Duncan: I'm thinking it might be something like that. I'm also seeing just
a ton of garbage collection on the box, could it be pulling rows for all
100k attrs for a given row_time into memory since only row_time is the
partition key?
Jens: I'm not using EBS (although I used to until I read up on how useless
it is). I'm not sure what constitutes proper paging but my client has a
pretty small amount of available memory so I'm doing pages of size 5k using
the C++ Datastax driver.
Thanks for the replies!
-Dave
On Mon, Mar 23, 2015 at 2:00 AM, Jens Rantil <jens.ran...@tink.se> wrote:
> Also, two control questions:
>
> - Are you using EBS for data storage? It might introduce additional
> latencies.
> - Are you doing proper paging when querying the keyspace?
>
> Cheers,
> Jens
>
> On Mon, Mar 23, 2015 at 5:56 AM, Dave Galbraith <
> david92galbra...@gmail.com> wrote:
>
>> Hi! So I've got a table like this:
>>
>> CREATE TABLE "default".metrics (row_time int,attrs varchar,offset
>> int,value double, PRIMARY KEY(row_time, attrs, offset)) WITH COMPACT
>> STORAGE AND bloom_filter_fp_chance=0.01 AND caching='KEYS_ONLY' AND
>> comment='' AND dclocal_read_repair_chance=0 AND gc_grace_seconds=864000 AND
>> index_interval=128 AND read_repair_chance=1 AND replicate_on_write='true'
>> AND populate_io_cache_on_flush='false' AND default_time_to_live=0 AND
>> speculative_retry='NONE' AND memtable_flush_period_in_ms=0 AND
>> compaction={'class':'DateTieredCompactionStrategy','timestamp_resolution':'MILLISECONDS'}
>> AND compression={'sstable_compression':'LZ4Compressor'};
>>
>> and I'm running Cassandra on an EC2 m3.2xlarge out in the cloud, with 4
>> GB of heap space. So it's timeseries data that I'm doing so I increment
>> "row_time" each day, "attrs" is additional identifying information about
>> each series, and "offset" is the number of milliseconds into the day for
>> each data point. So for the past 5 days, I've been inserting 3k
>> points/second distributed across 100k distinct "attrs"es. And now when I
>> try to run queries on this data that look like
>>
>> "SELECT * FROM "default".metrics WHERE row_time = 5 AND attrs =
>> 'potatoes_and_jam'"
>>
>> it takes an absurdly long time and sometimes just times out. I did
>> "nodetool cftsats default" and here's what I get:
>>
>> Keyspace: default
>> Read Count: 59
>> Read Latency: 397.12523728813557 ms.
>> Write Count: 155128
>> Write Latency: 0.3675690719921613 ms.
>> Pending Flushes: 0
>> Table: metrics
>> SSTable count: 26
>> Space used (live): 35146349027
>> Space used (total): 35146349027
>> Space used by snapshots (total): 0
>> SSTable Compression Ratio: 0.10386468749216264
>> Memtable cell count: 141800
>> Memtable data size: 31071290
>> Memtable switch count: 41
>> Local read count: 59
>> Local read latency: 397.126 ms
>> Local write count: 155128
>> Local write latency: 0.368 ms
>> Pending flushes: 0
>> Bloom filter false positives: 0
>> Bloom filter false ratio: 0.00000
>> Bloom filter space used: 2856
>> Compacted partition minimum bytes: 104
>> Compacted partition maximum bytes: 36904729268
>> Compacted partition mean bytes: 986530969
>> Average live cells per slice (last five minutes):
>> 501.66101694915255
>> Maximum live cells per slice (last five minutes): 502.0
>> Average tombstones per slice (last five minutes): 0.0
>> Maximum tombstones per slice (last five minutes): 0.0
>>
>> Ouch! 400ms of read latency, orders of magnitude higher than it has any
>> right to be. How could this have happened? Is there something fundamentally
>> broken about my data model? Thanks!
>>
>>
>
>
> --
> Jens Rantil
> Backend engineer
> Tink AB
>
> Email: jens.ran...@tink.se
> Phone: +46 708 84 18 32
> Web: www.tink.se
>
> Facebook <https://www.facebook.com/#!/tink.se> Linkedin
> <http://www.linkedin.com/company/2735919?trk=vsrp_companies_res_photo&trkInfo=VSRPsearchId%3A1057023381369207406670%2CVSRPtargetId%3A2735919%2CVSRPcmpt%3Aprimary>
> Twitter <https://twitter.com/tink>
>
