git.net

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [DISCUSS] Unified Core API for Streaming and Batch


Hi all,

Really excited to see this discussion really happens, I also want to share
my two cents here.
Lets first focus on this question: “What Flink API Stack Should be for a
Unified Engine".

There are multiply ways to judge whether an engine is unified or not. From
user's perspective, as long as you provides api for
both stream and batch processing, can be considered unified. But that's
definitely not the way a developer sees. I think developers
cares more about the implementations. How many infrastructures are shared
between different compute modes, network, functions,
or even operators? Sharing more things is not a free lunch, it can make
things more complicated, but the potential benefits is also
huge.

In current Flink's implementation, two things are shared. One is the
network stack, and the other one is job scheduling. If we want to
push the unify effort a little further, next thing we should consider is
tasks and operators (BatchTask and Driver for current batch implementation).
Here are the benefits I can see if we try to unify them:
1. State is open to batch processing, make batch failover more
possibilities.
2. Stream processing can borrow some efficient ideas from batch, such as
memory management, binary data processing.
3. Batch & stream operators can be mixed together, to meet more complicated
computation requirements, such as progressive computation,
    which is not pure stream processing or batch processing
4. Make all developers a joint effort, working on same technique stack no
matter you are mainly focus on stream or batch, even ML.

And once the operator api is unified, we can next consider to have a more
formal DAG api for various processing modes. I think we both agree
the idea which Flink built upon: "stream is the basic, batch is just a
special case of streaming". I think it's also make sense to have the DAG
api mainly
focus to describe a stream, whether it is bounded or not. I found
StreamTransformation is good fit for this requirement. It has no semantics,
just tell
you the physical transformation we did on the stream. Like
OneInputStreamTransfomation, all we should know is this takes one stream as
input, have a
operator to process the elements it received, and the output can be further
transformed by another  OneInputStreamTransfomation, or be one input to a
TwoInputStreamTransfomation. It describes how data flows, but have very
limited information about how data be processed, is it be mapped, or be
flatmapped.

All the user API (DataStream, DataSet, Table) we now have, have semantics,
or even consists of optimizers. Based on these thoughts, I will try to
answer the questions Stephan has raised:

  - Relationship of DataStream, DataSet, and Table API
I think these three APIs can be independent for now. DataSet for pure batch
processing, DataStream for pure stream processing and you want to deal with
state explicitly,
Table API for relational data processing.

  - Where do we want automatic optimization, and where not
DataStream, DataSet, and Table API can all have their own optimizations,
but StreamTransformation does not.

  - Future of DataSet (subsumed in data stream, or remains independent)
I think it's better to remains independent for now, and subsumed in data
stream in the future.

  - What happens with iterations
I think the more important question is how to describe iteration on stream
transformations, what information can be hided, and what information must
be exposed to transformation.

  - What happens with the collection execution mode
I think this mode can be fully replaced by mini cluster.

Best,
Kurt


On Tue, Dec 4, 2018 at 12:14 PM Wang Feng <feng.wang@xxxxxxxxxxx> wrote:

> Hi Stephan:
> I totally agree with you, this discussion covers too many topics, so we
> can cut it into a series of sub-discussions proposed by you,  firstly we
> can focus on phrase-1: “What Flink API Stack Should be for a Unified
> Engine”.
> Best,
> Feng Wang
>
> On Dec 3, 2018, at 19:36, Stephan Ewen <sewen@xxxxxxxxxx<mailto:
> sewen@xxxxxxxxxx>> wrote:
>
> Hi all!
>
> This is a great discussion to start and I agree with the idea behind it. We
> should get started designing what the Flink stack should look like in the
> future.
>
> This discussion is very big, though, and from past experiences if the scope
> is too big, the discussions and up falling apart when everyone goes into
> different details.
> So my suggestion would be to stage this discussion and take this aspect
> after aspect, starting with what we want to expose to users and then going
> into the internal details.
>
> *Discussion (1) What should the API stack look like*
>  - Relationship of DataStream, DataSet, and Table API
>  - Where do we want automatic optimization, and where not
>  - Future of DataSet (subsumed in data stream, or remains independent)
>  - What happens with iterations
>  - What happens with the collection execution mode
>
> *Discussion (2) What should the abstractions look like.*
>  - This is based on the outcome of (1)
>  - Operator DAG
>  - Operator Interface
>  - what is sent to the REST API when a job is submitted, etc.
>  - modules and dependency structure
>
> *Discussion (3) what is special for Batch*
>  - I would like to follow the philosophy that "batch allows us to activate
> additional optimizations" made possible by the bounded nature of the
> inputs.
>  - special case scheduling
>  - additional runtime algorithms (like hybrid hash joins)
>  - no watermarks / late data / etc.
>  - Special casing in failover (or possibly not, could be still the same
> core mechanism
>
> What do you think?
>
> Best,
> Stephan
>
>
>
> On Mon, Dec 3, 2018 at 12:17 PM Haibo Sun <sunhaibotb@xxxxxxx<mailto:
> sunhaibotb@xxxxxxx>> wrote:
>
> Thanks, zhijiang.
>
>
> For the optimization, such as cost-based estimation, we still want to keep
> it in the data set layer,
> but your suggestion is also a thought that can be considered.
>
>
> As I know, currently these batch scenarios have been contained in DataSet,
> such as
> the sort-merge join algorithm. So I think that the unification should
> consider such features
> as input selection at reading.
>
>
> Best,
> Haibo
>
>
> At 2018-12-03 16:38:13, "zhijiang" <wangzhijiang999@xxxxxxxxxx
> .INVALID<mailto:wangzhijiang999@xxxxxxxxxx.INVALID>>
> wrote:
> Hi haibo,
>
> Thanks for bringing this discussion!
>
> I reviewd the google doc and really like the idea of unifying the stream
> and batch in all stacks. Currently only network runtime stack is unified
> for both stream and batch jobs, but the compilation, operator and runtime
> task stacks are all separate. The stream stack developed frequently and
> behaved dominantly these years, but the batch stack was touched less. If
> they are unified into one stack, the batch jobs can also get benefits from
> all the improvements. I think it is a very big work but worth doing, left
> some concerns:
>
> 1. The current job graph generation for batch covers complicated
> optimization such as cost-based estimate, plan etc. Would this part also be
> considered retaining during integrating with stream graph generation?
>
> 2. I saw some other special improvements for batch scenarios in the doc,
> such as input selection while reading. I acknowledge these roles for
> special batch scenarios, but they seem not the blocker for unification
> motivation, because current batch jobs can also work without these
> improvements. So the further improvments can be separated into individual
> topics after we reaching the unification of stream and batch firstly.
>
> Best,
> Zhijiang
>
>
> ------------------------------------------------------------------
> 发件人:孙海波 <sunhaibotb@xxxxxxx<mailto:sunhaibotb@xxxxxxx>>
> 发送时间:2018年12月3日(星期一) 10:52
> 收件人:dev <dev@xxxxxxxxxxxxxxxx<mailto:dev@xxxxxxxxxxxxxxxx>>
> 主 题:[DISCUSS] Unified Core API for Streaming and Batch
>
> Hi all,
> This post proposes unified core API for Streaming and Batch.
> Currently DataStream and DataSet adopt separated compilation processes,
> execution tasks
> and basic programming models in the runtime layer, which complicates the
> system implementation.
> We think that batch jobs can be processed in the same way as streaming
> jobs, thus we can unify
> the execution stack of DataSet into that of DataStream.  After the
> unification the DataSet API will
> also be built on top of StreamTransformation, and its basic programming
> model will be changed
> from "UDF on Driver" to "UDF on StreamOperator". Although the DataSet
> operators will need to
> implement the interface StreamOperator instead after the unification,
> user jobs do not need to change
> since DataSet uses the same UDF interfaces as DataStream.
>
> The unification has at least three benefits:
> 1. The system will be greatly simplified with the same execution stack
> for both streaming and batch jobs.
> 2. It is no longer necessary to implement two sets of Driver(s) (operator
> strategies) for batch, namely chained and non-chained.
> 3. The unified programming model enables streaming and batch jobs to
> share the same operator implementation.
>
>
> The following is the design draft. Any feedback is highly appreciated.
>
>
> https://docs.google.com/document/d/1G0NUIaaNJvT6CMrNCP6dRXGv88xNhDQqZFrQEuJ0rVU/edit?usp=sharing
>
> Best,
> Haibo
>
>
>