My assumption is that you have something like log.dirs=/mnt/diska,/mnt/diskb,/mnt/diskc
Let's say that diskc goes bad. The expected behavior is that the kafka broker gets an error on write and kills itself. When that happens you have two options. Either you remove the drive and add a new one and remount it as /mnt/diskc (but with no data, just an empty directory). When the broker restarts it will pull down the data for the partitions that belong on diskc. Your other option, which you ask about, is to remove /mnt/diskc from the log.dirs list. If you do this when the broker comes online it will recreate the partitions it needs but will add them to /mnt/diska and /mnt/diskc (roughly round-robin). Hopefully that makes sense. -Jay On Thu, Aug 15, 2013 at 11:27 AM, Vadim Keylis <vkeylis2...@gmail.com>wrote: > Good morning Jay. When you say delete directory from the list, did you mean > from file system? Can I see through JMX which partitions online and which > one are not? > > Thanks > > > On Thu, Aug 15, 2013 at 10:04 AM, Jason Rosenberg <j...@squareup.com> > wrote: > > > Thanks Jay, I'll do some testing with this and report back. > > > > Jason > > > > > > On Thu, Aug 15, 2013 at 7:10 AM, Jay Kreps <jay.kr...@gmail.com> wrote: > > > > > I believe either should work. The broker has a record of what it should > > > have in zk and will recreate any missing logs. Try it to make sure > > though. > > > > > > Sent from my iPhone > > > > > > On Aug 15, 2013, at 12:52 AM, Jason Rosenberg <j...@squareup.com> > wrote: > > > > > > > Ok, that makes sense that the broker will shut itself down. > > > > > > > > If we bring it back up, can this be with an altered set of log.dirs? > > > Will > > > > the destroyed partitions get rebuilt on a new log.dir? Or do we have > > to > > > > bring it back up with a new or repaired disk, matching the old > log.dir, > > > in > > > > order for those replicas to be rebuilt? > > > > > > > > Jason > > > > > > > > > > > > On Wed, Aug 14, 2013 at 4:16 PM, Jay Kreps <jay.kr...@gmail.com> > > wrote: > > > > > > > >> If you get a disk error that results in an IOException the broker > will > > > >> shut itself down. You would then have the option of replacing the > disk > > > or > > > >> deleting that data directory from the list. When the broker is > brought > > > back > > > >> up the intact partitions will quickly catch up and be online; the > > > destroyed > > > >> partitions will have to fully rebuild off the other replicas and > will > > > take > > > >> a little longer but will automatically come back online once they > have > > > >> restored off the replicas. > > > >> > > > >> -jay > > > >> > > > >> Sent from my iPhone > > > >> > > > >> On Aug 14, 2013, at 1:49 PM, Jason Rosenberg <j...@squareup.com> > > wrote: > > > >> > > > >>> I'm getting ready to try out this configuration (use multiple > disks, > > no > > > >>> RAID, per broker). One concern is the procedure for recovering if > > > there > > > >> is > > > >>> a disk failure. > > > >>> > > > >>> If a disk fails, will the broker go offline, or will it continue > > > serving > > > >>> partitions on its remaining good disks? And if so, is there a > > > procedure > > > >>> for moving the partitions that were on the failed disk, but not > > > >> necessarily > > > >>> all the others on that broker? > > > >>> > > > >>> Jason > > > >>> > > > >>> > > > >>> On Thu, Jun 20, 2013 at 3:15 PM, Jason Rosenberg <j...@squareup.com > > > > > >> wrote: > > > >>> > > > >>>> yeah, that would work! > > > >>>> > > > >>>> > > > >>>> On Thu, Jun 20, 2013 at 1:20 PM, Jay Kreps <jay.kr...@gmail.com> > > > wrote: > > > >>>> > > > >>>>> Yeah we didn't go as far as adding weighting or anything like > > that--I > > > >>>>> think we'd be open to a patch that did that as long as it was > > > >>>>> optional. In the short term you can obviously add multiple > > > directories > > > >>>>> on the same disk to increase its share. > > > >>>>> > > > >>>>> -Jay > > > >>>>> > > > >>>>> On Thu, Jun 20, 2013 at 12:59 PM, Jason Rosenberg < > > j...@squareup.com> > > > >>>>> wrote: > > > >>>>>> This sounds like a great idea, to just disks as "just a bunch of > > > >> disks" > > > >>>>> or > > > >>>>>> JBOD.....hdfs works well this way. > > > >>>>>> > > > >>>>>> Do all the disks need to be the same size, to use them evenly? > > > Since > > > >> it > > > >>>>>> will allocate partitions randomly? > > > >>>>>> > > > >>>>>> It would be nice if you had 2 disks, with one twice as large as > > the > > > >>>>> other, > > > >>>>>> if the larger would be twice as likely to receive partitions as > > the > > > >>>>> smaller > > > >>>>>> one, etc. > > > >>>>>> > > > >>>>>> I suppose this goes into my earlier question to the list, > > vis-a-vis > > > >>>>>> heterogeneous brokers (e.g. utilize brokers with different sized > > > >>>>> storage, > > > >>>>>> using some sort of weighting scheme, etc.). > > > >>>>>> > > > >>>>>> Jason > > > >>>>>> > > > >>>>>> > > > >>>>>> On Thu, Jun 20, 2013 at 11:07 AM, Jay Kreps < > jay.kr...@gmail.com> > > > >>>>> wrote: > > > >>>>>> > > > >>>>>>> The intention is to allow the use of multiple disks without > RAID > > or > > > >>>>>>> logical volume management. We have found that there are a lot > of > > > >>>>>>> downsides to RAID--in particular a huge throughput hit. Since > we > > > >>>>>>> already have a parallelism model due to partitioning and a > fault > > > >>>>>>> tolerance model with replication RAID doesn't actually buy > much. > > > With > > > >>>>>>> this feature you can directly mount multiple disks as their own > > > >>>>>>> directory and the server will randomly assign partitions to > them. > > > >>>>>>> > > > >>>>>>> Obviously this will only work well if there are enough > > > >> high-throughput > > > >>>>>>> partitions to make load balance evenly (e.g. if you have only > one > > > big > > > >>>>>>> partition per server then this isn't going to work). > > > >>>>>>> > > > >>>>>>> -Jay > > > >>>>>>> > > > >>>>>>> On Wed, Jun 19, 2013 at 11:01 PM, Jason Rosenberg < > > > j...@squareup.com> > > > >>>>>>> wrote: > > > >>>>>>>> is it possible for a partition to have multiple replicas on > > > >> different > > > >>>>>>>> directories on the same broker? (hopefully no!) > > > >>>>>>>> > > > >>>>>>>> > > > >>>>>>>> On Wed, Jun 19, 2013 at 10:47 PM, Jun Rao <jun...@gmail.com> > > > wrote: > > > >>>>>>>> > > > >>>>>>>>> It takes a comma separated list and partition replicas are > > > randomly > > > >>>>>>>>> distributed to the list. > > > >>>>>>>>> > > > >>>>>>>>> Thanks, > > > >>>>>>>>> > > > >>>>>>>>> Jun > > > >>>>>>>>> > > > >>>>>>>>> > > > >>>>>>>>> On Wed, Jun 19, 2013 at 10:25 PM, Jason Rosenberg < > > > >> j...@squareup.com > > > >>>>>> > > > >>>>>>>>> wrote: > > > >>>>>>>>> > > > >>>>>>>>>> In the 0.8 config, log.dir is now log.dirs. It looks like > the > > > >>>>>>> singular > > > >>>>>>>>>> log.dir is still supported, but under the covers the > property > > is > > > >>>>>>>>> log.dirs. > > > >>>>>>>>>> > > > >>>>>>>>>> I'm curious, does this take a comma separated list of > > > directories? > > > >>>>>>> The > > > >>>>>>>>> new > > > >>>>>>>>>> config page just says: > > > >>>>>>>>>> "The directories in which the log data is kept" > > > >>>>>>>>>> > > > >>>>>>>>>> Also, how does kafka handle multiple directories? Does it > > treat > > > >>>>> each > > > >>>>>>>>>> directory as a separate replica partition, or what? > > > >>>>>>>>>> > > > >>>>>>>>>> Jason > > > >> > > > > > >
