Thanks, Fabian. I just want to check one thing again. As you said, [Distribute By] is partitionByHash(). and [Sort By] should be sortGroup on Flink. However, [Cluster By] is consist of partitionByHash(). *sortPartition()*.
As far as I know, [Cluster By] is same as the combination with [Distribute By] + [Sort By]. Therefore, according to your suggestion, should it be partitionByHash() + sortGroup() instead of sortPartition() ? Or probably I did not still get much difference between Partition and scope within a reduce. Regards, Philip On Mon, Oct 19, 2015 at 2:17 PM, Fabian Hueske <fhue...@gmail.com> wrote: > Hi Philip, > > here a few additions to what Max said: > - ORDER BY: As Max said, Flink's sortPartition() does only sort with a > partition and does not produce a total order. You can either set the > parallelism to 1 as Max suggested or use a custom partitioner to range > partition the data. > - SORT BY: From your description, the semantics are not 100% clear. If > SORT BY refers to the order of tuples WITHIN a reduce function call, it > should be groupBy().sortGroup() in Flink instead of sortPartition > - DISTRIBUTE BY: This should be partitionByHash() instead of groupBy(). > GroupBy() will also sort the data which is not required for DISTRIBUTE BY. > - CLUSTER BY: This should be partitionByHash().sortPartition(). > - Reduce vs. GroupReduce: A ReduceFunction is always combinable. This is > optional for GroupReduceFunctions. > > Cheers, Fabian > > > > 2015-10-19 13:01 GMT+02:00 Maximilian Michels <m...@apache.org>: > >> Hi Philip, >> >> Thank you for your questions. I think you have mapped the HIVE >> functions to the Flink ones correctly. Just a remark on the ORDER BY. >> You wrote that it produces a total order of all the records. In this >> case, you'd have do a SortPartition operation with parallelism set to >> 1. This is necessary because we need to have all records in one place >> to perform a sort on them. >> >> Considering your reduce question: There is no fundamental >> advantage/disadvantage of using GroupReduce over Reduce. It depends on >> your use case which one is more convenient or efficient. For the >> regular reduce, you just get two elements and produce one. You can't >> easily keep state between the reduces other than in the value itself. >> The GroupReduce, on the other hand, may produce none, one, or multiple >> elements per grouping and keep state in between emitting values. Thus, >> GroupReduce is a more powerful operator and can be seen as a superset >> of the Reduce operator. I would advise you to use the one you find >> easiest to use. >> >> Best regards, >> Max >> >> On Sun, Oct 18, 2015 at 9:16 PM, Philip Lee <philjj...@gmail.com> wrote: >> > Hi, Flink people, a question about translation from HIVE Query to Flink >> > fucntioin by using Table API. In sum up, I am working on some benchmark >> for >> > flink >> > >> > I am Philip Lee majoring in Computer Science in Master Degree of TUB. , >> I >> > work on translation from Hive Query of Benchmark to Flink codes. >> > >> > As I stuided it, I have a few of questions. >> > >> > First of all, if there are people who do no know Hive functions, let me >> > briefly explan. >> > >> > ORDER BY: it just guarntees total order in the output. >> > SORT BY: it only guarntess ordering of the rows within a reducer. >> > GROUP BY: this is just groupBy function in SQL. >> > DISTRIBUTE BY: all rows with the same distributed by columns will go to >> the >> > same reducer. >> > CLUSTER BY: this is just consisted of Distribute By the same column + >> Sort >> > By the same column. >> > >> > I just want to check that the flink functions I use are equal to Hive >> one. >> > < Hive SQL Query = Flink functions > >> > >> > ORDER BY = sortPartition(,) >> > SORT BY= groupBy(`col).sortPartition(,) >> > GROUP BY: this is just groupBy function. >> > DISTRIBUTE BY = groupBy(`col) >> > CLUSTER BY = groupBy(`col).sortPartition(,) >> > >> > I do not see much difference between groupBy and distributed by if I >> apply >> > it to flink function. >> > If this is hadoop version, we could say mapper is distribute by on >> hadoop. >> > However, I am not much sure what could be DISTRIBUTE BY on flink. I >> tried to >> > guess groupBy on Flink could be the function which is to distribute the >> rows >> > by the specified key. >> > >> > Please feel free to correct what I suggested. >> > >> > >> > Secondly, I just want to make sure the difference between reduce >> function >> > and reduceGroup. I guess there must be a trade-off between two >> functinos. I >> > know reduceGroup is invoked with an Iterator, but which case is more >> proper >> > and benifical to use reduceGroup function rather than reduce function? >> > >> > Best Regards, >> > Philip >> > >> > -- >> > >> > ========================================================== >> > >> > Hae Joon Lee >> > >> > >> > Now, in Germany, >> > >> > M.S. Candidate, Interested in Distributed System, Iterative Processing >> > >> > Dept. of Computer Science, Informatik in German, TUB >> > >> > Technical University of Berlin >> > >> > >> > In Korea, >> > >> > M.S. Candidate, Computer Architecture Laboratory >> > >> > Dept. of Computer Science, KAIST >> > >> > >> > Rm# 4414 CS Dept. KAIST >> > >> > 373-1 Guseong-dong, Yuseong-gu, Daejon, South Korea (305-701) >> > >> > >> > Mobile) 49) 015-251-448-278 in Germany, no cellular in Korea >> > >> > ========================================================== >> > > -- ========================================================== *Hae Joon Lee* Now, in Germany, M.S. Candidate, Interested in Distributed System, Iterative Processing Dept. of Computer Science, Informatik in German, TUB Technical University of Berlin In Korea, M.S. Candidate, Computer Architecture Laboratory Dept. of Computer Science, KAIST Rm# 4414 CS Dept. KAIST 373-1 Guseong-dong, Yuseong-gu, Daejon, South Korea (305-701) Mobile) 49) 015-251-448-278 in Germany, no cellular in Korea ==========================================================
