Jeff, Joao, Thanks for the pointer! I think i am getting close here: 1. set up a cluster of flume agent with redundancies, source as avro, sink as HDFS. 2 use storm(not quite necessary) to read from our socket server, then in the bolt, using flume client (load balancing rpc client) to send the event to the agent set up in step 1.
Then I thus get all the benefit of storm and flume. Does this set up look right to you? thank you very much, Chen On Thu, Jan 9, 2014 at 8:58 PM, Joao Salcedo <[email protected]> wrote: > Hi Chen, > > Maybe it would be worth checking this > http://flume.apache.org/FlumeDeveloperGuide.html#loadbalancing-rpc-client > > Regards, > > Joao > > > On Fri, Jan 10, 2014 at 3:50 PM, Jeff Lord <[email protected]> wrote: > >> Have you taken a look at the load balancing rpc client? >> >> >> On Thu, Jan 9, 2014 at 8:43 PM, Chen Wang <[email protected]>wrote: >> >>> Jeff, >>> I have read this ppt at the beginning, but didn't find solution to my >>> user case. To simplify my case, I only have 1 data source(composed of 5 >>> socket server) and i am looking for a fault tolerant deployment of flume, >>> that can read from this single data source and sink to hdfs in fault >>> tolerant mode: when one node dies, another flume node can pick up and >>> continue; >>> Thanks, >>> Chen >>> >>> >>> On Thu, Jan 9, 2014 at 7:49 PM, Jeff Lord <[email protected]> wrote: >>> >>>> Chen, >>>> >>>> Have you taken a look at this presentation on Planning and Deploying >>>> Flume from ApacheCon? >>>> >>>> >>>> http://archive.apachecon.com/na2013/presentations/27-Wednesday/Big_Data/11:45-Mastering_Sqoop_for_Data_Transfer_for_Big_Data-Arvind_Prabhakar/Arvind%20Prabhakar%20-%20Planning%20and%20Deploying%20Apache%20Flume.pdf >>>> >>>> It may have the answers you need. >>>> >>>> Best, >>>> >>>> Jeff >>>> >>>> >>>> On Thu, Jan 9, 2014 at 7:24 PM, Chen Wang >>>> <[email protected]>wrote: >>>> >>>>> Thanks Saurabh. >>>>> If that is the case, I am actually thinking about using storm spout to >>>>> talk to our socket server so that the storm cluster can take care of the >>>>> reading socket server part. Then in each storm node, start a flume agent, >>>>> listening on a RPC port and write to HDFS(with fail over) .Then in the >>>>> storm bolt, simply send the data to RPC so that flume can get it. >>>>> How do you think of this setup? It takes care of both failover on the >>>>> source(by storm) and on the sink(by flume) But It looks a little >>>>> complicated for me. >>>>> Chen >>>>> >>>>> >>>>> On Thu, Jan 9, 2014 at 7:18 PM, Saurabh B >>>>> <[email protected]>wrote: >>>>> >>>>>> Hi Chen, >>>>>> >>>>>> I think Flume doesn't have a way to configure multiple sources >>>>>> pointing to same data source. Of course you can do that, but you will end >>>>>> up with duplicate data. Flume offers fail over at the sink level. >>>>>> >>>>>> On Thu, Jan 9, 2014 at 6:56 PM, Chen Wang <[email protected] >>>>>> > wrote: >>>>>> >>>>>>> Ok. so after more researching:) It seems that what i need is the >>>>>>> failover for agent source, (not fail over for sink): >>>>>>> If one agent dies, another same kind of agent will start running. >>>>>>> Does flume support this scenario? >>>>>>> Thanks, >>>>>>> Chen >>>>>>> >>>>>>> >>>>>>> On Thu, Jan 9, 2014 at 3:12 PM, Chen Wang < >>>>>>> [email protected]> wrote: >>>>>>> >>>>>>>> After reading more docs, it seems that if I want to achieve my >>>>>>>> goal, i have to do the following: >>>>>>>> 1. Having one agent with the custom source running on one node. >>>>>>>> This agent reads from those 5 socket server, and sink to some kind of >>>>>>>> sink(maybe another socket?) >>>>>>>> 2. On another(or more) machines, setting up collectors that read >>>>>>>> from the agent sink in 1, and sink to hdfs. >>>>>>>> 3. Having a master node managing nodes in 1,2. >>>>>>>> >>>>>>>> But it seems to be overskilled in my case: in 1, i can already sink >>>>>>>> to hdfs. Since the data available at socket server are much faster >>>>>>>> than the >>>>>>>> data translation part. I want to be able to later add more nodes to >>>>>>>> do the >>>>>>>> translation job. so what is the correct setup? >>>>>>>> Thanks, >>>>>>>> Chen >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> On Thu, Jan 9, 2014 at 2:38 PM, Chen Wang < >>>>>>>> [email protected]> wrote: >>>>>>>> >>>>>>>>> Guys, >>>>>>>>> In my environment, the client is 5 socket servers. Thus i wrote a >>>>>>>>> custom source spawning 5 threads reading from each of them >>>>>>>>> infinitely,and >>>>>>>>> the sink is hdfs(hive table). The work fine by running flume-ng >>>>>>>>> agent. >>>>>>>>> >>>>>>>>> But how can i deploy this in distributed mode(cluster)? I am >>>>>>>>> confused about the 3 ties(agent,collector,storage) mentioned in the >>>>>>>>> doc. >>>>>>>>> Does it apply to my case? How can I separate my agent/collect/storage? >>>>>>>>> Apparently i can only have one agent running: multiple agent will >>>>>>>>> result in >>>>>>>>> getting duplicates from the socket server. But I want that if one >>>>>>>>> agent >>>>>>>>> dies, other agent can take it up. I would also like to be able to add >>>>>>>>> horizontal scalability for writing to hdfs. How can I achieve all >>>>>>>>> this? >>>>>>>>> >>>>>>>>> thank you very much for your advice. >>>>>>>>> Chen >>>>>>>>> >>>>>>>> >>>>>>>> >>>>>>> >>>>>> >>>>>> >>>>>> -- >>>>>> Mailing List Archives, >>>>>> QnaList.com >>>>>> >>>>> >>>>> >>>> >>> >> >
