I have a similar data pattern and 100ms response time is fairly consistent.
I’ve been trying hard to find the right set of configs to get closer to 10-20ms with no luck, but I’m finding that 100ms average is pretty reasonable. From: Willem Conradie [mailto:willem.conra...@pbtgroup.co.za] Sent: Wednesday, January 20, 2016 8:31 AM To: jamestay...@apache.org Cc: user@phoenix.apache.org Subject: RE: Telco HBase POC Hi James, Thanks for being willing to assist. This is what the input data record will look like (test data) : UserID DateTime TXNID DeviceID IPAddress UsageArray URIArray 12345678901 20151006124945 992194978 123456789012345 111.111.111.111 10-4000:26272:1019324|0-4000:0:0|10-4000:25780:498309|420-4000:152:152|500-500:1258:2098|9001-9001:120:0|0-4000:0:0|502-4000:154:0|10-4000:73750:448374|420-4000:608:608|1-4000:364:550|358-4000:40:52 www.facebook.com|www.whatsapp.com|www.google.co.nz|ssl.gstatic.com|www.google.com<http://www.facebook.com|www.whatsapp.com|www.google.co.nz|ssl.gstatic.com|www.google.com/> Unique key on record is “UserID,DateTime,TXNID”. Read access pattern is as follows: User queries by UserID, DateTime range to supply them with usage stats (derived from ‘UsageArray’) for websites (derived from ‘URIArray’) visited over their selected time range. Just to recap the data volumes: Expected data volume : 60 000 files per day 1 –to 10 MB per file 500 million records per day 500 GB total volume per day I need to be flexible in the amount of data stored. Initially it will be 5 days, but can increase to 30 days and possibly 90 days. One concern I have (not founded in any way) with the phoenix client is whether it will be able to support data access for above queries within 100ms range. Regards, Willem From: James Taylor [mailto:jamestay...@apache.org] Sent: 19 January 2016 10:07 PM To: user <user@phoenix.apache.org<mailto:user@phoenix.apache.org>> Subject: Re: Telco HBase POC Hi Willem, Let us know how we can help as you start getting into this, in particular with your schema design based on your query requirements. Thanks, James On Mon, Jan 18, 2016 at 8:50 AM, Pariksheet Barapatre <pbarapa...@gmail.com<mailto:pbarapa...@gmail.com>> wrote: Hi Willem, Use Phoenix bulk load. I guess your source is csv so phoenixcsvbulk loader can be used. How frequently you want to load these files. If you can wait for certain interval to merge these files and map reduce will bulk load to Phoenix table. Cheers Pari On 18-Jan-2016 4:17 pm, "Willem Conradie" <willem.conra...@pbtgroup.co.za<mailto:willem.conra...@pbtgroup.co.za>> wrote: Hi Pari, My comments in blue. Few notes from my experience : 1. Use bulk load rather than psql.py. Load larger files(merge) instead of small files. Are you referring to native HBase bulk load or Phoenix MapReduce bulk load? Unfortunately we can’t change how the files are received from source. Must we pre-process to merge the files before running the bulk load utility? 2. Increase HBase block cache 3. Turn off HBase auto compaction 4. Select primary key correctly 5. Don't use salting . As table will be huge, your phoenix query will fork may scanners. Try something like hash on userid. 6. Define TTL to purge data periodically Regards, Willem From: Pariksheet Barapatre [mailto:pbarapa...@gmail.com<mailto:pbarapa...@gmail.com>] Sent: 15 January 2016 03:17 PM To: user@phoenix.apache.org<mailto:user@phoenix.apache.org> Subject: Re: Telco HBase POC Hi Willem, Looking at your use case. Phoenix would be a handy client. Few notes from my experience : 1. Use bulk load rather than psql.py. Load larger files(merge) instead of small files. 2. Increase HBase block cache 3. Turn off HBase auto compaction 4. Select primary key correctly 5. Don't use salting . As table will be huge, your phoenix query will fork may scanners. Try something like hash on userid. 6. Define TTL to purge data periodically Cheers Pari On 15 January 2016 at 17:48, Pedro Gandola <pedro.gand...@gmail.com<mailto:pedro.gand...@gmail.com>> wrote: Hi Willem, Just to give you my short experience as phoenix user. I'm using Phoenix4.4 on top of a HBase cluster where I keep 3 billion entries. In our use case Phoenix is doing very well and it saved a lot of code complexity and time. If you guys have already decided that HBase is the way to go then having phoenix as a SQL layer it will help a lot, not only in terms of code simplicity but It will help you to create and maintain your indexes and views which can be hard&costly tasks using the plain HBase API. Joining tables it's just a simple SQL join :). And there are a lot of more useful features that make your life easier with HBase. In terms of performance and depending on the SLAs that you have you need to benchmark, however I think your main battles are going to be with HBase, JVM GCs, Network, FileSystem, etc... I would say to give Phoenix a try, for sure. Cheers Pedro On Fri, Jan 15, 2016 at 9:12 AM, Willem Conradie <willem.conra...@pbtgroup.co.za<mailto:willem.conra...@pbtgroup.co.za>> wrote: Hi, I am currently consulting at a client with the following requirements. They want to make available detailed data usage CDRs for customers to verify their data usage against the websites that they visited. In short this can be seen as an itemised bill for data usage. The data is currently not loaded into a RDBMS due to the volumes of data involved. The proposed solution is to load the data into HBase, running on a HDP cluster, and make it available for querying by the subscribers. It is critical to ensure low latency read access to the subscriber data, which possibly will be exposed to 25 million subscribers. We will be running a scaled down version first for a proof of concept with the intention of it becoming an operational data store. Once the solution is functioning properly for the data usage CDRs other CDR types will be added, as such we need to build a cost effective, scalable solution . I am thinking of using Apache Phoenix for the following reasons: 1. 1. Current data loading into RDBMS is file based (CSV) via a staging server using the RDBMS file load drivers 2. 2. Use Apache Phoenix bin/psql.py script to mimic above process to load to HBase 3. 3. Expected data volume : 60 000 files per day 1 –to 10 MB per file 500 million records per day 500 GB total volume per day 4. 4. Use Apache Phoenix client for low latency data retrieval Is Apache Phoenix a suitable candidate for this specific use case? Regards, Willem -- Cheers, Pari