I would really like to see us converging the stack and the functionality here. Meaning to try and use the same sinks in the Table API as for the DataStream API, and using the same sink for batch and streaming.
The StreamingFileSink has a lot of things that can help with that. If possible, it would be nice to extend it (which would help move towards the above goal) rather than build a second sink. Building a second sink leads us further away from unification. I am a bit puzzled by the statement that sinks are primarily for Hive. The Table API should not be coupled to Hive, it should be an independent batch/streaming API for many use cases, supporting very well for batch and streaming interplay. Supporting Hive is great, but we should not be building this towards Hive, as just yet another Hive runtime. Why "yet another Hive runtime" when what we have a unique streaming engine that can do much more? We would drop our own strength and reduce ourselves to a limited subset. Let's build a File Sink that can also support Hive, but can do so much more. For example, efficient streaming file ingestion as materialized views from changelogs. *## Writing Files in Streaming* To write files in streaming, I don't see another way than using the streaming file sink. If you want to write files across checkpoints, support exactly-once, and support consistent "stop with savepoint", it is not trivial. A part of the complexity comes from the fact that not all targets are actually file systems, and not all have simple semantics for persistence. S3 for example does not support renames (only copies, which may take a lot of time) and it does not support flush/sync of data (the S3 file system in Hadoop exposes that but it does not work. flush/sync, followed by a failure, leads to data loss). You need to devise a separate protocol for that, which is exactly what has already been done and abstracted behind the recoverable writers. If you re-engineer that in the, you will end up either missing many things (intermediate persistence on different file systems, and atomic commit in the absence of renames, etc.), or you end up doing something similar as the recoverable writers do. *## Atomic Commit in Batch* For batch sinks, it is also desirable to write the data first and then atomically commit it once the job is done. Hadoop has spent a lot of time making this work, see this doc here, specifically the section on 'The "Magic" Committer'. [1] What Flink has built in the RecoverableWriter is in some way an even better version of this, because it works without extra files (we pass data through checkpoint state) and it supports not only committing once at the end, but committing multiple time intermediate parts during checkpoints. Meaning using the recoverable writer mechanism in batch would allow us to immediately get the efficient atomic commit implementations on file:// hdfs:// and s3://, with a well defined way to implement it also for other file systems. *## Batch / Streaming Unification* It would be great to start looking at these things in the same way: - streaming (exactly-once): commits files (after finished) at the next checkpoint - batch: single commit at the end of the job *## DataStream / Table API Stack Unification* Having the same set of capabilities would make it much easier for users to understand the system. Especially when it comes to consistent behavior across external systems. Having a different file sink in Table API and DataStream API means that DataStream can write correctly to S3 while Table API cannot. *## What is missing?* It seems there are some things that get in the way of naturally Can you make a list of what features are missing in the StreamingFileSink that make it usable for the use cases you have in mind? Best, Stephan [1] https://hadoop.apache.org/docs/current/hadoop-aws/tools/hadoop-aws/committer_architecture.html On Mon, Mar 16, 2020 at 12:31 PM Jingsong Li <jingsongl...@gmail.com> wrote: > Hi Piotr, > > I am very entangled. > > Let me re-list the table streaming sink requirements: > - In table, maybe 90% sinks are for Hive. The parquet and orc are the most > important formats. Hive provide RecordWriters, it is easy to support all > hive formats by using it, and we don't need concern hive version > compatibility too, but it can not work with FSDataOutputStream. > - Hive table maybe use external HDFS. It means, hive has its own hadoop > configuration. > - In table, partition commit is needed, we can not just move files, it is > important to complete table semantics to update catalog. > > You are right DataStream and Table streaming sink will not be fully > compatible, each with its own set of limitations, quirks and features. > But if re-using DataStream, batch and streaming also will not be fully > compatible. Provide a unify experience to batch and streaming is also > important. > > Table and DataStream have different concerns, and they tilt in different > directions. > > Of course, it is very good to see a unify implementation to solve batch > sink and hive things, unify DataStream batch sink and DataStream streaming > sink and Table batch sink and Table streaming sink. > > Le's see what others think. > > Best, > Jingsong Lee > > > On Mon, Mar 16, 2020 at 4:15 PM Piotr Nowojski <pi...@ververica.com> > wrote: > > > Hi Jingsong, > > > > > First way is reusing Batch sink in FLINK-14254, It has handled the > > partition and metastore logic well. > > > - unify batch and streaming > > > - Using FileOutputFormat is consistent with FileInputFormat. > > > - Add exactly-once related logic. Just 200+ lines code. > > > - It's natural to support more table features, like partition commit, > > auto compact and etc.. > > > > > > Second way is reusing Datastream StreamingFileSink: > > > - unify streaming sink between table and Datastream. > > > - It maybe hard to introduce table related features to > StreamingFileSink. > > > > > > I prefer the first way a little. What do you think? > > > > I would be surprised if adding “exactly-once related logic” is just 200 > > lines of code. There are things like multi part file upload to s3 and > there > > are also some pending features like [1]. I would suggest to ask/involve > > Klou in this discussion. > > > > If it’s as easy to support exactly-once streaming with current batch > sink, > > that begs the question, why do we need to maintain StreamingFileSink? > > > > The worst possible outcome from my perspective will be, if we have > another > > example of an operator/logic implemented independently both in DataStream > > API and Table API. Because I’m pretty sure they will not be fully > > compatible, each with it’s own set of limitations, quirks and features. > > Especially that we have on our long term roadmap and wish list to unify > > such kind of operators. > > > > Piotrek > > > > [1] https://issues.apache.org/jira/browse/FLINK-11499 < > > https://issues.apache.org/jira/browse/FLINK-11499> > > > > > On 16 Mar 2020, at 06:55, Jingsong Li <jingsongl...@gmail.com> wrote: > > > > > > Thanks Jinhai for involving. > > > > > >> we need add 'connector.sink.username' for UserGroupInformation when > data > > > is written to HDFS > > > > > > Yes, I am not an expert of HDFS, but it seems we need do this "doAs" in > > the > > > code for access external HDFS. I will update document. > > > > > > Best, > > > Jingsong Lee > > > > > > On Mon, Mar 16, 2020 at 12:01 PM Jingsong Li <jingsongl...@gmail.com> > > wrote: > > > > > >> Thanks Piotr and Yun for involving. > > >> > > >> Hi Piotr and Yun, for implementation, > > >> > > >> FLINK-14254 [1] introduce batch sink table world, it deals with > > partitions > > >> thing, metastore thing and etc.. And it just reuse Dataset/Datastream > > >> FileInputFormat and FileOutputFormat. Filesystem can not do without > > >> FileInputFormat, because it need deal with file things, split things. > > Like > > >> orc and parquet, they need read whole file and have different split > > logic. > > >> > > >> So back to file system connector: > > >> - It needs introducing FilesystemTableFactory, FilesystemTableSource > and > > >> FilesystemTableSink. > > >> - For sources, reusing Dataset/Datastream FileInputFormats, there are > no > > >> other interface to finish file reading. > > >> > > >> For file sinks: > > >> - Batch sink use FLINK-14254 > > >> - Streaming sink has two ways. > > >> > > >> First way is reusing Batch sink in FLINK-14254, It has handled the > > >> partition and metastore logic well. > > >> - unify batch and streaming > > >> - Using FileOutputFormat is consistent with FileInputFormat. > > >> - Add exactly-once related logic. Just 200+ lines code. > > >> - It's natural to support more table features, like partition commit, > > auto > > >> compact and etc.. > > >> > > >> Second way is reusing Datastream StreamingFileSink: > > >> - unify streaming sink between table and Datastream. > > >> - It maybe hard to introduce table related features to > > StreamingFileSink. > > >> > > >> I prefer the first way a little. What do you think? > > >> > > >> Hi Yun, > > >> > > >>> Watermark mechanism might not be enough. > > >> > > >> Watermarks of subtasks are the same in the "snapshotState". > > >> > > >>> we might need to also do some coordination between subtasks. > > >> > > >> Yes, JobMaster is the role to control subtasks. Metastore is a very > > >> fragile single point, which can not be accessed by distributed, so it > is > > >> uniformly accessed by JobMaster. > > >> > > >> [1]https://issues.apache.org/jira/browse/FLINK-14254 > > >> > > >> Best, > > >> Jingsong Lee > > >> > > >> On Fri, Mar 13, 2020 at 6:43 PM Yun Gao <yungao...@aliyun.com> wrote: > > >> > > >>> Hi, > > >>> > > >>> Very thanks for Jinsong to bring up this discussion! It should > > >>> largely improve the usability after enhancing the FileSystem > connector > > in > > >>> Table. > > >>> > > >>> I have the same question with Piotr. From my side, I think it > > >>> should be better to be able to reuse existing StreamingFileSink. I > > think We > > >>> have began > > >>> enhancing the supported FileFormat (e.g., ORC, Avro...), and > > >>> reusing StreamFileSink should be able to avoid repeat work in the > Table > > >>> library. Besides, > > >>> the bucket concept seems also matches the semantics of > partition. > > >>> > > >>> For the notification of adding partitions, I'm a little > wondering > > >>> that the Watermark mechanism might not be enough since > Bucket/Partition > > >>> might spans > > >>> multiple subtasks. It depends on the level of notification: if > we > > >>> want to notify for the bucket on each subtask, using watermark to > > notifying > > >>> each subtask > > >>> should be ok, but if we want to notifying for the whole > > >>> Bucket/Partition, we might need to also do some coordination between > > >>> subtasks. > > >>> > > >>> > > >>> Best, > > >>> Yun > > >>> > > >>> > > >>> > > >>> ------------------------------------------------------------------ > > >>> From:Piotr Nowojski <pi...@ververica.com> > > >>> Send Time:2020 Mar. 13 (Fri.) 18:03 > > >>> To:dev <dev@flink.apache.org> > > >>> Cc:user <u...@flink.apache.org>; user-zh <user...@flink.apache.org> > > >>> Subject:Re: [DISCUSS] FLIP-115: Filesystem connector in Table > > >>> > > >>> Hi, > > >>> > > >>> > > >>> Which actual sinks/sources are you planning to use in this feature? > Is > > it about exposing StreamingFileSink in the Table API? Or do you want to > > implement new Sinks/Sources? > > >>> > > >>> Piotrek > > >>> > > >>>> On 13 Mar 2020, at 10:04, jinhai wang <jinhai...@gmail.com> wrote: > > >>>> > > >>> > > >>>> Thanks for FLIP-115. It is really useful feature for platform > > developers who manage hundreds of Flink to Hive jobs in production. > > >>> > > >>>> I think we need add 'connector.sink.username' for > > UserGroupInformation when data is written to HDFS > > >>>> > > >>>> > > >>>> 在 2020/3/13 下午3:33，“Jingsong Li”<jingsongl...@gmail.com> 写入: > > >>>> > > >>>> Hi everyone, > > >>>> > > >>> > > >>>> I'd like to start a discussion about FLIP-115 Filesystem connector > > in Table > > >>>> [1]. > > >>>> This FLIP will bring: > > >>>> - Introduce Filesystem table factory in table, support > > >>>> csv/parquet/orc/json/avro formats. > > >>>> - Introduce streaming filesystem/hive sink in table > > >>>> > > >>> > > >>>> CC to user mail list, if you have any unmet needs, please feel > free > > to > > >>>> reply~ > > >>>> > > >>>> Look forward to hearing from you. > > >>>> > > >>>> [1] > > >>>> > > >>> > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-115%3A+Filesystem+connector+in+Table > > >>>> > > >>>> Best, > > >>>> Jingsong Lee > > >>>> > > >>>> > > >>>> > > >>> > > >>> > > >>> > > >> > > >> -- > > >> Best, Jingsong Lee > > >> > > > > > > > > > -- > > > Best, Jingsong Lee > > > > > > -- > Best, Jingsong Lee >
