I've naively tried few other key types, it seems to be unrelated to key type.
As for now I have two workarounds and ignorance: 1. If there is one dominant dataset and other datasets are small (size << GB) then I use SideInput. 2. If I have multiple datasets of similar size I enclose it in a common container, flatten it and GroupByKey. 3. I measure occurrences and ignore the bug for now. Do you have an idea how a test for this may be constructed? It seems handy, I think. I also found two things, maybe they help you: 1. issue doesn't appear without parallelism 2. issue doesn't appear with a tiny datasets 2016-05-30 17:13 GMT+02:00 Aljoscha Krettek <[email protected]>: > You're right. I'm still looking into this, unfortunately I haven't made > progress so far. I'll keep you posted. > > On Sun, 29 May 2016 at 18:20 Pawel Szczur <[email protected]> wrote: > >> Hi, >> >> I used the config as in the repo. >> Please grep the the log for "hereGoesLongStringID0,2", you will see that >> this key is processed multiple times. >> >> This is how I understand CoGroupByKey: one has two (or more) >> PCollection<KV<K,?>>. Both sets are grouped by key. For each unique key a >> KV<K, CoGbkResult> is produced, a given CoGbkResult contains all values >> from all input PCollections which have the given key. >> >> But from the log it seems that each key produced more than one >> CoGbkResult. >> >> The final counters didn't catch the bug because in your case, the value >> from dataset1 was replicated for each key. >> >> Cheers, Pawel >> >> 2016-05-29 15:59 GMT+02:00 Aljoscha Krettek <[email protected]>: >> >>> Hi, >>> I ran your data generator with these configs: >>> p.apply(Create.of(new Config(3, 5, 600_000, 1))) >>> .apply(ParDo.of(new Generator())).apply( >>> AvroIO.Write.to >>> ("/tmp/dataset1").withSchema(DumbData.class).withNumShards(6)); >>> >>> p.apply(Create.of(new Config(3, 5, 600_000, 2))). >>> apply(ParDo.of(new Generator())).apply( >>> AvroIO.Write.to >>> ("/tmp/dataset2").withSchema(DumbData.class).withNumShards(6)); >>> >>> Then I ran the job with parallelism=6. I couldn't reproduce the problem, >>> this is the log file from one of several runs: >>> https://gist.github.com/aljoscha/ef1d804f57671cd472c75b92b4aee51b >>> >>> Could you please send me the exact config that you used. Btw, I ran it >>> inside an IDE, do the problems also occur in the IDE for you or only when >>> you execute on a cluster? >>> >>> Cheers, >>> Aljoscha >>> >>> On Sun, 29 May 2016 at 01:51 Pawel Szczur <[email protected]> wrote: >>> >>>> Hi Aljoscha. >>>> >>>> I've created a repo with fake dataset to allow easily reproduce the >>>> problem: >>>> https://github.com/orian/cogroup-wrong-grouping >>>> >>>> What I noticed: if the dataset is too small the bug doesn't appear. >>>> >>>> You can modify the size of dataset, but in ideal case it should be few >>>> hundred thousands records per key (I guess it depends on the machine you >>>> run it). >>>> >>>> Cheers, Pawel >>>> >>>> 2016-05-28 12:45 GMT+02:00 Aljoscha Krettek <[email protected]>: >>>> >>>>> Hi, >>>>> which version of Beam/Flink are you using. >>>>> >>>>> Could you maybe also provide example data and code that showcases the >>>>> problem? If you have concerns about sending it to a public list you can >>>>> also send it to me directly. >>>>> >>>>> Cheers, >>>>> Aljoscha >>>>> >>>>> On Fri, 27 May 2016 at 20:53 Pawel Szczur <[email protected]> >>>>> wrote: >>>>> >>>>>> *Data description.* >>>>>> >>>>>> I have two datasets. >>>>>> >>>>>> Records - the first, containes around 0.5-1M of records per >>>>>> (key,day). For testing I use 2-3 keys and 5-10 days of data. What I shoot >>>>>> for is 1000+ keys. Each record contains key, timestamp in μ-seconds and >>>>>> some other data. >>>>>> Configs - the second, is rather small. It describes the key in time, >>>>>> e.g. you can think about it as a list of tuples: (key, start date, end >>>>>> date, description). >>>>>> >>>>>> For the exploration I've encoded the data as files of length-prefixed >>>>>> Protocol Buffer binary encoded messages. Additionally the files are >>>>>> packed >>>>>> with gzip. Data is sharded by date. Each file is around 10MB. >>>>>> >>>>>> *Pipeline* >>>>>> >>>>>> First I add keys to both datasets. For Records dataset it's (key, day >>>>>> rounded timestamp). For Configs a key is (key, day), where day is each >>>>>> timestamp value between start date and end date (pointing midnight). >>>>>> The datasets are merged using CoGroupByKey. >>>>>> >>>>>> As a key type I use import org.apache.flink.api.java.tuple.Tuple2 >>>>>> with a Tuple2Coder from this repo. >>>>>> >>>>>> *The problem* >>>>>> >>>>>> If the Records dataset is tiny like 5 days, everything seems fine >>>>>> (check normal_run.log). >>>>>> >>>>>> INFO [main] (FlinkPipelineRunner.java:124) - Final aggregator values: >>>>>> INFO [main] (FlinkPipelineRunner.java:127) - item count : 4322332 >>>>>> INFO [main] (FlinkPipelineRunner.java:127) - missing val1 : 0 >>>>>> INFO [main] (FlinkPipelineRunner.java:127) - multiple val1 : 0 >>>>>> >>>>>> When I run the pipeline against 10+ days I encounter an error >>>>>> pointing that for some Records there's no Config (wrong_run.log). >>>>>> >>>>>> INFO [main] (FlinkPipelineRunner.java:124) - Final aggregator values: >>>>>> INFO [main] (FlinkPipelineRunner.java:127) - item count : 8577197 >>>>>> INFO [main] (FlinkPipelineRunner.java:127) - missing val1 : 6 >>>>>> INFO [main] (FlinkPipelineRunner.java:127) - multiple val1 : 0 >>>>>> >>>>>> Then I've added some extra logging messages: >>>>>> >>>>>> (ConvertToItem.java:144) - 68643 items for KeyValue3 on: >>>>>> 1462665600000000 >>>>>> (ConvertToItem.java:140) - no items for KeyValue3 on: 1463184000000000 >>>>>> (ConvertToItem.java:123) - missing for KeyValue3 on: 1462924800000000 >>>>>> (ConvertToItem.java:142) - 753707 items for KeyValue3 on: >>>>>> 1462924800000000 marked as no-loc >>>>>> (ConvertToItem.java:123) - missing for KeyValue3 on: 1462752000000000 >>>>>> (ConvertToItem.java:142) - 749901 items for KeyValue3 on: >>>>>> 1462752000000000 marked as no-loc >>>>>> (ConvertToItem.java:144) - 754578 items for KeyValue3 on: >>>>>> 1462406400000000 >>>>>> (ConvertToItem.java:144) - 751574 items for KeyValue3 on: >>>>>> 1463011200000000 >>>>>> (ConvertToItem.java:123) - missing for KeyValue3 on: 1462665600000000 >>>>>> (ConvertToItem.java:142) - 754758 items for KeyValue3 on: >>>>>> 1462665600000000 marked as no-loc >>>>>> (ConvertToItem.java:123) - missing for KeyValue3 on: 1463184000000000 >>>>>> (ConvertToItem.java:142) - 694372 items for KeyValue3 on: >>>>>> 1463184000000000 marked as no-loc >>>>>> >>>>>> You can spot that in first line 68643 items were processed for >>>>>> KeyValue3 and time 1462665600000000. >>>>>> Later on in line 9 it seems the operation processes the same key >>>>>> again, but it reports that no Config was available for these Records. >>>>>> The line 10 informs they've been marked as no-loc. >>>>>> >>>>>> The line 2 is saying that there were no items for KeyValue3 and time >>>>>> 1463184000000000, but in line 11 you can read that the items for this >>>>>> (key,day) pair were processed later and they've lacked a Config. >>>>>> >>>>>> *Work-around (after more testing, doesn't work, staying with Tuple2)* >>>>>> >>>>>> I've switched from using Tuple2 to a Protocol Buffer message: >>>>>> >>>>>> message KeyDay { >>>>>> optional ByteString key = 1; >>>>>> optional int64 timestamp_usec = 2; >>>>>> } >>>>>> >>>>>> But using Tuple2.of() was just easier than: >>>>>> KeyDay.newBuilder().setKey(...).setTimestampUsec(...).build(). >>>>>> >>>>>> // The original description comes from: >>>>>> http://stackoverflow.com/questions/37473682/items-not-groupped-correctly-cogroupbykey >>>>>> >>>>> >>>> >>
