As a follow-up, is operating a Cassandra cluster with machines on multiple racks and vnodes bound to cause load imbalance ? Shouldn't token-ranges assigned to individual machines via their vnodes be approximately balanced ? We're otherwise unable to explain why this imbalance occurs. ( it shouldn't be the fault of the Murmur3 partitioner which guarantees a uniform distribution of keys across token-ranges according to the doc. )
On Thu, Sep 19, 2013 at 8:56 PM, Jayadev Jayaraman <jdisal...@gmail.com>wrote: > We're using NetworkTopologyStrategy for our placement_strategy and our > primary motivation for placing nodes on more than 1 rack is to prevent > showstoppers caused by localized faults leading to multiple EC2 instances > in the same availability zone (rack) becoming unreachable (which we have > encountered before). > > I've attached the output of "nodetool ring" here. > > > On Thu, Sep 19, 2013 at 8:35 PM, Mohit Anchlia <mohitanch...@gmail.com>wrote: > >> Other thing I noticed is that you are using mutiple RACKS and that might >> be contributing factor to it. However, I am not sure. >> >> Can you paste the output of nodetool cfstats and ring? Is it possible to >> run the same test but keeping all the nodes in one rack? >> >> I think you should open a JIRA if you are able to reproduce this. >> >> On Thu, Sep 19, 2013 at 4:41 PM, Jayadev Jayaraman >> <jdisal...@gmail.com>wrote: >> >>> We ran nodetool repair on all nodes for all Keyspaces / CFs, restarted >>> cassandra and this is what we get for nodetool status : >>> >>> bin/nodetool -h localhost status >>> Datacenter: us-east >>> =================== >>> Status=Up/Down >>> |/ State=Normal/Leaving/Joining/Moving >>> -- Address Load Tokens Owns (effective) Host ID >>> Rack >>> UN 10.238.133.174 885.36 MB 256 8.4% >>> e41d8863-ce37-4d5c-a428-bfacea432a35 1a >>> UN 10.238.133.97 468.66 MB 256 7.7% >>> 1bf42b5e-4aed-4b06-bdb3-65a78823b547 1a >>> UN 10.151.86.146 1.08 GB 256 8.0% >>> 8952645d-4a27-4670-afb2-65061c205734 1a >>> UN 10.138.10.9 941.44 MB 256 8.6% >>> 25ccea82-49d2-43d9-830c-b9c9cee026ec 1a >>> UN 10.87.87.240 99.69 MB 256 8.6% >>> ea066827-83bc-458c-83e8-bd15b7fc783c 1b >>> UN 10.93.5.157 87.44 MB 256 7.6% >>> 4ab9111c-39b4-4d15-9401-359d9d853c16 1b >>> UN 10.238.137.250 561.42 MB 256 7.8% >>> 84301648-afff-4f06-aa0b-4be421e0d08f 1a >>> UN 10.92.231.170 893.75 MB 256 9.3% >>> a18ce761-88a0-4407-bbd1-c867c4fecd1f 1b >>> UN 10.138.2.20 31.89 MB 256 7.9% >>> a6d4672a-0915-4c64-ba47-9f190abbf951 1a >>> UN 10.93.31.44 312.52 MB 256 7.8% >>> 67a6c0a6-e89f-4f3e-b996-cdded1b94faf 1b >>> UN 10.93.91.139 30.46 MB 256 8.1% >>> 682dd848-7c7f-4ddb-a960-119cf6491aa1 1b >>> UN 10.236.138.169 260.15 MB 256 9.1% >>> cbbf27b0-b53a-4530-bfdf-3764730b89d8 1a >>> UN 10.137.7.90 38.45 MB 256 7.4% >>> 17b79aa7-64fc-4e16-b96a-955b0aae9bb4 1a >>> UN 10.93.77.166 867.15 MB 256 8.8% >>> 9a821d1e-40e5-445d-b6b7-3cdd58bdb8cb 1b >>> UN 10.120.249.140 863.98 MB 256 9.4% >>> e1fb69b0-8e66-4deb-9e72-f901d7a14e8a 1b >>> UN 10.90.246.128 242.63 MB 256 8.4% >>> 054911ec-969d-43d9-aea1-db445706e4d2 1b >>> UN 10.123.95.248 171.51 MB 256 7.2% >>> a17deca1-9644-4520-9e62-ac66fc6fef60 1b >>> UN 10.136.11.40 33.8 MB 256 8.5% >>> 66be1173-b822-40b5-b650-cb38ae3c7a51 1a >>> UN 10.87.90.42 38.01 MB 256 8.0% >>> dac0c6ea-56c6-44da-a4ec-6388f39ecba1 1b >>> UN 10.87.75.147 579.29 MB 256 8.3% >>> ac060edf-dc48-44cf-a1b5-83c7a465f3c8 1b >>> UN 10.151.49.88 151.06 MB 256 8.9% >>> 57043573-ab1b-4e3c-8044-58376f7ce08f 1a >>> UN 10.87.83.107 512.91 MB 256 8.3% >>> 0019439b-9f8a-4965-91b8-7108bbb55593 1b >>> UN 10.238.170.159 85.04 MB 256 9.4% >>> 32ce322e-4f7c-46c7-a8ce-bd73cdd54684 1a >>> UN 10.137.20.183 167.41 MB 256 8.4% >>> 15951592-8ab2-473d-920a-da6e9d99507d 1a >>> >>> It doesn't seem to have changed by much. The loads are still highly >>> uneven. >>> >>> As for the number of keys in each node's CFs : the largest node now >>> has 5589120 keys for the column-family that had 6527744 keys before >>> (load is now 1.08 GB as compares to 1.05 GB before), while the smallest >>> node now has 71808 keys as compared to 3840 keys before (load is now >>> 31.89 MB as compares to 1.12 MB before). >>> >>> >>> On Thu, Sep 19, 2013 at 5:18 PM, Mohit Anchlia >>> <mohitanch...@gmail.com>wrote: >>> >>>> Can you run nodetool repair on all the nodes first and look at the keys? >>>> >>>> >>>> On Thu, Sep 19, 2013 at 1:22 PM, Suruchi Deodhar < >>>> suruchi.deod...@generalsentiment.com> wrote: >>>> >>>>> Yes, the key distribution does vary across the nodes. For example, on >>>>> the node with the highest data, Number of Keys (estimate) is 6527744 for a >>>>> particular column family, whereas for the same column family on the node >>>>> with least data, Number of Keys (estimate) = 3840. >>>>> >>>>> Is there a way to control this distribution by setting some parameter >>>>> of cassandra. >>>>> >>>>> I am using the Murmur3 partitioner with NetworkTopologyStrategy. >>>>> >>>>> Thanks, >>>>> Suruchi >>>>> >>>>> >>>>> >>>>> On Thu, Sep 19, 2013 at 3:59 PM, Mohit Anchlia <mohitanch...@gmail.com >>>>> > wrote: >>>>> >>>>>> Can you check cfstats to see number of keys per node? >>>>>> >>>>>> >>>>>> On Thu, Sep 19, 2013 at 12:36 PM, Suruchi Deodhar < >>>>>> suruchi.deod...@generalsentiment.com> wrote: >>>>>> >>>>>>> Thanks for your replies. I wiped out my data from the cluster and >>>>>>> also cleared the commitlog before restarting it with num_tokens=256. I >>>>>>> then >>>>>>> uploaded data using sstableloader. >>>>>>> >>>>>>> However, I am still not able to see a uniform distribution of data >>>>>>> across nodes of the clusters. >>>>>>> >>>>>>> The output of the bin/nodetool -h localhost status commands looks >>>>>>> like follows. Some nodes have data as low as 1.12MB while some have as >>>>>>> high >>>>>>> as 912.57 MB. >>>>>>> >>>>>>> Datacenter: us-east >>>>>>> =================== >>>>>>> Status=Up/Down >>>>>>> |/ State=Normal/Leaving/Joining/Moving >>>>>>> -- Address Load Tokens Owns (effective) Host >>>>>>> ID Rack >>>>>>> UN 10.238.133.174 856.66 MB 256 8.4% >>>>>>> e41d8863-ce37-4d5c-a428-bfacea432a35 1a >>>>>>> UN 10.238.133.97 439.02 MB 256 7.7% >>>>>>> 1bf42b5e-4aed-4b06-bdb3-65a78823b547 1a >>>>>>> UN 10.151.86.146 1.05 GB 256 8.0% >>>>>>> 8952645d-4a27-4670-afb2-65061c205734 1a >>>>>>> UN 10.138.10.9 912.57 MB 256 8.6% >>>>>>> 25ccea82-49d2-43d9-830c-b9c9cee026ec 1a >>>>>>> UN 10.87.87.240 70.85 MB 256 8.6% >>>>>>> ea066827-83bc-458c-83e8-bd15b7fc783c 1b >>>>>>> UN 10.93.5.157 60.56 MB 256 7.6% >>>>>>> 4ab9111c-39b4-4d15-9401-359d9d853c16 1b >>>>>>> UN 10.92.231.170 866.73 MB 256 9.3% >>>>>>> a18ce761-88a0-4407-bbd1-c867c4fecd1f 1b >>>>>>> UN 10.238.137.250 533.77 MB 256 7.8% >>>>>>> 84301648-afff-4f06-aa0b-4be421e0d08f 1a >>>>>>> UN 10.93.91.139 478.45 KB 256 8.1% >>>>>>> 682dd848-7c7f-4ddb-a960-119cf6491aa1 1b >>>>>>> UN 10.138.2.20 1.12 MB 256 7.9% >>>>>>> a6d4672a-0915-4c64-ba47-9f190abbf951 1a >>>>>>> UN 10.93.31.44 282.65 MB 256 7.8% >>>>>>> 67a6c0a6-e89f-4f3e-b996-cdded1b94faf 1b >>>>>>> UN 10.236.138.169 223.66 MB 256 9.1% >>>>>>> cbbf27b0-b53a-4530-bfdf-3764730b89d8 1a >>>>>>> UN 10.137.7.90 11.36 MB 256 7.4% >>>>>>> 17b79aa7-64fc-4e16-b96a-955b0aae9bb4 1a >>>>>>> UN 10.93.77.166 837.64 MB 256 8.8% >>>>>>> 9a821d1e-40e5-445d-b6b7-3cdd58bdb8cb 1b >>>>>>> UN 10.120.249.140 838.59 MB 256 9.4% >>>>>>> e1fb69b0-8e66-4deb-9e72-f901d7a14e8a 1b >>>>>>> UN 10.90.246.128 216.75 MB 256 8.4% >>>>>>> 054911ec-969d-43d9-aea1-db445706e4d2 1b >>>>>>> UN 10.123.95.248 147.1 MB 256 7.2% >>>>>>> a17deca1-9644-4520-9e62-ac66fc6fef60 1b >>>>>>> UN 10.136.11.40 4.24 MB 256 8.5% >>>>>>> 66be1173-b822-40b5-b650-cb38ae3c7a51 1a >>>>>>> UN 10.87.90.42 11.56 MB 256 8.0% >>>>>>> dac0c6ea-56c6-44da-a4ec-6388f39ecba1 1b >>>>>>> UN 10.87.75.147 549 MB 256 8.3% >>>>>>> ac060edf-dc48-44cf-a1b5-83c7a465f3c8 1b >>>>>>> UN 10.151.49.88 119.86 MB 256 8.9% >>>>>>> 57043573-ab1b-4e3c-8044-58376f7ce08f 1a >>>>>>> UN 10.87.83.107 484.3 MB 256 8.3% >>>>>>> 0019439b-9f8a-4965-91b8-7108bbb55593 1b >>>>>>> UN 10.137.20.183 137.67 MB 256 8.4% >>>>>>> 15951592-8ab2-473d-920a-da6e9d99507d 1a >>>>>>> UN 10.238.170.159 49.17 MB 256 9.4% >>>>>>> 32ce322e-4f7c-46c7-a8ce-bd73cdd54684 1a >>>>>>> >>>>>>> Is there something else that I should be doing differently? >>>>>>> >>>>>>> Thanks for your help! >>>>>>> >>>>>>> Suruchi >>>>>>> >>>>>>> >>>>>>> >>>>>>> On Thu, Sep 19, 2013 at 3:20 PM, Richard Low <rich...@wentnet.com>wrote: >>>>>>> >>>>>>>> The only thing you need to guarantee is that Cassandra doesn't >>>>>>>> start with num_tokens=1 (the default in 1.2.x) or, if it does, that you >>>>>>>> wipe all the data before starting it with higher num_tokens. >>>>>>>> >>>>>>>> >>>>>>>> On 19 September 2013 19:07, Robert Coli <rc...@eventbrite.com>wrote: >>>>>>>> >>>>>>>>> On Thu, Sep 19, 2013 at 10:59 AM, Suruchi Deodhar < >>>>>>>>> suruchi.deod...@generalsentiment.com> wrote: >>>>>>>>> >>>>>>>>>> Do you suggest I should try with some other installation >>>>>>>>>> mechanism? Are there any known problems with the tar installation of >>>>>>>>>> cassandra 1.2.9 that I should be aware of? >>>>>>>>>> >>>>>>>>> >>>>>>>>> I was asking in the context of this JIRA : >>>>>>>>> >>>>>>>>> https://issues.apache.org/jira/browse/CASSANDRA-2356 >>>>>>>>> >>>>>>>>> Which does not seem to apply in your case! >>>>>>>>> >>>>>>>>> =Rob >>>>>>>>> >>>>>>>> >>>>>>>> >>>>>>> >>>>>> >>>>> >>>> >>> >> >
