kylin-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alberto Ramón <a.ramonporto...@gmail.com>
Subject Re:
Date Mon, 16 Jan 2017 11:45:31 GMT
Hi Phon, I'm not expert but I have some suggestions:

- All Dim en are using Dict: you can change a lot to Integer (Fix length)
- Re-Order row key its a good idea. I always try to first fields of key
have Fix Length. Put mandatory the First its a good Idea
- See hierarchy optimizations, will be very interesting for you: Country,
Region, City, site . Perhaps Company  and Account also can be included (I
don't know your data)
- If you use Left join, the first step of building cube (flat table) will
be more slow
- Check if your ORC input table is compressed
- Try to use derived DIm with very low cardinality columns, perhaps: TypeID,
NetworkID, LanguajeID, IsMovileDevice.
   I understand that Affiliated, Account, Company, ... will growth in the
future, because you are working with test data ?

Check this references:
http://kylin.apache.org/docs/howto/howto_optimize_cubes.html
http://mail-archives.apache.org/mod_mbox/kylin-user/201611.mbox
/%3Ctencent_F5A1E061EFFB778CC5BF9909%40qq.com%3E
http://mail-archives.apache.org/mod_mbox/kylin-user/201607.mbox
/%3C004201d1d4ef%240151b7e0%2403f527a0%24%40fishbowl.com%3E
http://mail-archives.apache.org/mod_mbox/kylin-user/201612.mbox
/%3CCAEcyM171RGhk0QoXJUjjZJeSxXwgUGu0vO%2B_T71KXMU1k00L%2Bg%40mail.gmail.com
%3E
Check this tunning example:  https://github.com/albertoRamon/Kylin
/tree/master/KylinPerformance

BR, Alb


2017-01-16 3:47 GMT+01:00 Phong Pham <phongpham1805@gmail.com>:

> Hi all,
>     Hi all,
>    * We still meet problems with query performance. Here is the cube info
> of one cube*:
> {
>  "uuid": "6b2f4643-72a3-4a51-b9f2-47aa8e1322a5",
>  "last_modified": 1484533219336,
>  "version": "1.6.0",
>  "name": "account_global_convtrack_summary_daily_test",
>  "owner": "ADMIN",
>  "descriptor": "account_global_convtrack_summary_daily_test",
>  "cost": 50,
>  "status": "READY",
>  "segments": [
> {
>  "uuid": "85fa970e-6808-47c8-ae35-45d1975bb3bc",
>  "name": "20160101000000_20161226000000",
>  "storage_location_identifier": "KYLIN_7E4KIJ3YGX",
>  "date_range_start": 1451606400000,
>  "date_range_end": 1482710400000,
>  "source_offset_start": 0,
>  "source_offset_end": 0,
>  "status": "READY",
>  "size_kb": 9758001,
>  "input_records": 8109122,
>  "input_records_size": 102078756,
>  "last_build_time": 1484533219335,
>  "last_build_job_id": "a4f67403-17cb-4474-84d1-21ad64ed17a8",
>  "create_time_utc": 1484527504660,
>  "cuboid_shard_nums": {},
>  "total_shards": 4,
>  "blackout_cuboids": [],
>  "binary_signature": null,
>  "dictionaries": {
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/CITYID":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> CITYID/0015e15c-9336-4040-b8ad-b7afba71d51c.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/TYPE":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> TYPE/56cc3576-3c19-40fb-8704-29dba88e3511.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/NETWORKID":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> NETWORKID/edc1b900-8b8a-4834-a8ab-4d23e0087d61.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/WEEKGROUP":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> WEEKGROUP/3c3ae7e2-05a0-49a3-b396-ded7b1faaebd.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/DATESTATSBIGINT":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> DATESTATSBIGINT/b2003335-f10c-48b5-ac98-6d2ddd25854b.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/COUNTRYID":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> COUNTRYID/233a3b35-9e0f-46e3-bb01-3330c907ab33.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/ACCOUNTID":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> ACCOUNTID/612d8a57-8ed8-4fdd-bf99-c64fb2a583fe.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/DEVICEID":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> DEVICEID/8813544c-aac3-4f26-849b-3e3d1b71d9e2.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/LANGUAGEID":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> LANGUAGEID/02dea027-86cf-44e6-9bcf-9dbd4c33e54b.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/COMPANYID":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> COMPANYID/75a5566e-b419-4fc8-9184-757b207a35d2.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/REGIONID":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> REGIONID/81d5b463-8639-4633-83b9-9ac9e43e32cb.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_
> SUMMARY_DAILY_ORC/AFFILIATEID": "/dict/METRIXA_GLOBAL_
> DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> AFFILIATEID/0a35d5ce-dabb-4e32-ad5f-b87ef4c18ee3.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/SITEID":
> "/dict/MTX_SYSTEM.TBL_CONVTRACK_SITES_ORC/SITEID/07e4f091-f6aa-4520-9069-
> 416ee4c904de.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/MONTHGROUP":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> MONTHGROUP/e3bf45aa-3ff3-477b-aafd-d2c38a70caea.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/DATESTATS":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> DATESTATS/5a3d3dc6-90eb-493b-84d0-b1b8ca8b70ec.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/ISMOBILEDEVICE":
> "/dict/METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> ISMOBILEDEVICE/eba9f8db-c5f0-4283-8a77-5f72d75c5867.dict",
> "METRIXA_GLOBAL_DATABASE.ACCOUNT_GLOBAL_CONVTRACK_
> SUMMARY_DAILY_ORC/SOURCEURLID": "/dict/METRIXA_GLOBAL_
> DATABASE.ACCOUNT_GLOBAL_CONVTRACK_SUMMARY_DAILY_ORC/
> SOURCEURLID/3f90d0de-6d04-4bc6-af20-0030a91326f0.dict"
>  },
>  "snapshots": {
> "MTX_SYSTEM.TBL_MCM_COUNTRY_CITY_ORC": "/table_snapshot/MTX_SYSTEM.
> TBL_MCM_COUNTRY_CITY_ORC/f32ec683-f83f-423a-820e-1bfd4b65426f.snapshot",
> "METRIXA_GLOBAL_DATABASE.GLOBAL_SOURCEURL_ORC": "/table_snapshot/METRIXA_
> GLOBAL_DATABASE.GLOBAL_SOURCEURL_ORC/32e8df3f-7188-4646-9eff-6c96792897f4.
> snapshot",
> "MTX_SYSTEM.TBL_MCM_COUNTRY_REGION_ORC": "/table_snapshot/MTX_SYSTEM.
> TBL_MCM_COUNTRY_REGION_ORC/e4378b9c-ff08-4207-92fa-3f0cf37f00d5.snapshot",
> "MTX_SYSTEM.TBL_MCM_COUNTRY_ORC": "/table_snapshot/MTX_SYSTEM.
> TBL_MCM_COUNTRY_ORC/2f2ffb19-d675-43a2-bb08-66a83801f875.snapshot",
> "MTX_SYSTEM.GLOBAL_ACCOUNT_SEARCH_ENGINE_ORC":
> "/table_snapshot/MTX_SYSTEM.GLOBAL_ACCOUNT_SEARCH_ENGINE_
> ORC/53ef6022-7249-4ef8-8518-b7d84c65fdfa.snapshot",
> "MTX_SYSTEM.TBL_CONVTRACK_SITES_ORC": "/table_snapshot/MTX_SYSTEM.
> TBL_CONVTRACK_SITES_ORC/0cbb0323-d434-44de-8891-85b024589743.snapshot",
> "MTX_SYSTEM.TBL_MCM_LANGUAGE_ORC": "/table_snapshot/MTX_SYSTEM.
> TBL_MCM_LANGUAGE_ORC/957e6a54-c618-4e5c-bc8d-c89952cafe1e.snapshot",
> "MTX_SYSTEM.TBL_CONVTRACK_AFFILIATES_ORC": "/table_snapshot/MTX_SYSTEM.
> TBL_CONVTRACK_AFFILIATES_ORC/f794bce2-dcb1-41b0-b9bf-
> fe3c9e1ad661.snapshot"
>  },
>  "index_path": "/kylin/kylin_metadata/kylin-a4f67403-17cb-4474-84d1-
> 21ad64ed17a8/account_global_convtrack_summary_daily_clone/
> secondary_index/",
>  "rowkey_stats": [
> [
>  "DATESTATS",
>  360,
>  2
> ],
> [
>  "CITYID",
>  60804,
>  2
> ],
> [
>  "SOURCEURLID",
>  38212,
>  2
> ],
> [
>  "REGIONID",
>  2822,
>  2
> ],
> [
>  "COUNTRYID",
>  238,
>  1
> ],
> [
>  "LANGUAGEID",
>  173,
>  1
> ],
> [
>  "AFFILIATEID",
>  36,
>  1
> ],
> [
>  "ACCOUNTID",
>  62,
>  1
> ],
> [
>  "COMPANYID",
>  19,
>  1
> ],
> [
>  "SITEID",
>  103,
>  1
> ],
> [
>  "WEEKGROUP",
>  52,
>  1
> ],
> [
>  "MONTHGROUP",
>  12,
>  1
> ],
> [
>  "TYPE",
>  2,
>  1
> ],
> [
>  "ISMOBILEDEVICE",
>  2,
>  1
> ],
> [
>  "DEVICEID",
>  338,
>  2
> ],
> [
>  "NETWORKID",
>  161,
>  1
> ],
> [
>  "DATESTATSBIGINT",
>  360,
>  2
> ]
>  ]
> }
>  ],
>  "create_time_utc": 1484286587541,
>  "size_kb": 9758001,
>  "input_records_count": 8109122,
>  "input_records_size": 102078756
> }
> *+ We have 2 colums that is high cardinality*: [
>  "CITYID",
>  60804,
>  2
> ],
> [
>  "SOURCEURLID",
>  38212,
>  2
> ],
> *+ We define left-join from model for all relations*
> *+ With new aggregation:*
>         Includes
> ["SITEID","COMPANYID","SOURCEURLID","DATESTATS","WEEKGROUP","MONTHGROUP","
> COUNTRYID","REGIONID","TYPE","ISMOBILEDEVICE","LANGUAGEID","
> DEVICEID","NETWORKID","ACCOUNTID","AFFILIATEID","CITYID"]
>
> Mandatory Dimensions
> ["DATESTATS"]: Because we always use datestats as a filter
>
> Hierarchy Dimensions: None < Maybe wee will put CountryId, RegionId, and
> CityId
> Joint Dimensions
> ["LANGUAGEID","ACCOUNTID","AFFILIATEID","SITEID","CITYID"
> ,"REGIONID","COUNTRYID","SOURCEURLID"]: Please explain to me more about
> join dimensions? I don't understand fully about this theory.
> *+ Rowkeys:*
> We defined all rows is dict, because all of them are not ultra high
> cardinality
>
> The query that is very slow is that:
> + We get all dims and metrics, left join all dim tables and group by all
> dims
> + We set datetstats condition for 1 year
>
> And query often take a long time to executed: >10s
>
> So are there problems with our cube design? I would like to hear your
> reply soon.
> Thanks so much for your help.
>
> 2017-01-12 21:28 GMT+07:00 ShaoFeng Shi <shaofengshi@apache.org>:
>
>> Obviously there are too many segments (24*3=72), try to merge them as
>> Billy suggested.
>>
>> Secondly if possible try to review and optimize the cube design
>> (especially the rowkey sequence, put high-cardinality filter column to the
>> begin position to minimal the scan range), see
>> http://www.slideshare.net/YangLi43/design-cube-in-apache-kylin
>>
>> Thirdly try to give more power to the cluster, e.g use physical machines;
>> and also use multiple kylin query nodes to balance the concurrent work
>> load.
>>
>> Just some cents, hope it can help.
>>
>> 2017-01-12 22:16 GMT+08:00 Billy Liu <billyliu@apache.org>:
>>
>>> I have concerns with so many segments. Please try query only one cube
>>> with one segment first.
>>>
>>> 2017-01-12 13:36 GMT+08:00 Phong Pham <phongpham1805@gmail.com>:
>>>
>>>> Hi,
>>>> Thank you so much for your help. I really appreciate it. Im really
>>>> impressed with your project and trying to apply it to our product. Our live
>>>> product is still working on Mysql and MongoDb, but data is growing fast.
>>>> That's why we need your product for the database engine replacement.
>>>> About our problem with many queries on same time on Apache Kylin, I'm
>>>> trying to monitor some elements on our system and review cubes. So are
>>>> there some tutorials about concurrency of Kylin or HBase?
>>>> I will give you more details abour our system:
>>>> Hardware:
>>>> 2 physical machines -> 7 vitural machines
>>>> Each vitural machine:
>>>> CPU: 8cores
>>>> RAM: 24GB
>>>> We are setup hadoop env with  hortonwork 2.5 and setup HBase with 5
>>>> RegionServer, 2 Hbase masters
>>>> Apahce Kylin we setup on 2 machines:
>>>> + Node 1: using for build cubes
>>>> + Node 2: using for only queries (this node also contain RegionServer)
>>>> Cube and Queries:
>>>> + Size of Cubes:
>>>>   - Cube 1: 20GB/14M rows - 24 segments (maybe we need to meger them
>>>> into 2-3 segments)
>>>>   - Cube 2: 460MB/3M rows - 24 segments
>>>>   - Cube 3: 1.3GB/1.4M rows - 24 segments
>>>> + We use one query to read data from 3 cubes and union all into 1 result
>>>> Test case:
>>>> + On single request: 3s
>>>> + On 5 requests on same times: (submit multi-requests from client):
>>>> 20s/request
>>>> And that is not acceptable when we go live.
>>>> So hope you all review our struture and give us some best pratices with
>>>> Kylin And Hbase.
>>>> Thanks
>>>>
>>>> 2017-01-12 8:24 GMT+07:00 ShaoFeng Shi <shaofengshi@apache.org>:
>>>>
>>>>> In this case you need do some profiling to see what's the bottleneck:
>>>>> Kylin or HBase or other factors like CPU, memory or network; maybe it
is
>>>>> related with the cube design, try to optimize the cube design with the
>>>>> executed query is also a way; It is hard to give you good answer with
a
>>>>> couple words.
>>>>>
>>>>> 2017-01-11 19:50 GMT+08:00 Phong Pham <phongpham1805@gmail.com>:
>>>>>
>>>>>> Heres about detail on our system:
>>>>>>
>>>>>> Hbase: 5 nodes
>>>>>> Data size: 24M rows
>>>>>>
>>>>>> Query result:
>>>>>> *Success: true*
>>>>>> *Duration: 20s*
>>>>>> *Project: metrixa_global_database*
>>>>>> *Realization Names: [xxx, xxx, xxx]*
>>>>>> *Cuboid Ids: [45971, 24]*
>>>>>>
>>>>>>
>>>>>> 2017-01-11 18:34 GMT+07:00 Phong Pham <phongpham1805@gmail.com>:
>>>>>>
>>>>>>> Hi all,
>>>>>>>     I have a problem with concurrency on Apache Kylin. Execute
>>>>>>> single query, it takes about 3s. Howerver,when i run multiple
queries on
>>>>>>> the same time, each query take about 13-15s. So how can i solve
problems?
>>>>>>> My Kylin Version is 1.6.1
>>>>>>> Thanks
>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> Best regards,
>>>>>
>>>>> Shaofeng Shi 史少锋
>>>>>
>>>>>
>>>>
>>>
>>
>>
>> --
>> Best regards,
>>
>> Shaofeng Shi 史少锋
>>
>>
>

Mime
View raw message