Re: [DISCUSS] FLIP-95: New TableSource and TableSink interfaces

Wed, 25 Mar 2020 02:58:32 -0700 

Hi Becket,
Let me clarify a few things first: Historically we thought of Table API/SQL as a library on top of DataStream API. Similar to Gelly or CEP. We used TypeInformation in Table API to integrate nicely with DataStream API. However, the last years have shown that SQL is not just a library. It is an entire ecosystem that defines data types, submission behavior, execution behavior, and highly optimized SerDes. SQL is a way to declare data processing end-to-end such that the planner has the full control over the execution. 


But I totally agree with your concerns around connectors. There is no 
big difference between your concerns and the current design.



1. "native connector interface is a generic abstraction of doing IO and 
Serde":



This is the case in our design. We are using SourceFunction, 
DeserializationSchema, WatermarkAssigner, etc. all pluggable interfaces 
that the DataStream API offers for performing runtime operations.


2. "advanced features ... could be provided in a semantic free way":


I agree here. But this is an orthogonal topic that each connector 
implementer should keep in mind. If a new connector is developed, it 
should *not* be developed only for SQL in mind but with good abstraction 
such that also DataStream API users can use it. A connector should have 
a builder pattern to plugin all capabilities like Parque filters etc. 
There should be no table-specific native/runtime connectors. I think 
this discussion is related to the discussion of FLIP-115.



However, as I mentioned before: This FLIP only discusses the interfaces 
for communication between planner and connector factory. As Dawid said 
earlier, a DynamicTableSource can be more seen as a factory that calls 
pluggable interfaces of a native connextor in the end:


KafkaConnector.builder()
  .watermarkAssigner(...)
  .keyDeser(...)
  .valueDeser(...)
  ....
  .build()

Regards,
Timo


On 25.03.20 09:05, Becket Qin wrote:

Hi Kurt,

I do not object to promote the concepts of SQL, but I don't think we should
do that by introducing a new dedicate set of connector public interfaces
that is only for SQL. The same argument can be applied to Gelly, CEP, and
Machine Learning, claiming that they need to introduce a dedicated public
set of interfaces that fits their own concept and ask the the connector
developers to learn and follow their design. As an analogy, if we want to
promote Chinese, we don't want to force people to learn ancient Chinese
poem while they only need to know a few words like "hello" and "goodbye".

As some design principles, here are what I think what  Flink connectors
should look like:

1. The native connector interface is a generic abstraction of doing IO and
Serde, without semantic for high level use cases such as SQL, Gelly, CEP,
etc.

2. Some advanced features that may help accelerate the IO and Serde could
be provided in the native connector interfaces in a semantic free way so
all the high level use cases can leverage.

3. Additional semantics can be built on top of the native source interface
through providing different plugins. These plugins could be high level use
case aware. For example, to provide a filter to the source, we can do the
following

// An interface for all the filters that take an expression.
interface ExpressionFilter {
     FilterResult applyFilterExpression();
}

// An filter plugin implementation that translate the SQL Expression to a
ParquetFilterPredicate.
Class ParquetExpressionFilter implements Supplier<ParquetFilterPredicate>,
ExpressionFilter {
     // Called by the high level use case,
     FilterResult applyFilterExpression() { ... }

     // Used by the native Source interface.
     ParquetFilterPredicate get() { ... }
}

In this case, the connector developer just need to write the logic of
translating an Expression to Parquet FilterPredicate. They don't have to
understand the entire set of interfaces that we want to promote. Just like
they only need to know how to say "Hello" without learning ancient Chinese
poem.

Again, I am not saying this is necessarily the best approach. But so far it
seems a reasonable design principle to tell the developers.

Thanks,

Jiangjie (becket) Qin



On Wed, Mar 25, 2020 at 11:53 AM Kurt Young <ykt...@gmail.com> wrote:


Hi Becket,

I don't think we should discuss this in pure engineering aspects. Your
proposal is trying
to let SQL connector developers understand as less SQL concepts as
possible. But quite
the opposite, we are designing those interfaces to emphasize the SQL
concept, to bridge
high level concepts into real interfaces and classes.

We keep talking about time-varying relations and dynamic table when
introduce SQL concepts,
sources and sinks are most critical part playing with those concepts. It's
essential to let
Flink SQL developers to learn these concepts and connect them with real
codes by introducing
these connector interfaces and can further write *correct* connectors based
on such domain
knowledge.

So this FLIP is a very important chance to express these concepts and make
most SQL developers
be align with concepts and on same page. It's mostly for different level of
abstractions and for domains
like SQL, it's becoming more important.  It helps Flink SQL go smoothly in
the future, and also
make it easier for new contributors. But I would admit this is not that
obvious for others who don't work
with SQL frequently.

Best,
Kurt


On Wed, Mar 25, 2020 at 11:07 AM Becket Qin <becket....@gmail.com> wrote:


Hi Jark,

It is good to know that we do not expect the end users to touch those
interfaces.

Then the question boils down to whether the connector developers should

be

aware of the interfaces that are only used by the SQL optimizer. It

seems a

win if we can avoid that.

Two potential solutions off the top of my head are:
1. An internal helper class doing the instanceOf based on DataStream

source

interface and create pluggables for that DataStream source.
2. codegen the set of TableSource interfaces given a DataStream Source

and

its corresponding TablePluggablesFactory.

Thanks,

Jiangjie (Becket) Qin

On Wed, Mar 25, 2020 at 10:07 AM Jark Wu <imj...@gmail.com> wrote:


Hi Becket,

Regarding to Flavor1 and Flavor2, I want to clarify that user will

never

use table source like this:

{
    MyTableSource myTableSource = MyTableSourceFactory.create();
    myTableSource.setSchema(mySchema);
    myTableSource.applyFilterPredicate(expression);
    ...
}

TableFactory and TableSource are not directly exposed to end users, all

the

methods are called by planner, not users.
Users always use DDL or descriptor to register a table, and planner

will

find the factory and create sources according to the properties.
All the optimization are applied automatically, e.g. filter/projection
pushdown, users don't need to call `applyFilterPredicate` explicitly.



On Wed, 25 Mar 2020 at 09:25, Becket Qin <becket....@gmail.com> wrote:


Hi Timo and Dawid,

Thanks for the clarification. They really help. You are right that we

are

on the same page regarding the hierarchy. I think the only difference
between our view is the flavor of the interfaces. There are two

flavors

of

the source interface for DataStream and Table source.

*Flavor 1. Table Sources are some wrapper interfaces around

DataStream

source.*
Following this way, we will reach the design of the current proposal,

i.e.

each pluggable exposed in the DataStream source will have a

corresponding

TableSource interface counterpart, which are at the Factory level.

Users

will write code like this:

{
     MyTableSource myTableSource = MyTableSourceFactory.create();
     myTableSource.setSchema(mySchema);
     myTableSource.applyFilterPredicate(expression);
     ...
}

The good thing for this flavor is that from the SQL / Table's

perspective,

there is a dedicated set of Table oriented interface.

The downsides are:
A. From the user's perspective, DataStream Source and Table Source

are

just

two different sets of interfaces, regardless of how they are the same
internally.
B. The source developers have to develop for those two sets of

interfaces

in order to support both DataStream and Table.
C. It is not explicit that DataStream can actually share the

pluggable

in

Table / SQL. For example, in order to provide a filter pluggable with

SQL

expression, users will have to know the actual converter class that
converts the expression to the filter predicate and construct that
converter by themselves.

---------------

*Flavor 2. A TableSource is a DataStream source with a bunch of

pluggables.

No Table specific interfaces at all.*
Following this way, we will reach another design where you have a
SourceFactory and a single Pluggable factory for all the table

pluggables.

And users will write something like:

{
     Deserializer<Row> myTableDeserializer =
MyTablePluggableFactory.createDeserializer(schema)
     MySource<Row> mySource = MySourceFactory.create(properties,
myTableDeserializer);








mySource.applyPredicate(MyTablePluggableFactory.createFilterPredicate(expression));

}

The good thing for this flavor is that there is just one set of

interface

that works for both Table and DataStream. There is no difference

between

creating a DataStream source and creating a Table source. DataStream

can

easily reuse the pluggables from the Table sources.

The downside is that Table / SQL won't have a dedicated API for
optimization. Instead of writing:

if (MyTableSource instanceOf FilterableTableSource) {
      // Some filter push down logic.
      MyTableSource.applyPredicate(expression)
}

One have to write:

if (MySource instanceOf FilterableSource) {
     // Some filter push down logic.








mySource.applyPredicate(MyTablePluggableFactory.createFilterPredicate(expression));

}

-------------------------

Just to be clear, I am not saying flavor 2 is necessarily better than
flavor 1, but I want to make sure flavor 2 is also considered and
discussed.

Thanks,

Jiangjie (Becket) Qin.

On Tue, Mar 24, 2020 at 10:53 PM Dawid Wysakowicz <

dwysakow...@apache.org>

wrote:


Hi Becket,

I really think we don't have a differing opinions. We might not see

the

changes in the same way yet. Personally I think of the

DynamicTableSource

as of a factory for a Source implemented for the DataStream API.

The

important fact about the DynamicTableSource and all feature traits
(SupportsFilterablePushDown, SupportsProjectPushDown etc.) work

with

Table

API concepts such as e.g. Expressions, SQL specific types etc. In

the

end

what the implementation would resemble is (bear in mind I

tremendously

simplified the example, just to show the relation between the two

APIs):


SupportsFilterablePushDown {

   applyFilters(List<ResolvedExpression> filters) {

     this.filters = convertToDataStreamFilters(filters);

   }

   Source createSource() {

         return Source.create()

           .applyFilters(this.filters);

    }

}

or exactly as you said for the computed columns:


SupportsComputedColumnsPushDown {



   applyComputedColumn(ComputedColumnConverter converter) {

     this.deserializationSchema = new DeserializationSchema<Row> {

       Row deserialize(...) {

         RowData row = format.deserialize(bytes); // original

format,

e.g

json, avro, etc.

         RowData enriched = converter(row)

       }

     }

   }

   Source createSource() {

         return Source.create()

           .withDeserialization(deserializationSchema);

    }

}

So to sum it up again, all those interfaces are factories that

configure

appropriate parts of the DataStream API using Table API concepts.

Finally

to answer you question for particular comparisons:

DynamicTableSource v.s. Source<Row, SourceSplitT, EnumChkT>
SupportsFilterablePushDown v.s. FilterableSource
SupportsProjectablePushDown v.s. ProjectableSource
SupportsWatermarkPushDown v.s. WithWatermarkAssigner
SupportsComputedColumnPushDown v.s. ComputedColumnDeserializer
ScanTableSource v.s. ChangeLogDeserializer.

pretty much you can think of all on the left as factories for the

right

side, left side works with Table API classes (Expressions,

DataTypes).

I

hope this clarifies it a bit.

Best,

Dawid
On 24/03/2020 15:03, Becket Qin wrote:

Hey Kurt,

I don't think DataStream should see some SQL specific concepts such

as


Filtering or ComputedColumn.

Projectable and Filterable seems not necessarily SQL concepts, but

could

be

applicable to DataStream source as well to reduce the network load.

For

example ORC and Parquet should probably also be readable from

DataStream,

right?

ComputedColumn is not part of the Source, it is an interface

extends

the

Deserializer, which is a pluggable for the Source. From the SQL's
perspective it has the concept of computed column, but from the

Source

perspective, It is essentially a Deserializer which also converts

the

records internally, assuming we allow some conversion to be

embedded

to

the source in addition to just deserialization.

Thanks,

Jiangjie (Becket) Qin

On Tue, Mar 24, 2020 at 9:36 PM Jark Wu <imj...@gmail.com> <

imj...@gmail.com> wrote:



Thanks Timo for updating the formats section. That would be very

helpful

for changelog supporting (FLIP-105).

I just left 2 minor comment about some method names. In general,

I'm

+1

to

start a voting.










--------------------------------------------------------------------------------------------------


Hi Becket,

I agree we shouldn't duplicate codes, especiall the runtime
implementations.
However, the interfaces proposed by FLIP-95 are mainly used during
optimization (compiling), not runtime.
I don't think there is much to share for this. Because table/sql
is declarative, but DataStream is imperative.
For example, filter push down, DataStream FilterableSource may

allow

to

accept a FilterFunction (which is a black box for the source).
However, table sources should pick the pushed filter expressions,

some

sources may only support "=", "<", ">" conditions.
Pushing a FilterFunction doesn't work in table ecosystem. That

means,

the

connectors have to have some table-specific implementations.


Best,
Jark

On Tue, 24 Mar 2020 at 20:41, Kurt Young <ykt...@gmail.com> <

ykt...@gmail.com> wrote:



Hi Becket,

I don't think DataStream should see some SQL specific concepts such

as

Filtering or ComputedColumn. It's
better to stay within SQL area and translate to more generic

concept

when

translating to DataStream/Runtime
layer, such as use MapFunction to represent computed column logic.

Best,
Kurt


On Tue, Mar 24, 2020 at 5:47 PM Becket Qin <becket....@gmail.com>

<

becket....@gmail.com> wrote:



Hi Timo and Dawid,

It's really great that we have the same goal. I am actually

wondering


if

we

can go one step further to avoid some of the interfaces in Table as

well.

For example, if we have the FilterableSource, do we still need the
FilterableTableSource? Should DynamicTableSource just become a
Source<*Row*,
SourceSplitT, EnumChkT>?

Can you help me understand a bit more about the reason we need the
following relational representation / wrapper interfaces v.s. the
interfaces that we could put to the Source in FLIP-27?

DynamicTableSource v.s. Source<Row, SourceSplitT, EnumChkT>
SupportsFilterablePushDown v.s. FilterableSource
SupportsProjectablePushDown v.s. ProjectableSource
SupportsWatermarkPushDown v.s. WithWatermarkAssigner
SupportsComputedColumnPushDown v.s. ComputedColumnDeserializer
ScanTableSource v.s. ChangeLogDeserializer.
LookUpTableSource v.s. LookUpSource

Assuming we have all the interfaces on the right side, do we still

need


the

interfaces on the left side? Note that the interfaces on the right

can


be

used by both DataStream and Table. If we do this, there will only

be


one

set of Source interfaces Table and DataStream, the only difference

is


that

the Source for table will have some specific plugins and

configurations.

An

omnipotent Source can implement all the the above interfaces and

take a

Deserializer that implements both ComputedColumnDeserializer and
ChangeLogDeserializer.

Would the SQL planner work with that?

Thanks,

Jiangjie (Becket) Qin






On Tue, Mar 24, 2020 at 5:03 PM Jingsong Li <

jingsongl...@gmail.com>

<

jingsongl...@gmail.com>

wrote:


+1. Thanks Timo for the design doc.

We can also consider @Experimental too. But I am +1 to

@PublicEvolving,

we

should be confident in the current change.

Best,
Jingsong Lee

On Tue, Mar 24, 2020 at 4:30 PM Timo Walther <twal...@apache.org>

<

twal...@apache.org>


wrote:

@Becket: We totally agree that we don't need table specific

connectors

during runtime. As Dawid said, the interfaces proposed here are

just

for

communication with the planner. Once the properties (watermarks,
computed column, filters, projecttion etc.) are negotiated, we can
configure a regular Flink connector.

E.g. setting the watermark assigner and deserialization schema of a
Kafka connector.

For better separation of concerns, Flink connectors should not

include

relational interfaces and depend on flink-table. This is the
responsibility of table source/sink.

@Kurt: I would like to mark them @PublicEvolving already because we

need

to deprecate the old interfaces as early as possible. We cannot

redirect

to @Internal interfaces. They are not marked @Public, so we can

still

evolve them. But a core design shift should not happen again, it

would

leave a bad impression if we are redesign over and over again.

Instead

we should be confident in the current change.

Regards,
Timo


On 24.03.20 09:20, Dawid Wysakowicz wrote:

Hi Becket,

Answering your question, we have the same intention not to

duplicate

connectors between datastream and table apis. The interfaces

proposed

in

the FLIP are a way to describe relational properties of a source.

The

intention is as you described to translate all of those expressed

as

expressions or other Table specific structures into a DataStream

source.

In other words I think what we are doing here is in line with

what

you

described.

Best,

Dawid

On 24/03/2020 02:23, Becket Qin wrote:

Hi Timo,

Thanks for the proposal. I completely agree that the current

Table

connectors could be simplified quite a bit. I haven't finished

reading

everything, but here are some quick thoughts.

Actually to me the biggest question is why should there be two

different

connector systems for DataStream and Table? What is the

fundamental

reason

that is preventing us from merging them to one?

The basic functionality of a connector is to provide

capabilities

to

do

IO

and Serde. Conceptually, Table connectors should just be

DataStream

connectors that are dealing with Rows. It seems that quite a few

of

the

special connector requirements are just a specific way to do IO

/

Serde.

Taking SupportsFilterPushDown as an example, imagine we have the

following

interface:

interface FilterableSource<PREDICATE> {
      void applyFilterable(Supplier<PREDICATE> predicate);
}

And if a ParquetSource would like to support filterable, it will

become:

class ParquetSource implements Source,

FilterableSource(FilterPredicate> {

      ...
}

For Table, one just need to provide an predicate supplier that

converts

an

Expression to the specified predicate type. This has a few

benefit:

1. Same unified API for filterable for sources, regardless of

DataStream or

Table.
2. The  DataStream users now can also use the

ExpressionToPredicate

supplier if they want to.

To summarize, my main point is that I am wondering if it is

possible

to

have a single set of connector interface for both Table and

DataStream,

rather than having two hierarchies. I am not 100% sure if this

would

work,

but if it works, this would be a huge win from both code

maintenance

and

user experience perspective.

Thanks,

Jiangjie (Becket) Qin



On Tue, Mar 24, 2020 at 2:03 AM Dawid Wysakowicz <

dwysakow...@apache.org>

wrote:


Hi Timo,

Thank you for the proposal. I think it is an important

improvement

that

will benefit many parts of the Table API. The proposal looks

really

good

to me and personally I would be comfortable with voting on the

current

state.

Best,

Dawid

On 23/03/2020 18:53, Timo Walther wrote:

Hi everyone,

I received some questions around how the new interfaces play

together

with formats and their factories.

Furthermore, for MySQL or Postgres CDC logs, the format should

be

able

to return a `ChangelogMode`.

Also, I incorporated the feedback around the factory design in

general.

I added a new section `Factory Interfaces` to the design

document.

This should be helpful to understand the big picture and

connecting

the concepts.

Please let me know what you think?

Thanks,
Timo


On 18.03.20 13:43, Timo Walther wrote:

Hi Benchao,

this is a very good question. I will update the FLIP about

this.

The legacy planner will not support the new interfaces. It

will

only

support the old interfaces. With the next release, I think

the

Blink

planner is stable enough to be the default one as well.

Regards,
Timo

On 18.03.20 08:45, Benchao Li wrote:

Hi Timo,

Thank you and others for the efforts to prepare this FLIP.

The FLIP LGTM generally.

+1 for moving blink data structures to table-common, it's

useful

to

udf too
in the future.
A little question is, do we plan to support the new

interfaces

and

data

types in legacy planner?
Or we only plan to support these new interfaces in blink

planner.

And using primary keys from DDL instead of derived key

information

from

each query is also a good idea,
we met some use cases where this does not works very well

before.

This FLIP also makes the dependencies of table modules more

clear, I

like
it very much.

Timo Walther <twal...@apache.org> <twal...@apache.org>

于2020年3月17日周二

上午1:36写道：



Hi everyone,

I'm happy to present the results of long discussions that

we

had

internally. Jark, Dawid, Aljoscha, Kurt, Jingsong, me, and

many

more

have contributed to this design document.

We would like to propose new long-term table source and

table

sink

interfaces:












https://cwiki.apache.org/confluence/display/FLINK/FLIP-95%3A+New+TableSource+and+TableSink+interfaces


This is a requirement for FLIP-105 and finalizing FLIP-32.

The goals of this FLIP are:

- Simplify the current interface architecture:
        - Merge upsert, retract, and append sinks.
        - Unify batch and streaming sources.
        - Unify batch and streaming sinks.

- Allow sources to produce a changelog:
        - UpsertTableSources have been requested a lot by

users.

Now

is the
time to open the internal planner capabilities via the new

interfaces.

        - According to FLIP-105, we would like to support

changelogs for

processing formats such as Debezium.

- Don't rely on DataStream API for source and sinks:
        - According to FLIP-32, the Table API and SQL should

be

independent
of the DataStream API which is why the `table-common`

module

has

no

dependencies on `flink-streaming-java`.
        - Source and sink implementations should only depend

on

the

`table-common` module after FLIP-27.
        - Until FLIP-27 is ready, we still put most of the

interfaces in

`table-common` and strictly separate interfaces that

communicate

with a
planner and actual runtime reader/writers.

- Implement efficient sources and sinks without planner

dependencies:

        - Make Blink's internal data structures available to

connectors.

        - Introduce stable interfaces for data structures

that

can

be

marked as `@PublicEvolving`.
        - Only require dependencies on `flink-table-common`

in

the

future

It finalizes the concept of dynamic tables and consideres

how

all

source/sink related classes play together.

We look forward to your feedback.

Regards,
Timo


--
Best, Jingsong Lee












































					Reply via email to