The Power and Simplicity of the Data Warehouse

“In many ways, dimensional modeling amounts to holding the fort against assaults on simplicity”
– Ralph Kimball, The Data Warehouse Toolkit

 

Although there are many reasons that an organization may consider building a “data warehouse”, most of the time the overwhelming reason is performance related… a report or web page is just taking too long to run and they want it to be faster. If that data could be pre-computed than that page would be really fast!

I’m here to tell you that speed is not the most important reason for building the warehouse.

When you’ve got a system where multiple consumers are reading the data from the transactional store and doing their own calculations, you create a whole bunch of other problems beyond just the speed issue:

  • You create the potential for multiple different and conflicting results in that system. At least a few of those consumers will inevitably do something wrong.
  • You put a considerable burden on your transactional system to recalculate that data over-and-over again for each custom request.
  • While those consumers are running their long running queries, that data is being simultaneously updated by a multitude of data collection and transformative processes… you are not only reading inconsistent sets of data in the consumers, you are blocking the collection and transformation processes from doing their job, getting in their way and slowing them down… and sometimes even causing them to fail.
  • You’ve created a multitude of intertwined dependencies in that system. This makes attempts to improve or otherwise change that transactional system extremely difficult to achieve without breaking things… and having to make massive system wide changes to accommodate even the smallest change.
  • The bottom line is this: You’ve just got a greatly over-complicated system that is much more prone to performance problems and defects. As Ralph Kimball states so eloquently, data warehouse efforts are a giant move towards simplicity. And simpler systems are better systems.

We recently launched a major warehouse initiative to, once and for all, pre-compute and store all our portfolio-level data. Although that data is already computed from data that’s been pre-aggregated at the account level, there is still considerable additional work required to aggregate that data further to the portfolio level.

Primarily, that’s a major performance problem. Pulling a single portfolio’s data can take as long as 5-7 minutes for larger portfolios. That’s a serious problem with our scalability and an overall burden on our system.

I’m happy to report that the portfolio warehouse initiative is nearing its conclusion and am confident it will do things for us far beyond the performance improvements we hoped to gain:

  • With every consumer pulling from the same warehouse for portfolio level information, we can guarantee they will get the same results… they are all “drinking water from the same well.”
  • The portfolio data can now be processed “incrementally”… i.e. rather than have to recalculate that data from the beginning of time for every request, we can reprocess only that data that has changed. This pays huge dividends on overall system performance and greatly decreases the burden on the system.
  • Our data will now be pulled from a snapshot-enabled data warehouse. This guarantees clean and consistent reads without blocking the transactional store from doing its work.
  • By having one system that reads transactional data and compiles and stores the portfolio data, we only have one system to change when we want to change something in the transactional store. This is hugely liberating to us when we want to modify those underlying systems.
  • The new published warehouse structure for portfolios is simple and easy to understand. It therefore opens up consumption of that data in new ways with less effort, opening the doors to new possibilities that were otherwise impossible. Looking at data for all the portfolios in a firm in one pass is now possible, performing cross-firm analytics that were unthinkable before. This opens a myriad of optionsfor us that we intend to take advantage of.
  • Oh, and the speed is nice also… it’s really fast!

While we are still in the final stages of implementation, we hope to bring this system fully into production over the next few months and are very excited about the possibilities… we hope you are too!

And if you’d like to read about data warehousing, I highly recommend what is, in my opinion, the bible of data warehousing:

The Data Warehouse Toolkit – By Ralph Kimball and Margy Ross

Fortigent

“In many ways, dimensional modeling amounts to holding the fort against assaults on simplicity”
– Ralph Kimball, The Data Warehouse Toolkit


Although there are many reasons that an organization may consider building a “data warehouse”, most of the time the overwhelming reason is performance related… a report or web page is just taking too long to run and they want it to be faster. If that data could be pre-computed than that page would be really fast!

I’m here to tell you that speed is not the most important reason for building the warehouse.

When you’ve got a system where multiple consumers are reading the data from the transactional store and doing their own calculations, you create a whole bunch of other problems beyond just the speed issue:

  • You create the potential for multiple different and conflicting results in that system. At least a few of those consumers will inevitably…

View original post 587 more words

The coolest SQL Server function you never heard of

Ever heard of the SQL_VARIANT_PROPERTY function? I didn’t think so.

SQL Server developers very often make the mistake of making their NUMERIC fields too large. When faced with a choice of how to size the column, they’ll often think “make it way larger than I need to be safe”.

This works OK as long as you simply store and read these values, but if you ever have to perform math with these columns, particularly some form of division or multiplication, you may find your results mysteriously losing precision.

This is because SQL Server can only store a maximum of 38 digits per number… if the results of your mathematic expression may yield a number larger than that, SQL Server will be forced to downsize it and remove digits from the mantissa as a result.

Continue reading