Yep. did not work entirely. I had a job to run on 1000 regions. And the caching was 200. The job crashed with a lot of ClosedChannelExceptions + LeaseExceptions.
Set the caching to 10 ==> the same. Set the caching to 1 ==> ~600 successfully completed tasks, but still a lot of them crashed ==> job crashed Set the hbase.rpc.timeout to 240000 (which is the lease timeout on the region server) ==> the job completed successfully, without any failed attempts. The problem was that we have some very large regions (2GB) and there are some of them with very few data, that's why it takes more than 60 seconds to get even the first row. As Daniel said, in the documentation of the lease timeout for regionserver and the hbase.rpc.timeout should be mentioned to be careful when modifying them, because you can get to problems, like in our case. Regards, Lucian On Wed, Oct 26, 2011 at 7:53 PM, Jean-Daniel Cryans <[email protected]>wrote: > Did you try setting the scanner caching down like I mentioned? > > J-D > > On Wed, Oct 26, 2011 at 8:48 AM, Lucian Iordache > <[email protected]> wrote: > > Problem solved. It was like I said, the server took more than the > > hbase.rpc.timeout to run the call and the client closed the connection. > > > > Best Regards, > > Lucian > > > > On Tue, Oct 25, 2011 at 11:15 AM, Lucian Iordache < > > [email protected]> wrote: > > > >> Yes, I will try to see the SocketTimeoutException after putting log on > >> debug, because, like it says here > >> https://issues.apache.org/jira/browse/HBASE-3154 , this is logged on > debug > >> on the client side. > >> > >> Regards, > >> Lucian > >> > >> > >> On Mon, Oct 24, 2011 at 8:22 PM, Jean-Daniel Cryans < > [email protected]>wrote: > >> > >>> So you should see the SocketTimeoutException in your *client* logs (in > >>> your case, mappers), not LeaseException. At this point yes you're > >>> going to timeout, but if you spend so much time cycling on the server > >>> side then you shouldn't set a high caching configuration on your > >>> scanner as IO isn't your bottle neck. > >>> > >>> J-D > >>> > >>> On Mon, Oct 24, 2011 at 10:15 AM, Lucian Iordache > >>> <[email protected]> wrote: > >>> > Hi, > >>> > > >>> > The servers have been restarted (I have this configuration for more > than > >>> a > >>> > month, so this is not the problem). > >>> > About the stack traces, they show exactly the same, a lot of > >>> > ClosedChannelConnections and LeaseExceptions. > >>> > > >>> > But I found something that could be the problem: hbase.rpc.timeout . > >>> This > >>> > defaults to 60 seconds, and I did not modify it in hbase-site.xml. So > it > >>> > could happen the next way: > >>> > - the mapper makes a scanner.next call to the region server > >>> > - the region servers needs more than 60 seconds to execute it (I use > >>> > multiple filters, and it could take a lot of time) > >>> > - the scan client gets the timeout and cuts the connection > >>> > - the region server tries to send the results to the client ==> > >>> > ClosedChannelConnection > >>> > > >>> > I will get a deeper look into it tomorrow. If you have other > >>> suggestions, > >>> > please let me know! > >>> > > >>> > Thanks, > >>> > Lucian > >>> > > >>> > On Mon, Oct 24, 2011 at 8:00 PM, Jean-Daniel Cryans < > >>> [email protected]>wrote: > >>> > > >>> >> Did you restart the region servers after changing the config? > >>> >> > >>> >> Are you sure it's the same exception/stack trace? > >>> >> > >>> >> J-D > >>> >> > >>> >> On Mon, Oct 24, 2011 at 8:04 AM, Lucian Iordache > >>> >> <[email protected]> wrote: > >>> >> > Hi all, > >>> >> > > >>> >> > I have exactly the same problem that Eran had. > >>> >> > But there is something I don't understand: in my case, I have set > the > >>> >> lease > >>> >> > time to 240000 (4 minutes). But most of the map tasks that are > >>> failing > >>> >> run > >>> >> > about 2 minutes. How is it possible to get a LeaseException if the > >>> task > >>> >> runs > >>> >> > less than the configured time for a lease? > >>> >> > > >>> >> > Regards, > >>> >> > Lucian Iordache > >>> >> > > >>> >> > On Fri, Oct 21, 2011 at 12:34 AM, Eran Kutner <[email protected]> > >>> wrote: > >>> >> > > >>> >> >> Perfect! Thanks. > >>> >> >> > >>> >> >> -eran > >>> >> >> > >>> >> >> > >>> >> >> > >>> >> >> On Thu, Oct 20, 2011 at 23:27, Jean-Daniel Cryans < > >>> [email protected] > >>> >> >> >wrote: > >>> >> >> > >>> >> >> > hbase.regionserver.lease.period > >>> >> >> > > >>> >> >> > Set it bigger than 60000. > >>> >> >> > > >>> >> >> > J-D > >>> >> >> > > >>> >> >> > On Thu, Oct 20, 2011 at 2:23 PM, Eran Kutner <[email protected]> > >>> wrote: > >>> >> >> > > > >>> >> >> > > Thanks J-D! > >>> >> >> > > Since my main table is expected to continue growing I guess > at > >>> some > >>> >> >> point > >>> >> >> > > even setting the cache size to 1 will not be enough. Is there > a > >>> way > >>> >> to > >>> >> >> > > configure the lease timeout? > >>> >> >> > > > >>> >> >> > > -eran > >>> >> >> > > > >>> >> >> > > > >>> >> >> > > > >>> >> >> > > On Thu, Oct 20, 2011 at 23:16, Jean-Daniel Cryans < > >>> >> [email protected] > >>> >> >> > >wrote: > >>> >> >> > > > >>> >> >> > > > On Wed, Oct 19, 2011 at 12:51 PM, Eran Kutner < > [email protected] > >>> > > >>> >> >> wrote: > >>> >> >> > > > > >>> >> >> > > > > Hi J-D, > >>> >> >> > > > > Thanks for the detailed explanation. > >>> >> >> > > > > So if I understand correctly the lease we're talking > about > >>> is a > >>> >> >> > scanner > >>> >> >> > > > > lease and the timeout is between two scanner calls, > correct? > >>> I > >>> >> >> think > >>> >> >> > that > >>> >> >> > > > > make sense because I now realize that jobs that fail > (some > >>> jobs > >>> >> >> > continued > >>> >> >> > > > > to > >>> >> >> > > > > fail even after reducing the number of map tasks as Stack > >>> >> >> suggested) > >>> >> >> > use > >>> >> >> > > > > filters to fetch relatively few rows out of a very large > >>> table, > >>> >> so > >>> >> >> > they > >>> >> >> > > > > could be spending a lot of time on the region server > >>> scanning > >>> >> rows > >>> >> >> > until > >>> >> >> > > > it > >>> >> >> > > > > reached my setCaching value which was 1000. Setting the > >>> caching > >>> >> >> value > >>> >> >> > to > >>> >> >> > > > 1 > >>> >> >> > > > > seem to allow these job to complete. > >>> >> >> > > > > I think it has to be the above, since my rows are small, > >>> with > >>> >> just > >>> >> >> a > >>> >> >> > few > >>> >> >> > > > > columns and processing them is very quick. > >>> >> >> > > > > > >>> >> >> > > > > >>> >> >> > > > Excellent! > >>> >> >> > > > > >>> >> >> > > > > >>> >> >> > > > > > >>> >> >> > > > > However, there are still a couple ofw thing I don't > >>> understand: > >>> >> >> > > > > 1. What is the difference between setCaching and > setBatch? > >>> >> >> > > > > > >>> >> >> > > > > >>> >> >> > > > * Set the maximum number of values to return for each call > to > >>> >> next() > >>> >> >> > > > > >>> >> >> > > > VS > >>> >> >> > > > > >>> >> >> > > > * Set the number of rows for caching that will be passed to > >>> >> scanners. > >>> >> >> > > > > >>> >> >> > > > The former is useful if you have rows with millions of > columns > >>> and > >>> >> >> you > >>> >> >> > > > could > >>> >> >> > > > setBatch to get only 1000 of them at a time. You could call > >>> that > >>> >> >> > intra-row > >>> >> >> > > > scanning. > >>> >> >> > > > > >>> >> >> > > > > >>> >> >> > > > > 2. Examining the region server logs more closely than I > did > >>> >> >> yesterday > >>> >> >> > I > >>> >> >> > > > see > >>> >> >> > > > > a log of ClosedChannelExceptions in addition to the > expired > >>> >> leases > >>> >> >> > (but > >>> >> >> > > > no > >>> >> >> > > > > UnknownScannerException), is that expected? You can see > an > >>> >> excerpt > >>> >> >> of > >>> >> >> > the > >>> >> >> > > > > log from one of the region servers here: > >>> >> >> > http://pastebin.com/NLcZTzsY > >>> >> >> > > > > >>> >> >> > > > > >>> >> >> > > > It means that when the server got to process that client > >>> request > >>> >> and > >>> >> >> > > > started > >>> >> >> > > > reading from the socket, the client was already gone. > Killing > >>> a > >>> >> >> client > >>> >> >> > does > >>> >> >> > > > that (or killing a MR that scans), so does > >>> SocketTimeoutException. > >>> >> >> This > >>> >> >> > > > should probably go in the book. We should also print > something > >>> >> nicer > >>> >> >> :) > >>> >> >> > > > > >>> >> >> > > > J-D > >>> >> >> > > > > >>> >> >> > > >>> >> >> > >>> >> > > >>> >> > >>> > > >>> > >> > >> > >> > >> -- > >> Numai bine, > >> Lucian > >> > > > > > > > > -- > > Numai bine, > > Lucian > > > -- Numai bine, Lucian
