Hi Jinxin, Thanks for your suggestions, I will try to use foreachpartition later. Best regards, maqy
发件人: Tang Jinxin 发送时间: 2020年4月23日 7:31 收件人: maqy 抄送: Andrew Melo; user@spark.apache.org 主题: 回复：Can I collect Dataset[Row] to driver without converting it toArray [Row]? Hi maqy, Thanks for your question.Through consideration,I have some ideas as follow:firstly,try not collect to driver if not nessessary,instead (use foreachpartition)send data from ececutors;secondly,if not use some high performance ser/deser like kryo, we could have a try.As a summary,I recommend the first point(desigh a more efficient way when the data is too large). Best wishes, littlestar xiaoxingstack 邮箱：xiaoxingst...@gmail.com 签名由 网易邮箱大师 定制 在2020年04月22日 23:24，maqy 写道： Hi Andrew, Thank you for your reply, I am using the scala api of spark, and the tensorflow machine is not in the spark cluster. Is this JIRA / PR still valid in this situation? In addition, the current bottleneck of the application is that the amount of data transferred through the network(use collect()) is too large, and the deserialization seems to take some time. Best wishes, maqy 发件人: Andrew Melo 发送时间: 2020年4月22日 21:02 收件人: maqy 抄送: Michael Artz; user@spark.apache.org 主题: Re: Can I collect Dataset[Row] to driver without converting it toArray [Row]? On Wed, Apr 22, 2020 at 3:24 AM maqy <454618...@qq.com> wrote: > > I will traverse this Dataset to convert it to Arrow and send it to Tensorflow > through Socket. (I presume you're using the python tensorflow API, if you're not, please ignore) There is a JIRA/PR ([1] [2]) which proposes to add native support for Arrow serialization to python, Under the hood, Spark is already serializing into Arrow format to transmit to python, it's just additionally doing an unconditional conversion to pandas once it reaches the python runner -- which is good if you're using pandas, not so great if pandas isn't what you operate on. The JIRA above would let you receive the arrow buffers (that already exist) directly. Cheers, Andrew [1] https://issues.apache.org/jira/browse/SPARK-30153 [2] https://github.com/apache/spark/pull/26783 > > I tried to use toLocalIterator() to traverse the dataset instead of collect > to the driver, but toLocalIterator() will create a lot of jobs and will bring > a lot of time consumption. > > > > Best regards, > > maqy > > > > 发件人: Michael Artz > 发送时间: 2020年4月22日 16:09 > 收件人: maqy > 抄送: user@spark.apache.org > 主题: Re: Can I collect Dataset[Row] to driver without converting it to Array > [Row]? > > > > What would you do with it once you get it into driver in a Dataset[Row]? > > Sent from my iPhone > > > > On Apr 22, 2020, at 3:06 AM, maqy <454618...@qq.com> wrote: > > > > When the data is stored in the Dataset [Row] format, the memory usage is very > small. > > When I use collect () to collect data to the driver, each line of the dataset > will be converted to Row and stored in an array, which will bring great > memory overhead. > > So, can I collect Dataset[Row] to driver and keep its data format? > > > > Best regards, > > maqy > > > > --------------------------------------------------------------------- To unsubscribe e-mail: user-unsubscr...@spark.apache.org r
