connector.topic", Are they really suitable for > table > > > hints? Looks weird to me. Because I think these properties are the > core of > > > table. > > > > > > Best, > > > Jingsong Lee > > > > > > On Mon, Mar 9, 2020 at 5:30 PM Jark Wu <imj...@gmail.com> wrote: > > > > > > > Thanks Danny for starting the discussion. > > > > +1 for this feature. > > > > > > > > If we just focus on the table hints not the query hints in this > release, > > > > could you split the FLIP into two FLIPs? > > > > Because it's hard to vote on partial part of a FLIP. You can keep > the table > > > > hints proposal in FLIP-113 and move query hints into another FLIP. > > > > So that we can focuse on the table hints in the FLIP. > > > > > > > > Thanks, > > > > Jark > > > > > > > > > > > > > > > > On Mon, 9 Mar 2020 at 17:14, DONG, Weike <kyled...@connect.hku.hk> > wrote: > > > > > > > > > Hi Danny, > > > > > > > > > > This is a nice feature, +1. > > > > > > > > > > One thing I am interested in but not mentioned in the proposal is > the > > > > error > > > > > handling, as it is quite common for users to write inappropriate > hints in > > > > > SQL code, if illegal or "bad" hints are given, would the system > simply > > > > > ignore them or throw exceptions? > > > > > > > > > > Thanks : ) > > > > > > > > > > Best, > > > > > Weike > > > > > > > > > > On Mon, Mar 9, 2020 at 5:02 PM Danny Chan <yuzhao....@gmail.com> > wrote: > > > > > > > > > > > Note: > > > > > > we only plan to support table hints in Flink release 1.11, so > please > > > > > focus > > > > > > mainly on the table hints part and just ignore the planner > hints, sorry > > > > > for > > > > > > that mistake ~ > > > > > > > > > > > > Best, > > > > > > Danny Chan > > > > > > 在 2020年3月9日 +0800 PM4:36，Danny Chan <yuzhao....@gmail.com>，写道： > > > > > > > Hi, fellows ~ > > > > > > > > > > > > > > I would like to propose the supports for SQL hints for our > Flink SQL. > > > > > > > > > > > > > > We would support hints syntax as following: > > > > > > > > > > > > > > select /*+ NO_HASH_JOIN, RESOURCE(mem='128mb', > parallelism='24') */ > > > > > > > from > > > > > > > emp /*+ INDEX(idx1, idx2) */ > > > > > > > join > > > > > > > dept /*+ PROPERTIES(k1='v1', k2='v2') */ > > > > > > > on > > > > > > > emp.deptno = dept.deptno > > > > > > > > > > > > > > Basically we would support both query hints(after the SELECT > keyword) > > > > > > and table hints(after the referenced table name), for 1.11, we > plan to > > > > > only > > > > > > support table hints with a hint probably named PROPERTIES: > > > > > > > > > > > > > > table_name /*+ PROPERTIES(k1='v1', k2='v2') *+/ > > > > > > > > > > > > > > I am looking forward to your comments. > > > > > > > > > > > > > > You can access the FLIP here: > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-113%3A+SQL+and+Planner+Hints > > > > > > > > > > > > > > Best, > > > > > > > Danny Chan > > > > > > > > > > > > > > > > > > > > > > > >
