Hi, thanks for the inputs. I wrote a google doc “Plan for decoupling Hive connector with Flink planner” [1] that shows how to decouple Hive connector with planner.
[1] https://docs.google.com/document/d/1LMQ_mWfB_mkYkEBCUa2DgCO2YdtiZV7YRs2mpXyjdP4/edit?usp=sharing Best, Yuxia.------------------------------------------------------------------ 发件人：Jark Wu<imj...@gmail.com> 日 期：2022年03月10日 19:59:30 收件人：Francesco Guardiani<france...@ververica.com>; dev<dev@flink.apache.org> 抄 送：Martijn Visser<martijnvis...@apache.org>; User<u...@flink.apache.org>; 罗宇侠(莫辞)<luoyuxia.luoyu...@alibaba-inc.com> 主 题：Re: Fwd: [DISCUSS] Flink's supported APIs and Hive query syntax Hi Francesco, Yes. The Hive syntax is a syntax plugin provided by Hive connector. > But right now I don't think It's a good idea adding new features on top, as it will create only more maintenance burden both for Hive developers and for table developers. We are not adding new Hive features, but fixing compatibility or behavior bugs, and almost all of them are just related to the Hive connector code, nothing to do with table planner. I agree we should investigate how to and how much work to decouple Hive connector and planner ASAP. We will come up with a google doc soon. But AFAIK, this may not be a huge work and not conflict with the bugfix works. Best, Jark On Thu, 10 Mar 2022 at 17:03, Francesco Guardiani <france...@ververica.com> wrote: > We still need some work to make the Hive dialect purely rely on public APIs, > and the Hive connector should be decopule with table planner. From the table perspective, I think this is the big pain point at the moment. First of all, when we talk about the Hive syntax, we're really talking about the Hive connector, as my understanding is that without the Hive connector in the classpath you can't use the Hive syntax [1]. The Hive connector is heavily relying on internals [2], and this is an important struggle for the table project, as sometimes is impedes and slows down development of new features and creates a huge maintenance burden for table developers [3]. The planner itself has some classes specific to Hive [4], making the codebase of the planner more complex than it already is. Some of these are just legacy, others exists because there are some abstractions missing in the table planner side, but those just need some work. So I agree with Jark, when the two Hive modules (connector-hive and sql-parser-hive) reach a point where they don't depend at all on flink-table-planner, like every other connector (except for testing of course), we should be good to move them in a separate repo and continue committing to them. But right now I don't think It's a good idea adding new features on top, as it will create only more maintenance burden both for Hive developers and for table developers. My concern with this plan is: how much realistic is to fix all the planner internal leaks in the existing Hive connector/parser? To me this seems like a huge task, including a non trivial amount of work to stabilize and design new entry points in Table API. [1] HiveParser [2] HiveParserCalcitePlanner [3] Just talking about code coupling, not even mentioning problems like dependencies and security updates [4] HiveAggSqlFunction On Thu, Mar 10, 2022 at 9:05 AM Martijn Visser <martijnvis...@apache.org> wrote: Thank you Yuxia for volunteering, that's really much appreciated. It would be great if you can create an umbrella ticket for that. It would be great to get some insights from currently Flink and Hive users which versions are being used. @Jark I would indeed deprecate the old Hive versions in Flink 1.15 and then drop them in Flink 1.16. That would also remove some tech debt and make it less work with regards to externalizing connectors. Best regards, Martijn On Thu, 10 Mar 2022 at 07:39, Jark Wu <imj...@gmail.com> wrote: Thanks Martijn for the reply and summary. I totally agree with your plan and thank Yuxia for volunteering the Hive tech debt issue. I think we can create an umbrella issue for this and target version 1.16. We can discuss details and create subtasks there. Regarding dropping old Hive versions, I'm also fine with that. But I would like to investigate some Hive users first to see whether it's acceptable at this point. My first thought was we can deprecate the old Hive versions in 1.15, and we can discuss dropping it in 1.16 or 1.17. Best, Jark On Thu, 10 Mar 2022 at 14:19, 罗宇侠(莫辞) <luoyuxia.luoyu...@alibaba-inc.com> wrote: Thanks Martijn for your insights. About the tech debt/maintenance with regards to Hive query syntax, I would like to chip-in and expect it can be resolved for Flink 1.16. Best regards, Yuxia ------------------原始邮件 ------------------ 发件人:Martijn Visser <martijnvis...@apache.org> 发送时间:Thu Mar 10 04:03:34 2022 收件人:User <u...@flink.apache.org> 主题:Fwd: [DISCUSS] Flink's supported APIs and Hive query syntax (Forwarding this also to the User mailing list as I made a typo when replying to this email thread) ---------- Forwarded message --------- From: Martijn Visser <martijnvis...@apache.org> Date: Wed, 9 Mar 2022 at 20:57 Subject: Re: [DISCUSS] Flink's supported APIs and Hive query syntax To: dev <dev@flink.apache.org>, Francesco Guardiani <france...@ververica.com>, Timo Walther <twal...@apache.org>, <us...@flink.apache.org> Hi everyone, Thank you all very much for your input. From my perspective, I consider batch as a special case of streaming. So with Flink SQL, we can support both batch and streaming use cases and I think we should use Flink SQL as our target. To reply on some of the comments: @Jing on your remark: > Since Flink has a clear vision of unified batch and stream processing, > supporting batch jobs will be one of the critical core features to help us > reach the vision and let Flink have an even bigger impact in the industry. I fully agree with that statement. I do think that having Hive syntax support doesn't help in that unified batch and stream processing. We're making it easier for batch users to run their Hive batch jobs on Flink, but that doesn't fit the "unified" part since it's focussed on batch, while Flink SQL focusses on batch and streaming. I would have rather invested time in making batch improvements to Flink and Flink SQL vs investing in Hive syntax support. I do understand from the given replies that Hive syntax support is valuable for those that are already running batch processing and would like to run these queries on Flink. I do think that's limited to mostly Chinese companies at the moment. @Jark I think you've provided great input and are spot on with: > Regarding the maintenance concern you raised, I think that's a good point and > they are in the plan. The Hive dialect has already been a plugin and option > now, and the implementation is located in hive-connector module. We still > need some work to make the Hive dialect purely rely on public APIs, and the > Hive connector should be decopule with table planner. At that time, we can > move the whole Hive connector into a separate repository (I guess this is > also in the externalize connectors plan). I'm looping in Francesco and Timo who can elaborate more in depth on the current maintenance issues. I think we need to have a proper plan on how this tech debt/maintenance can be addressed and to get commitment that this will be resolved in Flink 1.16, since we indeed need to move out all previously agreed connectors before Flink 1.16 is released. > From my perspective, Hive is still widely used and there exists many running > Hive SQL jobs, so why not to provide users a better experience to help them > migrate Hive jobs to Flink? Also, it doesn't conflict with Flink SQL as it's > just a dialect option. I do think there is a conflict with Flink SQL; you can't use both of them at the same time, so you don't have access to all features in Flink. That increases feature sparsity and user friction. It also puts a bigger burden on the Flink community, because having both options available means more maintenance work. For example, an upgrade of Calcite is more impactful. The Flink codebase is already rather large and CI build times are already too long. More code means more risk of bugs. If a user at some point wants to change his Hive batch job to a streaming Flink SQL job, there's still migration work for the user, it just needs to happen at a later stage. @Jingsong I think you have a good argument that migrating SQL for Batch ETL is indeed an expensive effort. Last but not least, there was no one who has yet commented on the supported Hive versions and security issues. I've reached out to the Hive community and from the info I've received so far is that only Hive 3.1.x and Hive 2.3.x are still supported. The older Hive versions are no longer maintained and also don't receive security updates. This is important because many companies scan the Flink project for vulnerabilities and won't allow using it when these types of vulnerabilities are included. My summary would be the following: * Like Jark said, in the short term, Hive syntax compatibility is the ticket for us to have a seat in the batch processing. Having improved Hive syntax support for that in Flink can help in this. * In the long term, we can and should drop it and focus on Flink SQL itself both for batch and stream processing. * The Hive maintainers/volunteers should come up with a plan on how the tech debt/maintenance with regards to Hive query syntax can be addressed and will be resolved for Flink 1.16. This includes stuff like using public APIs and decoupling it from the planner. This is also extremely important since we want to move out connectors with Flink 1.16 (next Flink release). I'm hoping that those who can help out with this will chip-in. * We follow the Hive versions that are still supported, which means we drop support for Hive 1.*, 2.1.x and 2.2.x and upgrade Hive 2.3 and Hive 3.1 to the latest version. Thanks again for your input and looking forward to your thoughts on this. Best regards, Martijn On Tue, 8 Mar 2022 at 10:39, 罗宇侠(莫辞) <luoyuxia.luoyu...@alibaba-inc.com> wrote: Hi Martijn, Thanks for driving this discussion. About your concerns, I would like to share my opinion. Actually, more exactly, FLIP-152 [1] is not to extend Flink SQL to support Hive query synax, it provides a Hive dialect option to enable users to switch to Hive dialect. From the commits about the corresponding FLINK-21529, it doesn't involve much about Flink itself. - About the struggling with maintaining. The current implementation is just to provide an option for user to use Hive dialect. I think there won't be much bother. - Although Apache Hive is less popular, it's widely used as an open source database over the years. There still exists many Hive SQL jobs in many companies. - As I said, the current implementation for Hive SQL synax is more like pluggable, we can also support for Snowflake and the others as long as it's necessary. - As for the know security vulnerabilities of Hive, maybe it's not a critical problem in this discuss. - For current implementation for Hive SQL syntax, it uses a pluggable HiveParser[3] to parse the SQL statement. I think there won't be much complexity brought to Flink to support Hive syntax. From my perspective, Hive is still widely used and there exists many running Hive SQL jobs, so why not to provide users a better experience to help them migrate Hive jobs to Flink? Also, it doesn't conflict with Flink SQL as it's just a dialect option. [1] https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=165227316 [2] https://issues.apache.org/jira/browse/FLINK-21529 [3] https://issues.apache.org/jira/browse/FLINK-21531 Best regards, Yuxia. ------------------原始邮件 ------------------ 发件人:Martijn Visser <martijnvis...@apache.org> 发送时间:Mon Mar 7 19:23:15 2022 收件人:dev <dev@flink.apache.org>, User <u...@flink.apache.org> 主题:[DISCUSS] Flink's supported APIs and Hive query syntax Hi everyone, Flink currently has 4 APIs with multiple language support which can be used to develop applications: * DataStream API, both Java and Scala * Table API, both Java and Scala * Flink SQL, both in Flink query syntax and Hive query syntax (partially) * Python API Since FLIP-152 [1] the Flink SQL support has been extended to also support the Hive query syntax. There is now a follow-up FLINK-26360 [2] to address more syntax compatibility issues. I would like to open a discussion on Flink directly supporting the Hive query syntax. I have some concerns if having a 100% Hive query syntax is indeed something that we should aim for in Flink. I can understand that having Hive query syntax support in Flink could help users due to interoperability and being able to migrate. However: - Adding full Hive query syntax support will mean that we go from 6 fully supported API/language combinations to 7. I think we are currently already struggling with maintaining the existing combinations, let another one more. - Apache Hive is/appears to be a project that's not that actively developed anymore. The last release was made in January 2021. It's popularity is rapidly declining in Europe and the United State, also due Hadoop becoming less popular. - Related to the previous topic, other software like Snowflake, Trino/Presto, Databricks are becoming more and more popular. If we add full support for the Hive query syntax, then why not add support for Snowflake and the others? - We are supporting Hive versions that are no longer supported by the Hive community with known security vulnerabilities. This makes Flink also vulnerable for those type of vulnerabilities. - The currently Hive implementation is done by using a lot of internals of Flink, making Flink hard to maintain, with lots of tech debt and making things overly complex. From my perspective, I think it would be better to not have Hive query syntax compatibility directly in Flink itself. Of course we should have a proper Hive connector and a proper Hive catalog to make connectivity with Hive (the versions that are still supported by the Hive community) itself possible. Alternatively, if Hive query syntax is so important, it should not rely on internals but be available as a dialect/pluggable option. That could also open up the possibility to add more syntax support for others in the future, but I really think we should just focus on Flink SQL itself. That's already hard enough to maintain and improve on. I'm looking forward to the thoughts of both Developers and Users, so I'm cross-posting to both mailing lists. Best regards, Martijn Visser https://twitter.com/MartijnVisser82 [1] https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=165227316 [2] https://issues.apache.org/jira/browse/FLINK-21529
