https://github.com/apache/flume/blob/trunk/flume-ng-core/src/main/java/org/apache/flume/source/SpoolDirectorySource.java
On Tue, Jan 22, 2013 at 9:23 PM, Roshan Naik <[email protected]> wrote: > Mike, > Where is the SpoolingFileSource that you are referring to ? > > -roshan > > > On Tue, Jan 22, 2013 at 6:39 PM, Mike Percy <[email protected]> wrote: > >> Hi Roshan, >> Yep in general I'd have concerns w.r.t. capacity planning and garbage >> collector behavior for large events. Flume holds at least one event batch >> in memory at once, depending on # of sources/sinks, and even with a batch >> size of 1 if you have unpredictably large events there is nothing >> preventing an OutOfMemoryError in extreme cases. But if you plan for >> capacity and test thoroughly then it can be made to work. >> >> Regards, >> Mike >> >> >> On Tue, Jan 22, 2013 at 3:38 PM, Roshan Naik <[email protected]>wrote: >> >>> i recall some discussion with regards to being cautious on the size of >>> the events (in this case the file being moved) as flume is not quite >>> intended for large events. Mike perhaps you can throw some light on that >>> aspect ? >>> >>> >>> On Tue, Jan 22, 2013 at 12:17 AM, Mike Percy <[email protected]> wrote: >>> >>>> Check out the latest changes to SpoolingFileSource w.r.t. >>>> EventDeserializers on trunk. You can deserialize a whole file that way if >>>> you want. Whether that is a good idea depends on your use case, though. >>>> >>>> It's on trunk, lacking user docs for the latest changes but I will try >>>> to hammer out updated docs soon. In the meantime, you can just look at the >>>> code and read the comments. >>>> >>>> Regards, >>>> Mike >>>> >>>> On Monday, January 21, 2013, Nitin Pawar wrote: >>>> >>>>> you cant configure it to send the entire file in an event unless you >>>>> have fixed number of events in each of the files. basically it reads the >>>>> entire file into a channel and then starts writing. >>>>> >>>>> so as long as you can limit the events in the file, i think you can >>>>> send entire file as a transaction but not as a single event >>>>> as long as I understand flume treats individual lines in the file as >>>>> event >>>>> >>>>> if you want to pull the entire file then you may want to implement >>>>> that with messaging queues where you send an event to activemq queue and >>>>> then your consumer may pull the file in one transaction with some other >>>>> mechanism like ftp or scp or something like that >>>>> >>>>> others will have better idea, i am just suggesting a crude way to get >>>>> the entire file as a single event >>>>> >>>>> >>>>> On Tue, Jan 22, 2013 at 12:19 PM, Henry Ma <[email protected]>wrote: >>>>> >>>>>> As far as I know, Directory Spooling Source will send the file line >>>>>> by line as an event, and File Roll Sink will receive these lines and roll >>>>>> up to a big file by a fixed interval. Is it right, and can we config it >>>>>> to >>>>>> send the whole file as an event? >>>>>> >>>>>> >>>>>> On Tue, Jan 22, 2013 at 1:22 PM, Nitin Pawar <[email protected] >>>>>> > wrote: >>>>>> >>>>>>> why don't you use directory spooling ? >>>>>>> >>>>>>> >>>>>>> On Tue, Jan 22, 2013 at 7:15 AM, Henry Ma >>>>>>> <[email protected]>wrote: >>>>>>> >>>>>>>> Hi, >>>>>>>> >>>>>>>> When using Flume to collect log files, we want to just COPY the >>>>>>>> original files from several servers to a central storage (unix file >>>>>>>> system), not to roll up to a big file. Because we must record some >>>>>>>> messages >>>>>>>> of the original file such as name, host, path, timestamp, etc. >>>>>>>> Besides, we >>>>>>>> want to guarantee total reliability: no file miss, no file >>>>>>>> reduplicated. >>>>>>>> >>>>>>>> It seems that, in Source, we must put a whole file (size may be >>>>>>>> between 100KB and 100MB) into a Flume event; and in Sink, we must write >>>>>>>> each event to a single file. >>>>>>>> >>>>>>>> Is it practicable? Thanks! >>>>>>>> >>>>>>>> -- >>>>>>>> Best Regards, >>>>>>>> Henry Ma >>>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> -- >>>>>>> Nitin Pawar >>>>>>> >>>>>> >>>>>> >>>>>> >>>>>> -- >>>>>> Best Regards, >>>>>> Henry Ma >>>>>> >>>>> >>>>> >>>>> >>>>> -- >>>>> Nitin Pawar >>>>> >>>> >>> >> >
