Hi Jincheng, Hequn & Jingsong, Thanks a lot for your suggestions. I have created FLIP-97[1] for this feature.
> One little suggestion: maybe it would be nice if we can add some performance explanation in the document? (I just very curious:)) Thanks for the suggestion. I have updated the design doc in the "BackGround" section about where the performance gains could be got from. > It seems that a batch should always in a bundle. Bundle size should always bigger than batch size. (if a batch can not cross bundle). Can you explain this relationship to the document? I have updated the design doc explaining more about these two configurations. > In the batch world, vectorization batch size is about 1024+. What do you think about the default value of "batch"? Is there any link about where this value comes from? I have performed a simple test for Pandas UDF which performs the simple +1 operation. The performance is best when the batch size is set to 5000. I think it depends on the data type of each column, the functionality the Pandas UDF does, etc. However I agree with you that we could give a meaningful default value for the "batch" size which works in most scenarios. > Can we only configure one parameter and calculate another automatically? For example, if we just want to "pipeline", "bundle.size" is twice as much as "batch.size", is this work? I agree with Jincheng that this is not feasible. I think that giving an meaningful default value for the "batch.size" which works in most scenarios is enough. What's your thought? Thanks, Dian [1] https://cwiki.apache.org/confluence/display/FLINK/FLIP-97%3A+Support+Scalar+Vectorized+Python+UDF+in+PyFlink On Mon, Feb 10, 2020 at 4:25 PM jincheng sun <sunjincheng...@gmail.com> wrote: > Hi Jingsong, > > Thanks for your feedback! I would like to share my thoughts regarding the > follows question: > > >> - Can we only configure one parameter and calculate another > automatically? For example, if we just want to "pipeline", "bundle.size" is > twice as much as "batch.size", is this work? > > I don't think this works. These two configurations are used for different > purposes and there is no direct relationship between them and so I guess we > cannot infer a configuration from the other configuration. > > Best, > Jincheng > > > Jingsong Li <jingsongl...@gmail.com> 于2020年2月10日周一 下午1:53写道： > > > Thanks Dian for your reply. > > > > +1 to create a FLIP too. > > > > About "python.fn-execution.bundle.size" and > > "python.fn-execution.arrow.batch.size", I got what are you mean about > > "pipeline". I agree. > > It seems that a batch should always in a bundle. Bundle size should > always > > bigger than batch size. (if a batch can not cross bundle). > > Can you explain this relationship to the document? > > > > I think default value is a very important thing, we can discuss: > > - In the batch world, vectorization batch size is about 1024+. What do > you > > think about the default value of "batch"? > > - Can we only configure one parameter and calculate another > automatically? > > For example, if we just want to "pipeline", "bundle.size" is twice as > much > > as "batch.size", is this work? > > > > Best, > > Jingsong Lee > > > > On Mon, Feb 10, 2020 at 11:55 AM Hequn Cheng <he...@apache.org> wrote: > > > > > Hi Dian, > > > > > > Thanks a lot for bringing up the discussion! > > > > > > It is great to see the Pandas UDFs feature is going to be introduced. I > > > think this would improve the performance and also the usability of > > > user-defined functions (UDFs) in Python. > > > One little suggestion: maybe it would be nice if we can add some > > > performance explanation in the document? (I just very curious:)) > > > > > > +1 to create a FLIP for this big enhancement. > > > > > > Best, > > > Hequn > > > > > > On Mon, Feb 10, 2020 at 11:15 AM jincheng sun < > sunjincheng...@gmail.com> > > > wrote: > > > > > > > Hi Dian, > > > > > > > > Thanks for bring up this discussion. This is very important for the > > > > ecological of PyFlink. Add support Pandas greatly enriches the > > available > > > > UDF library of PyFlink and greatly improves the usability of PyFlink! > > > > > > > > +1 for Support scalar vectorized Python UDF. > > > > > > > > I think we should to create a FLIP for this big enhancements. :) > > > > > > > > What do you think? > > > > > > > > Best, > > > > Jincheng > > > > > > > > > > > > > > > > dianfu <dia...@apache.org> 于2020年2月5日周三 下午6:01写道： > > > > > > > > > Hi Jingsong, > > > > > > > > > > Thanks a lot for the valuable feedback. > > > > > > > > > > 1. The configurations "python.fn-execution.bundle.size" and > > > > > "python.fn-execution.arrow.batch.size" are used for separate > purposes > > > > and I > > > > > think they are both needed. If they are unified, the Python > operator > > > has > > > > to > > > > > wait the execution results of the previous batch of elements before > > > > > processing the next batch. This means that the Python UDF execution > > can > > > > not > > > > > be pipelined between batches. With separate configuration, there > will > > > be > > > > no > > > > > such problems. > > > > > 2. It means that the Java operator will convert input elements to > > Arrow > > > > > memory format and then send them to the Python worker, vice verse. > > > > > Regarding to the zero-copy benefits provided by Arrow, we can gain > > them > > > > > automatically using Arrow. > > > > > 3. Good point! As all the classes of Python module is written in > Java > > > and > > > > > it's not suggested to introduce new Scala classes, so I guess it's > > not > > > > easy > > > > > to do so right now. But I think this is definitely a good > improvement > > > we > > > > > can do in the future. > > > > > 4. You're right and we will add a series of Arrow ColumnVectors for > > > each > > > > > type supported. > > > > > > > > > > Thanks, > > > > > Dian > > > > > > > > > > > 在 2020年2月5日，下午4:57，Jingsong Li <jingsongl...@gmail.com> 写道： > > > > > > > > > > > > Hi Dian, > > > > > > > > > > > > +1 for this, thanks driving. > > > > > > Documentation looks very good. I can imagine a huge performance > > > > > improvement > > > > > > and better integration to other Python libraries. > > > > > > > > > > > > A few thoughts: > > > > > > - About data split: "python.fn-execution.arrow.batch.size", can > we > > > > unify > > > > > it > > > > > > with "python.fn-execution.bundle.size"? > > > > > > - Use of Apache Arrow as the exchange format: Do you mean Arrow > > > support > > > > > > zero-copy between Java and Python? > > > > > > - ArrowFieldWriter seems we can implement it by code generation. > > But > > > it > > > > > is > > > > > > OK to initial version with virtual function call. > > > > > > - ColumnarRow for vectorization reading seems that we need > > implement > > > > > > ArrowColumnVectors. > > > > > > > > > > > > Best, > > > > > > Jingsong Lee > > > > > > > > > > > > On Wed, Feb 5, 2020 at 12:45 PM dianfu <dia...@apache.org> > wrote: > > > > > > > > > > > >> Hi all, > > > > > >> > > > > > >> Scalar Python UDF has already been supported in the coming > release > > > > 1.10 > > > > > >> (FLIP-58[1]). It operates one row at a time. It works in the way > > > that > > > > > the > > > > > >> Java operator serializes one input row to bytes and sends them > to > > > the > > > > > >> Python worker; the Python worker deserializes the input row and > > > > > evaluates > > > > > >> the Python UDF with it; the result row is serialized and sent > back > > > to > > > > > the > > > > > >> Java operator. > > > > > >> > > > > > >> It suffers from the following problems: > > > > > >> 1) High serialization/deserialization overhead > > > > > >> 2) It’s difficult to leverage the popular Python libraries used > by > > > > data > > > > > >> scientists, such as Pandas, Numpy, etc which provide high > > > performance > > > > > data > > > > > >> structure and functions. > > > > > >> > > > > > >> Jincheng and I have discussed offline and we want to introduce > > > > > vectorized > > > > > >> Python UDF to address the above problems. This feature has also > > been > > > > > >> mentioned in the discussion thread about the Python API plan[2]. > > For > > > > > >> vectorized Python UDF, a batch of rows are transferred between > JVM > > > and > > > > > >> Python VM in columnar format. The batch of rows will be > converted > > > to a > > > > > >> collection of Pandas.Series and given to the vectorized Python > UDF > > > > which > > > > > >> could then leverage the popular Python libraries such as Pandas, > > > > Numpy, > > > > > etc > > > > > >> for the Python UDF implementation. > > > > > >> > > > > > >> Please refer the design doc[3] for more details and welcome any > > > > > feedback. > > > > > >> > > > > > >> Regards, > > > > > >> Dian > > > > > >> > > > > > >> [1] > > > > > >> > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-58%3A+Flink+Python+User-Defined+Stateless+Function+for+Table > > > > > >> [2] > > > > > >> > > > > > > > > > > > > > > > http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-What-parts-of-the-Python-API-should-we-focus-on-next-tt36119.html > > > > > >> [3] > > > > > >> > > > > > > > > > > > > > > > https://docs.google.com/document/d/1ls0mt96YV1LWPHfLOh8v7lpgh1KsCNx8B9RrW1ilnoE/edit#heading=h.qivada1u8hwd > > > > > >> > > > > > >> > > > > > > > > > > > > -- > > > > > > Best, Jingsong Lee > > > > > > > > > > > > > > > > > > > > > > > -- > > Best, Jingsong Lee > > >
