Skip to Content

Introduction

I’d like to share my experience in developing “pseudo-realtime” in BW systems without HANA with large data volume.

I met this task on one retail project. The client wanted to update data in analytic reports every 15-20 minutes – main transactional data including checks from POS-system, movements from ERP, stocks, sales and so on.

Problem description

If we have a BW system on one of traditional DB (like Oracle, MSSQL, Sybase and so on) containing large volume of data (e.g. 50 mln lines of transactional data in a year), such task will bring us several serious problems:

  1. Index reload takes a long time, it can run during 10 and more minutes. During this time the speed of reports reduces heavily.
  2. Aggregates updating takes a long time too, and during this process data in reports can be unavailable for some time.
  3. Together reloading indexes and updating aggregates can takes much more time than required 20 minutes.

The task is even more complicated if we use the “slice scheme” of stock infocubes instead of non-cumulative. In this scheme we store stocks on every day instead of calculating them “on the fly” via non-cumulative.  Such method of storing stocks is quite popular in retail projects because it provides high speed of reading stock data, but it leads to huge number of lines in the infocube.

Solution

So, the idea of my solution was in storing data of the current day in separate infocubes. For example, we have a big infocube with POS data ZRT_C01, it contains 200 mln lines of data and has several aggregates.

We make a copy of this infocube ZRT_C02 with the same transformation to it. It’s useful to make an info-source and to put all logic before the infosource, this allows changing logic once in case of some changes.

The second infocube ZRT_C02 contains only today’s data. During the day we load delta in this cube. It has quite small volume of data so we don’t need to make aggregates on it. We can load data in one parallel process without index rebuilding, but anyway, index rebuilding doesn’t take a long time in such case.

Further, during the night data loading, we do the following:

  1. Load today’s data again from PSA (or some intermediate DSO) to the main infocube ZRT_C01, with index rebuilding, updating of aggregates and so on. It takes a long time, but at night it is not a problem.
  2. Delete data in secondary cube ZRT_C02.
  3. Initialize delta for the secondary cube ZRT_C02. This step is very important, because we want that the next delta-loading will load only tomorrow’s delta but not all data from PSA.

The 3rd step should be additionally commented, because it can’t be realized via standard elements of process chain. In this step we need to call data transfer process in the “Initialization” mode.

We can do it manually – to set the DTP mode to “No Data Transfer; Delta Status in Source: Fetched”, but there is no standard process for that in process chains. To automate it we need to write an ABAP-program and to include it to process chain. This program is described, for example, in this blog: https://blogs.sap.com/2014/08/22/automate-dtp-with-init-without-data-transfer-processing-mode/.

So, we have two main process chains:

  1. Night process chain. It loads master data and transactional data which are not critical for “real-time”. It loads delta to main infocubes (like ZRT_C01 in my example), delete data in the secondary infocubes, initialize delta in them and so on.
  2. Day “real-time” process chain. It works every 15-20 minutes and doesn’t contain operations with indexes and aggregates.

Both infocubes are included to one multi-cube and work at the same time. The main cube provides all data before today, the secondary cube provides today’s data.

In case of my project, such method allowed to make a pseudo-realtime with updating of critical transactional data every 15 minutes and without losses in report’s speed.

To report this post you need to login first.

2 Comments

You must be Logged on to comment or reply to a post.

  1. anuj pandit

    September  2017  calendar with holidays

    The task is even more complicated if we use the “slice scheme” of stock infocubes instead of non-cumulative. In this scheme we store stocks on every day instead of calculating them “on the fly” via non-cumulative.  Such method of storing stocks is quite popular in retail projects because it provides high speed of reading stock data, but it leads to huge number of lines in the infocube.

    October  2017  calendar with holidays

    November  2017  calendar with holidays

     

    (0) 
  2. Thomas Ward

    why can’t you use the delta once method for your secondary cube – hence removing your daily initialisation ?

     

    The safest option would be to load from your secondary cube to the main cube within the batch schedule.   Then delete the data in the secondary cube.  Using the delta once load into secondary cube again.

    This would remove any possibility of having the same records in both cubes.

    (0) 

Leave a Reply