Using Partitioning for Data Warehousing at TriTUX.com
TriTUX is based in Tunis, Tunisia. Founded in 2006 as Systems Analysis and Weblication Development, TriTUX is providing high-quality professional services for achieving flexible, creative and scalable systems. TriTUX engineers is a team of jack-of-all-trades, possessing intimate knowledge in networking, eXtreme programming, GNU/Linux and scalable solutions.
Our software helps organizations to gain better insight into their business, improving decision-making and enterprise performance.
Purpose of our application
We've been using partitioning at TriTUX.com since March 2008 and we're currently using MySQL 5.1.24.
Our application is used by a Telecom Operator, to calculate and track bonus points given to subscribers for incoming calls. The application is built on a strongly partitioned MySQL database to handle 60 million calls (INSERTs) per day and to deliver a daily per-subscriber call performance-meter.
How partitioning meets our needs
We realized that a simple SELECT for SUM() on a table with billions of rows would take an eternity. By partitioning our tables, we can now perform these querys in less than 10 seconds! We use both RANGE (by month/day) and HASH partitioning schemes. We also utilize MySQL triggers to perform updates on the partitioned summary tables.
Choosing and tuning the hardware was a difficult task too, sizing memory, CPU frequency and hard disks quality were the most time-consuming issues, along side with tuning MySQL's configurations to get the most out of it.
Our development platform is Ubuntu Linux. We develop in Java and PHP, using the Talend Open Studio ETLs development environment.
Our production system consists of one NAS connected to 6 HP Blade 64-bit multicore servers that run Red Hat Linux. Each of them is equipped with 32GB of RAM. After performing some benchmarking, we decided to use the InnoDB storage engine on top of an XFS file system. InnoDB was chosen because of its advantages to MyISAM when it comes to concurrency and data integrity. Backups are performed by taking LVM snapshots of the file systems.
Here are some relevant metrics about the size and type of our application:
- 60 million INSERTS per day (number of calls per day)
- 6 million subscribers
- The database is growing ~4GB per day
- We're using InnoDB with LVM snapshotting for hot backup
We did some software tuning to get the most of a 64bits multicores architecture such as using massive multithreaded software and get the most of the huge RAM space.
At TriTUX.com, we believe in the power of mature open source projects, so rejecting commercial databases from the beginning was a matter of love addiction to other alternatives. In our opinion, PostgreSQL lacks a bit when it comes to partitioning (in version 8.3) and it does not provide some of the advanced partitioning options that are available with MySQL 5.1.
Some additional information about how and why we selected MySQL with Partitioning for our requirements can be found on my blog here: http://tuxmining.blogspot.com/2008/05/mysql-51-2-weeks-of-stresstests.html
What we also do with the new features in MySQL
In addition to the application above, we also utilize MySQL in several other projects we are working on.
We're continuously working on improving our SOA middleware, called PodBridge which is based on a MySQL server (and can be configured to use PostgreSQL too), the event scheduler and partitioning were features which mostly helped us in designing robust and reliable software.
Replication is our key in succeeding real time synchronization between 3 distant sites for our ERP, called Alvanet and dedicated to automotive production companys. Replicating databases lets our customer have real time access to what's happening on the other side of the planet, in another site of the company without depending on bandwidth and internet connection quality, and especially without deploying a centralized platform for that purpose.
Replication and partitioning were also our keys to success in deploying our mobile banking software, called MobiBank.