I already patched our software and this object is ignored now. It only contains data no-one needs (it is an archive of references to archived messages). The sibling explosion was caused by the migration script which runs in the background converting old data from our mysql database, I'm still looking into this and the script is stopped in the mean time.
Ingo Am 08.04.2013 16:11, schrieb Evan Vigil-McClanahan:
You could try to read it by doing a HEAD request, which while cluster impacting wouldn't try to pass everything over the wire. This isn't necessary if you already have the latest vclock from some previous attempt or write. Then you could overwrite the object with the latest vclock (again, more impact), which would have the same effect as deleting the object. If that fails, there are other options we can explore. The worrying question here is how this object got so many siblings in the first place, which may be something that you want to look into and address. On Mon, Apr 8, 2013 at 3:36 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:Hi, I've finally been able to identify the big object (was a tough one), but unfortuneately, riak fails deleting it: irockel@bighead:~$ curl -v -X DELETE "http://172.22.3.22:8091/riak/m/Oa|1" * About to connect() to 172.22.3.22 port 8091 (#0) * Trying 172.22.3.22... connectedDELETE /riak/m/Oa|1 HTTP/1.1 User-Agent: curl/7.22.0 (x86_64-pc-linux-gnu) libcurl/7.22.0 OpenSSL/1.0.1 zlib/1.2.3.4 libidn/1.23 librtmp/2.3 Host: 172.22.3.22:8091 Accept: */*< HTTP/1.1 503 Service Unavailable < Server: MochiWeb/1.1 WebMachine/1.9.2 (someone had painted it blue) < Date: Mon, 08 Apr 2013 08:30:08 GMT < Content-Type: text/plain < Content-Length: 18 < Any suggestions how I could get rid of the object? It seems to be really big and has a lot of siblings (>100) which sums up to 2GB. Ingo Am 04.04.2013 17:51, schrieb Evan Vigil-McClanahan:Possible, but would need more information to make a guess. I'd keep a close eye on that node. On Thu, Apr 4, 2013 at 10:34 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:thanks, but it was a very obvious c&p error :) and we already have the ERL_MAX_ETS_TABLES set to 8192 as it is in the default vm.args. The only other messages were about a lot of handoff going on. Maybe the node was getting some data concerning the 2GB object? Ingo Am 04.04.2013 17:25, schrieb Evan Vigil-McClanahan:Major error on my part here!your vm.args: -env ERL_MAX_ETS_TABLES 819This should be -env ERL_MAX_ETS_TABLES 8192 Sorry for the sloppy cut and paste. Please do not do the former thing, or it will be very bad.This is a good idea for all systems but is especially important for people with large rings. Were there any other messages? Riak constantly spawns new processes, but they don't tend to build up unless the backend is misbehaving (or a few other less likely conditions), and a backup of spawned processes is the only thing I can think of that would make +P help with OOM issues. On Thu, Apr 4, 2013 at 9:21 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:A grep for "too many processes" didn't reveal anything. The process got killed by the oom-killer. Am 04.04.2013 16:12, schrieb Evan Vigil-McClanahan:That's odd. It was getting killed by the OOM killer, or crashing because it couldn't allocate more memory? That's suggestive of something else that's wrong, since the +P doesn't do any memory limiting. Are you getting 'too many processes' emulator errors on that node? On Thu, Apr 4, 2013 at 8:47 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:the crashing node seems to be caused by the raised +P param, after last crash I commented the param and now the node runs just fine. Am 04.04.2013 15:43, schrieb Ingo Rockel:Hi Evan, we added monitoring of the object sizes and there was one object on one of the three nodes mentioned which was > 2GB!! We just changed the application code to get the id of this object to be able to delete it. But is does happen only about once a day. We right now have another node constantly crashing with oom about 12 minutes after start (always the same time frame), could this be related to the big object issue? It is not one of the three nodes. The node logs a lot of handoff receiving is going on. Again, thanks for the help! Regards, Ingo Am 04.04.2013 15:30, schrieb Evan Vigil-McClanahan:If it's always the same three nodes it could well be same very large object being updated each day. Is there anything else that looks suspicious in your logs? Another sign of large objects is large_heap (or long_gc) messages from riak_sysmon. On Thu, Apr 4, 2013 at 3:58 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:Hi Evan, thanks for all the infos! I adjusted the leveldb-config as suggested, except the cache, which I reduced to 16MB, keeping this above the default helped a lot at least during load testing. And I added +P 130072 to the vm.args. Will be applied to the riak nodes the next hours. We have a monitoring using zabbbix, but haven't included the object sizes so far, will be added today. We double-checked the Linux-Performance-Doc to be sure everything is applied to the nodes, especially as the problems always are caused from the same three nodes. But everything looks fine. Ingo Am 03.04.2013 18:42, schrieb Evan Vigil-McClanahan:Another engineer mentions that you posted your eleveldb section and I totally missed it: The eleveldb section: %% eLevelDB Config {eleveldb, [ {data_root, "/var/lib/riak/leveldb"}, {cache_size, 33554432}, {write_buffer_size_min, 67108864}, %% 64 MB in bytes {write_buffer_size_max, 134217728}, %% 128 MB in bytes {max_open_files, 4000} ]}, This is likely going to make you unhappy as time goes on; Since all of those settings are per-vnode, your max memory utilization is well beyond your physical memory. I'd remove the tunings for the caches and buffers and drop max open files to 500, perhaps. Make sure that you've followed everything in: http://docs.basho.com/riak/latest/cookbooks/Linux-Performance-Tuning/, etc. On Wed, Apr 3, 2013 at 9:33 AM, Evan Vigil-McClanahan <emcclana...@basho.com> wrote:Again, all of these things are signs of large objects, so if you could track the object_size stats on the cluster, I think that we might see something. Even if you have no monitoring, a simple shell script curling /stats/ on each node once a minute should do the job for a day or two. On Wed, Apr 3, 2013 at 9:29 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:We just had it again (around this time of the day we have our highest user activity). I will set +P to 131072 tomorrow, anything else I should check or change? What about this memory-high-watermark which I get sporadically? Ingo Am 03.04.2013 17:57, schrieb Evan Vigil-McClanahan:As for +P it's been raised in R16 (which is on the current man page) on R15 it's only 32k. The behavior that you're describing does sound like a very large object getting put into the cluster (which may cause backups and push you up against the process limit, could have caused scheduler collapse on 1.2, etc.). On Wed, Apr 3, 2013 at 8:39 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:Evan, sys_process_count is somewhere between 5k and 11k on the nodes right now. Concerning your suggested +P config, according to the erlang-docs, the default for this param already is 262144, so setting it to 655536 would in fact lower it? We chose the ring size to be able to handle growth which was the main reason to switch from mysql to nosql/riak. We have 12 Nodes, so about 86 vnodes per node. No, we don't monitor object sizes, the majority of objects is very small (below 200 bytes), but we have objects storing references to this small objects which might grow to a few megabytes in size, most of these are paged though and should not exceed one megabyte. Only one type is not paged (implementation reasons). The outgoing/incoming traffic constantly is around 100 Mbit, if the peformance drops happen, we suddenly see spikes up to 1GBit. And these spikes constantly happen on three nodes as long as the performance drop exists. Ingo Am 03.04.2013 17:12, schrieb Evan Vigil-McClanahan:Ingo, riak-admin status | grep sys_process_count will tell you how many processes are running. The default process limit on erlang is a little low, and we'd suggest raising in (especially with your extra-large ring_size). Erlang processes are cheap, so 65535 or even double that will be fine. Busy dist ports are still worrying. Are you monitoring object sizes? Are there any spikes there associated with performance drops? On Wed, Apr 3, 2013 at 8:03 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:Hi Evan, I set swt very_low and zdbbl to 64MB, setting this params helped reducing the busy_dist_port and Monitor got {suppressed,... Messages a lot. But when the performance of the cluster suddenly drops we still see these messages. The cluster was updated to 1.3 in the meantime. The eleveldb section: %% eLevelDB Config {eleveldb, [ {data_root, "/var/lib/riak/leveldb"}, {cache_size, 33554432}, {write_buffer_size_min, 67108864}, %% 64 MB in bytes {write_buffer_size_max, 134217728}, %% 128 MB in bytes {max_open_files, 4000} ]}, the ring size is 1024 and the machines have 48GB of memory. Concerning the params from vm.args: -env ERL_MAX_PORTS 4096 -env ERL_MAX_ETS_TABLES 8192 +P isn't set Ingo Am 03.04.2013 16:53, schrieb Evan Vigil-McClanahan:For your prior mail, I thought that someone had answered. Our initial suggestion was to add +swt very_low to your vm.args, as well as setting the +zdbbl setting that Jon recommended in the list post you pointed to. If those help, moving to 1.3 should help more. Other limits in vm.args that can cause problems are +P, ERL_MAX_PORTS, and ERL_MAX_ETS_TABLES. Are any of these set? If so, to what? Can you also pate the eleveldb section of your app.config? On Wed, Apr 3, 2013 at 7:41 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:Hi Evan, I'm not sure, I find a lot of these: 2013-03-30 23:27:52.992 [error] <0.8036.323>@riak_api_pb_server:handle_info:141 Unrecognized message {22243034,{error,timeout}} and some of these at the same time one of the kind below gets logged (although the one has a different time stamp): 2013-03-30 23:27:53.056 [error] <0.9457.323>@riak_kv_console:status:178 Status failed error:terminated Ingo Am 03.04.2013 16:24, schrieb Evan Vigil-McClanahan:Resending to the list: Ingo, That is an indication that the protocol buffers server can't spawn a put fsm, which means that a put cannot be done for some reason or another. Are there any other messages that appear around this time that might indicate why? On Wed, Apr 3, 2013 at 12:09 AM, Ingo Rockel <ingo.roc...@bluelionmobile.com> wrote:Hi, we have some performance issues with our riak cluster, from time to time we have a sudden drop in performance (already asked the list about this, no-one had an idea though). Although not the same time but on the problematic nodes we have a lot of these messages from time to time: 2013-04-02 21:41:11.173 [warning] <0.25646.475> ** Can not start proc_lib:init_p ,[<0.14556.474>,[<0.9519.474>,riak_api_pb_sup,riak_api_sup,<0.1291.0>],riak_kv_p ut_fsm,start_link,[{raw,65032165,<0.9519.474>},{r_object,<<109>>,<<77,115,124,49 ,53,55,57,56,57,56,50,124,49,51,54,52,57,51,49,54,49,49,53,49,50,52,53,54>>,[{r_ content,{dict,0,16,16,8,80,48,{[],[],[],[],[],[],[],[],[],[],[],[],[],[],[],[]}, {{[],[],[],[],[],[],[],[],[],[],[],[],[],[],[],[]}}},<<>>}],[],{dict,2,16,16,8,8 0,48,{[],[],[],[],[],[],[],[],[],[],[],[],[],[],[],[]},{{[],[],[],[],[],[],[],[] ,[],[],[[<<99,111,110,116,101,110,116,45,116,121,112,101>>,97,112,112,108,105,99 ,97,116,105,111,110,47,106,115,111,110]],[],[],[],[],[[<<99,104,97,114,115,101,1 16>>,85,84,70,45,56]]}}},<<123,34,115,116,34,58,50,44,34,116,34,58,49,44,34,99,3 4,58,34,66,117,116,32,115,104,101,32,105,115,32,103,111,110,101,44,32,110,32,101 ,118,101,110,32,116,104,111,117,103,104,32,105,109,32,110,111,116,32,105,110,32, 117,114,32,99,105,116,121,32,105,32,108,111,118,101,32,117,32,110,100,32,105,32, 109,101,97,110,32,105,116,32,58,39,40,34,44,34,114,34,58,49,52,51,52,54,52,51,57 ,44,34,115,34,58,49,53,55,57,56,57,56,50,44,34,99,116,34,58,49,51,54,52,57,51,49 ,54,49,49,53,49,50,44,34,97,110,34,58,102,97,108,115,101,44,34,115,107,34,58,49, 51,54,52,57,51,49,54,49,49,53,49,50,52,53,54,44,34,115,117,34,58,48,125>>},[{tim eout,60000}]]] on 'riak@172.22.3.12' ** Can anyone explain to me what these messages mean and if I need to do something about it? Could these messages be in any way related to the performance issues? Ingo-- Software Architect Blue Lion mobile GmbH Tel. +49 (0) 221 788 797 14 Fax. +49 (0) 221 788 797 19 Mob. +49 (0) 176 24 87 30 89 ingo.roc...@bluelionmobile.comqeep: Hefferwolfwww.bluelionmobile.com www.qeep.net-- Software Architect Blue Lion mobile GmbH Tel. +49 (0) 221 788 797 14 Fax. +49 (0) 221 788 797 19 Mob. +49 (0) 176 24 87 30 89 ingo.roc...@bluelionmobile.comqeep: Hefferwolfwww.bluelionmobile.com www.qeep.net
-- Software Architect Blue Lion mobile GmbH Tel. +49 (0) 221 788 797 14 Fax. +49 (0) 221 788 797 19 Mob. +49 (0) 176 24 87 30 89 ingo.roc...@bluelionmobile.com >>> qeep: Hefferwolf www.bluelionmobile.com www.qeep.net _______________________________________________ riak-users mailing list riak-users@lists.basho.com http://lists.basho.com/mailman/listinfo/riak-users_lists.basho.com
