Thanks to the guys from ByteDance driving this topic, which could be another big story to extend Flink's ability.
In general, I think this is a great idea. However, before we move forward, I think we should first answer the question: which is the target for Flink OLAP? We run Presto/Trino and SparkSQL in the production environment for OLAP SQL analysis. Since Presto runs faster than SparkSQL in many cases, especially for ad-hoc queries at medium-sized data, we would run queries on Presto first or switch to SparkSQL for large-scale queries if necessary. Presto runs as a service and emphasis on query performance without node fault tolerance. Moreover, it leverages a pipeline-like data exchange mode instead of the classic stage blocking exchange mode, which is a bit like Flink's pipeline mode vs blocking mode. Can we say we hope Flink OLAP could target Presto/Trino in medium-sized data query, and switch to Flink batch SQL for large-scale analysis query? If so, I also think the naming of Flink OLAP looks a bit strange, as Flink batch SQL shall also serve for large-scale OLAP analysis. Best Yun Tang ________________________________ From: Jing Ge <j...@ververica.com.INVALID> Sent: Thursday, August 10, 2023 13:52 To: dev@flink.apache.org <dev@flink.apache.org> Subject: Re: [DISCUSS] How about adding OLAP to Flink Roadmap? Hi Shammon, Hi Xiangyu, Thanks for bringing this to our attention. I can see this is a great proposal born from real business scenarios. +1 for it. People have been keen to use one platform to cover all their data production and consumption requirements. Flink did a great job for the production, i.e. streaming and batch processing with all excellent ecosystems. This is the big advantage for Flink to go one step further and cover the consumption part. It will turn Flink into a unified compute platform like what the Ray project(the platform behind ChatGPT, if someone is not aware of it)[1] is doing and secure Flink to be one of the most interesting open source platforms for the next decade. Frankly speaking, it will be a big change. As far as I am concerned, the following should be considered(just thought about it at the first glance, there must be more). Architecture upgrade - since we will have three capabilities(I wanted to use "engines", but it might be too early to use the big word), i.e. streaming, batch, OLAP, it might make sense to upgrade the architecture while we are building the OLAP in Flink. The unified foundation or abstraction for distributed computation should be designed and implemented underneath those capabilities. In the future, new capabilities can leverage the foundation and could be developed at a very fast pace. MPP architecture - Flink session cluster is not the MMP architecture. Commonly speaking, SNA(shared nothing architecture) is the key that could implement MPP. Flink has everything to offer SNA. That is the reason why we can consider building OLAP into or on top of the Flink. And speaking of MPP, there will be a lot of things to do, e.g. the Retrieval Architecture[2], multiple level task split, dynamic retry or even split, etc. I will not expand all those topics at this early stage. OLAP queries syntax - at least some common syntax and statements need to be implemented, e.g. cube, grouping set, over partition by, you mention it. Last but not least, there will be a big effort to upgrade the runtime features to support OLAP wrt the performance and latency. Best regards, Jing [1] https://www.ray.io/ [2] https://www.tutorialsbook.com/teradata/teradata-architecture On Thu, Aug 10, 2023 at 11:39 AM Dan Zou <zoud...@163.com> wrote: > Thanks for bringing up this discussion, Shammon. I would like to share > some of my observations and experiences. > > Flink has almost become the de facto standard for streaming computing, and > Flink batch have been successfully applied in some companies. If Flink can > support OLAP scenarios well, a unified engine to support streaming, batch, > and OLAP will become a reality, which is very exciting. > > Based on the status quo, Flink can be used as a primary OLAP engine, > although there is still a lot of room for optimization. This means that we > do not need to carry out large-scale renovation at the beginning, but only > gradually and continuously enhance it without affecting streaming. > > Flink OLAP can largely reuse the capabilities of Flink Batch SQL and > optimizations in OLAP can also benefit Flink Batch. If we simplify job > startup overhead and increase cross-job resource reuse (Plan reuse, > Generated class reuse, Connection reuse, etc.) on this basis, Flink will > become a good OLAP engine. > > So, I am big +1 for adding OLAP to Flink Roadmap, and I am willing to > contribute to it. > > > > 2023年8月9日 15:35，xiangyu feng <xiangyu...@gmail.com> 写道： > > > > Thank you Shammon for initiating this discussion. As one of the Flink > OLAP > > developers in ByteDance, I would also like to share a real case of our > > users. > > > > About two years ago we found our first OLAP user internally by > integrating > > Flink OLAP with ByteHTAP. Users are willing to use Flink as an OLAP > engine > > mainly hoping to use Flink's rich cross-datasource join capability. In > the > > beginning, we only support simple query patterns with qps less than 2 and > > joins less than 5. With the business growing and our system capabilities > > evolving, users have moved more scenarios to Flink OLAP, and the query > > pattern is getting more and more complicated. Until early this year, the > > user's query pattern has changed to peak QPS greater than 20, join table > > number greater than 30 and scan data volume exceeding 1 billion rows. > Even > > with the evolution of our engine over the past two years, computing at > this > > scale is still very challenging. It is difficult to satisfy the > computation > > scale, system stability and query latency at the same time. > > > > Through talking to our user, we easily build some intermediate views by > > using Flink's streaming and batch engine. In a similar way to > materialized > > view, we have optimized user's query pattern to single query join less > than > > 10, scan data volume in tens of millions and QPS remains unchanged. In > this > > way, our OLAP service has not only perfectly met the business > requirements > > and also we have made this migration process very smooth, thanks to > Flink's > > powerful streaming and batch computing ecosystem. Finally, we are highly > > recognized by our users. > > > > *There are two points I want to make with this case:* > > > > 1, Although there are many OLAP engines out there, Flink may not always > > provide the best performance. But thanks to Flink's strong ecosystem, we > > are confident that we can build an OLAP engine that provides a great user > > experience. This is very important for many small and medium sized > > companies; > > > > 2, From another perspective, I personally believe that building OLAP will > > 'bring Flink closer to our end-users' and present a wider variety of > > computational challenges to Flink. As the case mentioned above, this is a > > very common case in data analytics, where the flink is used to precompute > > the data and feed it to OLAP services. These precalculations are often > > designed to compensate for the capabilities of other OLAP engines, such > as > > some engines may not have strong join capabilities, some may not have > > complete SQL support and some may have weak plan optimization support. In > > general, Flink will not face these users directly, and thus cannot build > a > > comprehensive end-to-end solution to solve this problem once and for > all. *By > > building Flink OLAP, we can finally fill the last missing block in the > > puzzle!* > > > > Of course, as we built Flink OLAP internally, we encountered many > > challenging issues, which is why we are putting this discussion out there > > and hoping to involve more contributors. Meanwhile, We also hope to > > contribute our optimizations back to the community through FLINK-25318. > > > > So for me, it is a big +1 to add OLAP to Flink Roadmap. ^-^ > > > > Best, > > Xiangyu > > > > Xintong Song <tonysong...@gmail.com> 于2023年8月8日周二 18:01写道： > > > >> Thanks for bringing this up, Shammon. > >> > >> In general, I'd be +1 to improve Flink's ability to serve as an OLAP > >> engine. > >> > >> I see a great value in Flink becoming a unified Large-scale Data > Processing > >> / Analysis tool. Through my interactions with users (Alibaba internal > >> users, external users on Alibaba Cloud, developers from other companies > via > >> conferences / meetups), it's commonly complained how complicated and > costly > >> it is to build a data processing platform out of a bunch of different > >> tools. That usually means higher learning / developing / operation and > >> maintenance cost. In addition, I also see a trend that many projects and > >> products are going along the same direction. > >> > >> I personally would not be concerned about losing focus. Unlike Apache > >> Paimon (Flink Table Store) which tries to solve a completely different > >> problem other than data processing, OLAP querying is just a special > case of > >> batch SQL data processing, where typically you have massive concurrent > >> short-lived queries. As Shammon mentioned, Flink already has most of the > >> essential building blocks: batch SQL processing, session mode, > sql-gateway, > >> etc. IMHO, the missing piece is mostly about improving the performance > in > >> the specific OLAP scenarios. That sounds like a reasonable extension to > me. > >> > >> I'd consider improving the OLAP capability as nice-to-have improvements. > >> That is to say, it must not come at the price of sacrificing the > >> experiences in other streaming / batch scenarios, nor significantly > >> complicate the system. I think one of the reasons that FLINK-25318 > became > >> stale was that some of the proposed solutions are too dedicated for the > >> OLAP scenarios and require extra efforts to carefully re-design in > order to > >> not affect other scenarios. I'd be glad to see such efforts being > revived. > >> > >> Regarding whether to include this in the roadmap, I'd suggest including > it > >> but maybe not as a top-level section. I think this is a good initiative. > >> But to be considered as one of the major directions, I think we need > some > >> more clear and concrete, and commonly agreed plans for it. > Alternatively, > >> we may consider putting it under the "Towards Streaming Warehouses" or > >> "Performance" sections, and using tones like "the community is trying to > >> explore the possibility of ..." to indicate its exploratory stage. > >> > >> > >> Best, > >> > >> Xintong > >> > >> > >> > >> On Tue, Aug 8, 2023 at 11:16 AM Yangze Guo <karma...@gmail.com> wrote: > >> > >>> Thanks for driving this, Shammon! And thanks for the valuable comments > >>> from Matthias. > >>> > >>> As one of the contributors to FLINK-25318, I would also like to share > >>> my thoughts. > >>> > >>> 1. Regarding using Flink as an OLAP engine: > >>> > >>> I understand the value of having a unified engine for streaming, > >>> batch, and OLAP, as described. From the user's perspective, using the > >>> same engine throughout different stages of their daily workflow can > >>> definitely save a lot of learning costs. It also gives us the > >>> opportunity to leverage fragment resources in a batch standalone > >>> cluster, thereby improving resource utilization efficiency. > >>> @Matthias I agree with the Unix philosophy you mentioned, but I would > >>> consider streaming/batch ETL and OLAP queries as a complete story. > >>> Just as the unified stream-batch engine of Flink frees users from > >>> maintaining two sets of business code, expanding Flink's use as an > >>> OLAP query engine can further reduce costs for users aiming to build > >>> an end-to-end streaming warehouse. AFAIK, within the ByteDance, users > >>> are already benefiting from this. > >>> > >>> > >>> 2. Regarding adding OLAP support to Flink 2.0 roadmap: > >>> > >>> Given the increasing user expectations for Flink's performance and > >>> functionality in OLAP scenarios, it seems reasonable to add OLAP > >>> support to the roadmap. However, I agree that introducing too many > >>> objectives can lead to scattered attention. Therefore, I'm partially > >>> in favor (+0.5) of it. > >>> > >>> BTW, in addressing concerns about the growing complexity of Flink's > >>> codebase, we can mitigate this issue through careful design choices > >>> and thorough code reviews. Another principle we need to follow is all > >>> of those improvements should not negatively impact stream processing. > >>> AFAICS, these enhancements will also benefit streaming users. > >>> > >>> Second to @Matthias opinion, at least in a first step, we can consider > >>> OLAP queries as short-lived batch processing and focus on driving > >>> optimizations specifically for these tasks. > >>> > >>> 3. Regarding the performance tests: > >>> > >>> I'm big +1 for adding a specific performance tests for OLAP or > >>> short-lived queries. In addition to add some micro benchmarks in [1], > >>> we can also develop an end-to-end benchmark similar to Nexmark[2]. I'm > >>> enthusiastic about participating in this project. > >>> > >>> [1] https://github.com/apache/flink-benchmarks > >>> [2] https://github.com/nexmark/nexmark > >>> > >>> > >>> Best, > >>> Yangze Guo > >>> > >>> On Mon, Aug 7, 2023 at 8:20 PM Matthias Pohl > >>> <matthias.p...@aiven.io.invalid> wrote: > >>>> > >>>> Thanks Shammon FY for starting this discussion. > >>>> > >>>> I'm not sure whether we have to expand the focus of Apache Flink in a > >> way > >>>> you're describing it with OLAP being a topic next to Batch and Stream > >>>> Processing. You've rightfully pointed out that there are already > >>> solutions > >>>> to cover OLAP. In this sense, I like the Unix philosophy: One tool for > >>> one > >>>> job [1]. My concern is that extending the scope of Flink would guide > us > >>>> into a direction where we lose focus: I see the risk in Flink's > >> codebase > >>>> becoming even more complex with the proposed change of the roadmap. > >> That > >>>> was one reason why a project like Flink Table Store ended up becoming > >> an > >>>> independent project Apache Paimon. > >>>> > >>>> That said, I still see value in some of the ideas you've collected in > >>>> FLINK-25318 [2]. Can't we just see OLAP queries as a specific form of > >>> Batch > >>>> Processing? The idea of the session cluster was to have the ability to > >>> run > >>>> multiple short-lived jobs, anyway. And there are definitely ways to > >>> improve > >>>> the execution of those queries as you pointed out. > >>>> > >>>> Therefore, I'm not rejecting the ideas you're suggesting in > FLINK-25318 > >>>> [2]. I'm just skeptical about expanding the scope of Apache Flink by > >>>> mentioning OLAP explicitly in Flink's roadmap. > >>>> > >>>> Independent of the roadmap discussion: If the community decides to put > >>> more > >>>> focus on short-lived queries as you suggested, I would like to have > >> that > >>>> reflected in the performance tests (like it was suggested by Piotr in > >> his > >>>> FLINK-25318 comment [3] and that's already covered by FLINK-25356 > [4]). > >>> We > >>>> should provide a performance test set before we proceed with > >> implementing > >>>> improvements to cover those scenarios. This would enable us to measure > >>> the > >>>> changes that were collected in FLINK-25318 [2]. > >>>> > >>>> I'm looking forward to other views on that topic. > >>>> > >>>> Best, > >>>> Matthias > >>>> > >>>> [1] http://www.catb.org/~esr/writings/taoup/html/ch01s06.html > >>>> [2] https://issues.apache.org/jira/browse/FLINK-25318 > >>>> [3] > >>>> > >>> > >> > https://issues.apache.org/jira/browse/FLINK-25318?focusedCommentId=17460615&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-17460615 > >>>> [4] https://issues.apache.org/jira/browse/FLINK-25356 > >>>> > >>>> On Mon, Aug 7, 2023 at 11:53 AM Shammon FY <zjur...@gmail.com> wrote: > >>>> > >>>>> Hi everyone, > >>>>> > >>>>> According to the discussion between @Matthias and me in FLINK-32667 > >>> [1], I > >>>>> would like to initiate a discussion about Flink OLAP. Let me > >> introduce > >>>>> ourselves first, we are the Streaming Compute team in ByteDance, and > >>> about > >>>>> half a year ago @Songxintong and we had an offline discussion on > >> Flink > >>> OLAP > >>>>> and created FLINK-25318 [2] to track the improvements to Flink OLAP. > >> We > >>>>> also have some simple online discussion with @Piotr on the issue. > >> Here > >>> I > >>>>> would like to share my thoughts on Flink OLAP and our practice in the > >>> past, > >>>>> and initiate a broader discussion to get more input and thoughts > >> about > >>>>> Flink OLAP in this thread. I hope if possible that OLAP could be the > >>> main > >>>>> capability besides Streaming and Batch for Flink and the community > >>> could > >>>>> consider adding it to Flink 2.0 Roadmap, which can attract more > >>> developers > >>>>> and users to promote and improve Flink OLAP in the future. > >>>>> > >>>>> 1. Why we want to support OLAP in Flink > >>>>> > >>>>> Currently, there are many excellent open-source OLAP engines, so why > >>> do we > >>>>> want to support OLAP in Flink? There are three main types in the > >>> current > >>>>> data processing progress: streaming, batch and OLAP. As an excellent > >>>>> streaming and batch processing engine, flink processes data and > >> writes > >>>>> results to storage such as lakehouse, key-value stores, databases and > >>> ect. > >>>>> After that, users need an OLAP engine to analyze their data. > >>>>> > >>>>> 1.1) As we all know, OLAP is a very important scenario and users need > >>> an > >>>>> OLAP engine indeed. > >>>>> > >>>>> 1.2) Flink OLAP can reduce team costs. We have encountered many small > >>> or > >>>>> medium-sized tech teams who are using Flink for streaming and batch > >>>>> processing. However, they have to introduce an additional OLAP engine > >>> for > >>>>> their users to analyze data, which increases the technical costs and > >>>>> operation costs for multiple engines and clusters. > >>>>> > >>>>> 1.3) Flink OLAP can reduce user usage costs. Many of our users are > >>>>> familiar with Flink and use Flink streaming and batch SQL to process > >>> data. > >>>>> After that, they are more focused on business. They are not very > >>> interested > >>>>> in introducing an additional OLAP engine which requires a new > >>> familiarity > >>>>> with the new engine and increases usage costs. > >>>>> > >>>>> 2. Why Flink can be an OLAP engine. > >>>>> > >>>>> So, technically speaking, can Flink support OLAP well? My answer is: > >> OF > >>>>> COURSE! > >>>>> > >>>>> 2.1) From Architecture, Flink supports jdbc driver, Sql-Gateway and > >>>>> session mode. Flink session cluster is a typical MPP architecture and > >>> each > >>>>> query does not need to require new resources. Users can easily submit > >>>>> SELECT statements by the jdbc driver and fetch results on the second > >> or > >>>>> even sub-second level. > >>>>> > >>>>> 2.2) Powerful batch processing power. Flink OLAP can take many batch > >>>>> operators and optimizations. At the same time, there are also large > >>> queries > >>>>> in OLAP and Flink can support them based on the ability of Flink > >> batch > >>>>> without the need to introduce an external batch engine like other > >> OLAP > >>>>> engines. > >>>>> > >>>>> 2.3) Flink supports standard SQL syntax such as QUERY/INSERT/UPDATE > >>> which > >>>>> meets the interaction requirements of OLAP users. > >>>>> > >>>>> 2.4) Powerful connector ecosystem. Flink has defined comprehensive > >>>>> interfaces for input and output and implemented many embedding > >>> connectors > >>>>> such as database, lakehouse. Users can easily implement customized > >>>>> connectors based on these interfaces too. > >>>>> > >>>>> 3. The main issues for Flink OLAP. > >>>>> > >>>>> We found there are two main types of issues that need to be addressed > >>> for > >>>>> Flink OLAP in our practice: Latency and QPS. Unlike streaming and > >>> batch, > >>>>> OLAP has the following two characteristics: > >>>>> > >>>>> a) OLAP jobs are usually short-lived and return results in seconds or > >>>>> sub-seconds which will have higher requirements for data processing > >>> latency. > >>>>> > >>>>> b) Users have QPS requirements for OLAP such as users often want to > >>>>> perform hundreds of simple queries or tens of complex queries in a > >>> Flink > >>>>> OLAP cluster concurrently in our practice. > >>>>> > >>>>> We need to make some improvements in Flink to support OLAP better. > >>> Based > >>>>> on our previously created FLINK-25318 [2] and our practical > >>> experience, I > >>>>> classify the improvements as follows: > >>>>> > >>>>> a) Improve the interaction between Flink and external components such > >>> as > >>>>> zk and filesystem. For example, Flink stores relevant information in > >>> zk to > >>>>> support failover and this is not needed by OLAP but rather causes > >>> delays > >>>>> due to the network jitter. > >>>>> > >>>>> b) Improvements in Flink internal progress, such as job submission > >>>>> progress, data fetch optimization, excessive number of interactive > >>> events > >>>>> among components(JM/TM/RM) etc. These will cause Flink to increase > >> job > >>>>> latency from a few milliseconds to tens of seconds when running > >> several > >>>>> small queries concurrently. > >>>>> > >>>>> c) Support OLAP specific features. Due to the OLAP characteristics, > >>> Flink > >>>>> may need to add some features only for OLAP. For example, we need to > >>>>> isolate resources between jobs, and ensure that large queries do not > >>> affect > >>>>> small queries. > >>>>> > >>>>> 4. Out practice about Flink OLAP. > >>>>> > >>>>> In addition to using Flink to support massive streaming businesses in > >>>>> ByteDance, we have also conducted many big improvements on Flink to > >>> support > >>>>> OLAP well. This can not only be seen as a POC, but we also provide > >>> Flink > >>>>> OLAP clusters in production for our users. > >>>>> > >>>>> After our improvements, we built a 500 core Flink OLAP cluster to > >> test > >>>>> Flink OLAP e2e latency and QPS. There will be nearly a thousand QPS > >> for > >>>>> simple queries of 128 subtasks with tens of milliseconds latency; for > >>>>> complex jobs with two JOIN operators and 700 subtasks, there will be > >>> about > >>>>> sixty QPS with hundreds of milliseconds latency. > >>>>> > >>>>> Our Flink OLAP in production supports more than ten internal > >> businesses > >>>>> and there are more than 6000 cores in all clusters and 500 thousand > >>> queries > >>>>> will be performed each day. > >>>>> > >>>>> 5. My Proposal > >>>>> > >>>>> Above all, I personally think it will provide great value for users > >>> and is > >>>>> technically feasible if Flink supports OLAP and this is also what > >> Flink > >>>>> engine can achieve through some improvements although this is not a > >>> simple > >>>>> matter. I hope the community could consider supporting OLAP as one of > >>> Flink > >>>>> technical directions and add it to the roadmap. As I mentioned at the > >>>>> beginning, this can attract more developers and users to participate > >>> and > >>>>> promote the improvement of Flink OLAP and build Flink as a unified > >>> engine > >>>>> for streaming, batch and OLAP. > >>>>> > >>>>> Looking forward to your feedback, thanks! > >>>>> > >>>>> > >>>>> [1] https://issues.apache.org/jira/browse/FLINK-32667 > >>>>> [2] https://issues.apache.org/jira/browse/FLINK-25318 > >>>>> > >>>>> Best, > >>>>> Shammon FY > >>>>> > >>>>> > >>> > >> > >
