Thank you , it has been enabled for the count_disctinct column has been
stored in HDFS.

What I warry about is that if a big dict is not in the cache , a query may
be very slow for having to fetch data from hbase or hdfs.

2017-11-29 16:58 GMT+08:00 ShaoFeng Shi <[email protected]>:

> Hi Hao,
>
> Kylin will automatically detect whether a resource size exceeds HBase
> cell's max size; if yes, it will save it to HDFS:
> https://github.com/apache/kylin/blob/master/storage-
> hbase/src/main/java/org/apache/kylin/storage/hbase/
> HBaseResourceStore.java#L419
>
> Please check whether it works on your side.
>
> 2017-11-29 16:01 GMT+08:00 杨浩 <[email protected]>:
>
>> I have generated the dict data size on TrieDictionaryForestBenchmark.
>> If cardinality is less than  20000, the dict size will be less than 802KB.
>> WIll the cardinality be less than 20000 to set a col as dict  if we want to
>> speed up query speed if the cell size (less than 1MB) is limit by hbase
>> admin?
>>
>> cardinality 0 10000 20000 30000 40000 50000 60000 70000
>> dict size  64B  406KB  802KB  1MB  1MB  1MB  2MB  2MB
>>
>> 2017-11-25 22:42 GMT+08:00 杨浩 <[email protected]>:
>>
>>> Thanks. The biggest number  has been writen in "Kylin Guide", but it may
>>> affect query performance for hbase limit of KV cell size.  As there are
>>> many cubes in the KYLIN , query server would fetch the dict from hbase many
>>> times. Our hbase admin says, if a KV size is under about 500 KB, the query
>>> perfmance can be guaranteed. So the dict size should be less than 500KB in
>>> our env.
>>>
>>> We may choose 1 million or half of that as the guide to use dict to
>>> ensure the query perfmance
>>>
>>> 2017-11-24 22:25 GMT+08:00 ShaoFeng Shi <[email protected]>:
>>>
>>>> The cap is 5 million I remember, But it's better to control that less
>>>> than 1 million.
>>>>
>>>> 2017-11-24 20:33 GMT+08:00 杨浩 <[email protected]>:
>>>>
>>>>> There are many cubes in our kylin env. Can any one give the numer
>>>>> of how big cardinal of a column if we want to code a column as dict?
>>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> Best regards,
>>>>
>>>> Shaofeng Shi 史少锋
>>>>
>>>>
>>>
>>
>
>
> --
> Best regards,
>
> Shaofeng Shi 史少锋
>
>

Reply via email to