Data Warehousing and Business Intelligence

24 April 2017

Choosing between Big Data and Data Warehousing

Filed under: Data Architecture,Data Warehousing — Vincent Rainardi @ 4:14 am

If we have 100 files each containing 10 million rows that we need to load to a repository so that we can analyse the data. What should we do? Do we put them into Hadoop (HDFS), or into a database (RDBMS)?

Last week I defined the difference between Big Data and Data Warehousing as: Big Data is Hadoop, and Data Warehousing is RDBMS. See my article here: link. Today I would like to illustrate using an example where we need to choose between the two.

There are 4 consideration factors:

  1. Data Structure
  2. Data Volume
  3. Unstructured Data
  4. Schema on Read

1. Data Structure: Simple vs Complex
If all 100 files have the same structure, and they all consist of the same 10 columns, then it is better to put them into Hadoop. We can then use Hive* and R to analyse the data. For example, to find patterns within the data, doing statistical analysis, or create forecasts. The development time will be shorter, because it is only 1 layer.
*or Phoenix, Impala, BigSQL, Stinger, Drill, Spark, depending on your Hadoop platform

If the 100 files contain 100 different tables, it is better to put them into a database, create a data warehouse, and use a Analytic/BI tool such as MicroStrategy or Tableau to analyse the data. For example to slice and dice the data, find percentage or anomalies and time series analysis. Yes we need to create 3 layers (staging, 3NF, star schema) but it enables us analyse each measure by different dimensions.

So if the data structure is simple, put into Hadoop, and if the structure is complex, but into a data warehouse. This is the general rule, but there are always exceptions. Can data with simple pattern be put into a data warehouse? Of course it can. Can data with complex pattern be put into Hadoop? Of course it can.

Using Hadoop and Hive/R we can also do slice and dice, find percentage or anomalies and do time series analysis. Using a data warehouse we can also do machine learning and data mining to find patterns in the data, do statistical analysis, and create forecasts. So basically, whether we store the data in Hadoop or in a data warehouse, we can still do complete analysis.

The issue here is storing it. Linking 100 tables in Hadoop is difficult and not natural. RDBMS such as SQL Server or Oracle is designed precisely for that task: linking and joining tables. Constructing a data model linking 100 tables is very suitable for an RDBMS. Can we design a data model linking 100 files with different structures in Hadoop? Of course we can. But it is much more difficult. For starters, it is Schema-on-Read, so the columns in the files have no data types. Schema-on-Read means that we don’t try to understand the relationship between the files when loading them into Hadoop. So yes we can load the 100 files into Hadoop, but we keep them as individual files, without relationships between them. This is the same as in Data Lake, which is also using Schema-on-Read, also using HDFS.

  1. Data Volume: Small vs Large

100 files containing 10 million rows each is 1 billion rows per day. If all 100 files have the same structure (say they all consists of the same 10 columns), then we will have a performance problem if we load them into an SMP database such as SQL Server or Oracle. Within 3 years, this table will have about 1 trillion rows. Even with partitioning and indexing, it will still be slow to query.

Hadoop on the other hand, will have no problem storing and querying 1 trillion rows. It is designed exactly for this task, by storing it in many files and querying it in parallel using Stinger, Drill, Phoenix, Impala or Spark. The file structure is simple (the same 10 columns each) which lends itself to Hadoop.

Teradata, Greenplum, Netezza, Exadata, and Azure SQL Data Warehouse are more than capable to handle this, with excellent query performance. But MPPs are more costly than Hadoop, which is why companies tend to choose Hadoop for this task. Using an MPP for this task is like killing a fly with a canon. Not only it is expensive and unnecesary, but also it is too sluggish and cumbersome for the task.

If the 100 source files have a complex structure (such as an export from SAP system) then yes an MPP is a suitable solution as we need to create relationship between the files/tables. But if the source files have a simple structure and we just need to union them, then Hadoop is more suitable and more economical for this task.

So if the data volume is large, like 1 billion per day, and the data structure is simple, put them into Hadoop. But if the data volume is large and the data structure is complex, put them into an MPP.

  1. Unstructured Data

If most of those 100 source files are MP4 (video) or MP3 (music), then Hadoop or Data Lake is an ideal platform to store them. An RDBMS be it SNP or MPP are not designed to store video or music files. They can (as a blob, or as externally-linked files), but they are not really designed for it.

If the source files are have different number of columns (such as Facebook or Twitter files,) then Hadoop or Data Lake is an ideal platform to store them. An RDBMS is not really designed for it.

Unstructured Data can also comes in the form of free-format text files (such as emails) and documents (such as journals and pattents). Again Hadoop or Data Lake is much better position to store them than a RDBMS.

  1. Schema-on-Read

One of the advantages of using Hadoop or Data Lake is that they are Schema-on-Read. Meaning that we just store those files without determining whether the first column is a numeric or a string. When we want to query it then we need to specify the data type.

Why is this an advantage? Because it makes it flexible. In Data Warehousing the first thing we need to do is to analyse the file structure, and design many tables to host the files in a Staging database. Then we design a normalised database to integrate those Staging tables. And then we design a Reporting layer in the form of Fact and Dimension tables and load those normalised tables into them. The whole thing can take a year if we have 100 files. The more number of files we have, the more complex the process and the longer it takes to design the databases for Integration layer and Reporting layer. It is good for the data architect (it gives them a job) but it is not good for the people who pay for this project.

Hadoop on the other hand is Schema-on-Read. After we put these 100 files in Hadoop, we query the first file. And when we query this first file, we specify the data types of each column. We don’t need to touch the other 99 files, yet. And we can already get the benefit. We can analyse the data straight away. On day one! If the other 99 files have the same structure, then we can union them, without extra effort of designing any database, and we can query them straight away. On day two! It is much simpler, we don’t need a team of 10 people designing any Staging, Normalised or Reporting layer for many many months. We can start analysing the data straight away and the project can finish in 2 to 3 months, with only 3 or 4 people. A lot less costly, a lot more agile, and a lot more flexible.

Just

17 April 2017

Definition of Big Data and Data Warehousing

Filed under: Data Architecture,Data Warehousing — Vincent Rainardi @ 5:38 pm
Tags: ,

I’m annoyed that many people label the normal Data Warehousing & BI stuff as “Big Data”. It is not. For me, Big Data is the “Hadoop stuff” (HDFS). If it is in SQL Server, Oracle or Teradata, it is just a normal database, not Big Data.

Big Data is stored on HDFS (Hadoop Distributed File System), not in RDBMS. Oracle, IBM, Microsoft, SAP, Teradata, all use Hadoop to store Big Data. Big Data is queried using MapReduce.

The reason why Big Data can’t be stored in RDBMS is because the format is not tabular. Sometimes it is 2 columns, sometimes it is 200 columns. Like Twitter data. The second reason is because it is too big. Sensors can make 100 measurements in a second, and in a year it could be Petabytes. Web Logs is another example. Tracking the ask and offer price of every transaction in every stock market is another example. Yes we can put Petabytes into SQL Server or Oracle, into Netezza or Teradata, but not at this speed. Hadoop on the other hand is designed exactly to cope with these kind of speed and volume.

Now the usage. What is Big Data Analytics? Big Data Analytics is when we do analytics on Hadoop Data.

Is Fraud Detection Big Data Analytics? Not always. Fraud Detection can be done on a normal Data Warehouse or a database. Is Machine Learning Big Data? Not always. Machine Learning can be done on a normal Data Warehouse or a database. If the Fraud Detection or the Machine Learning is done on data stored in Hadoop, then it is Big Data Analytics.

Even if it is only 200 GB, if it is stored in Hadoop, it is Big Data. Even if the data is 5 Petabyte, if it is stored in an SQL Server database, it is not Big Data, in my opinion.

Even if the data is in tabular format (i.e. columns and rows), if it is stored in Hadoop, it is Big Data. But if it is stored in an Oracle database, it is not Big Data.

Every Big Data architecture that I know uses Hadoop. No companies (or government) implement Big Data on an RDBMS. Or on a non HDFS files. Every single company, every single Big Data case I read implement the Big Data on Hadoop. I may be wrong and would be happy to be corrected. If you know a case which implements Big Data on a non-Hadoop system/architecture, I will grateful if you could let me know, either through comments, or via vrainardi@gmail.com.

 

 

18 September 2016

Rating Dimension

Filed under: Business Knowledge,Data Architecture,Data Warehousing — Vincent Rainardi @ 2:32 am

In an investment banking data warehouse, we have an instrument dimension (aka security dimension). This instrument dimension is used in many fact tables such as trade, risk, valuation, P&L, and performance.

In this instrument dimension we have rating attributes (columns), from the 3 agencies (S&P, Moody, Fitch), an in-house rating and many combination between them. We may have the ratings at both the issuer/obligor level, and at instrument/issue level. We could also have a hierarchy of ratings, e.g. AA+, AA and AA- are grouped into AA. There are separate ratings for cash instruments.

Usual Design

The usual design is to put all rating attributes in the instrument dimension. This is because the granularity of rating is at instrument level. In other words, rating is an attribute of an instrument.

Issues with the Usual Design

The issues with doing that are:

  1. Cash and FX may not be created as an intrument but they can have ratings.
  2. Some fact tables requires the rating outside the context of an instrument.

An example for point 1 is a USD bank balance (settled cash) in a GBP portfolio, which is given a rating of AAA. In the instrument dimension there is no instrument called USD bank balance. The other example is USD cash proceeds which is an unsettled cash.

An example for point 2 is a fact table used for analysing the average ratings of a portfolio across different date. This is called Portfolio Credit Quality analysis. On any given day, a credit portfolio has 3 attributes: highest rating, lowest rating and average rating. The average rating is calculated by converting the rating letter to a number (usually 1 to 27, or using Moody’s rating factor, link), multiply the rating number with the weight of each position, and sum them up. There can be several different criteria, for example: including or excluding debt derivatives, funds, cash, or equity, and doing lookthrough or not.

For example, portfolio X on 16th Sep could have these different attributes:

  1. Average rating: AA+
  2. Highest rating: AAA
  3. Lowest rating: BB-
  4. Average rating excluding cash: AA
  5. Average rating excluding cash and derivatives: BBB+
  6. Average rating excluding cash, derivatives and equity: BBB-
  7. Average rating including funds: AA-
  8. Average rating with lookthrough: A+

So the fact table grain is at portfolio level and date. We therefore cannot use instrument dimension in this fact table, and therefore it is necessary to have rating in its own dimension.

Another example of a fact table which requires rating not at instrument dimension is sectoral credit quality. A credit portfolio consists of many bonds and credit derivatives. Each of these bonds (or CDS) are of certain industry sector. A bond issued by BP is in Energy sector. A bond issued by Santander is in Financial sector. The average rating for each sector is calculated by converting each rating to a number, then multiplying it by the weight and sum up for all positions in that sector.

If we use GICS (link) as an example, and government, on any given day a credit portfolio can have sectoral credit quality like this: (usually excluding cash and equity, but including credit derivatives and debt funds)

  1. Consumer Discretionary: AA+
  2. Consumer Staples: AA-
  3. Energy: A+
  4. Financials:AAA
  5. Health Care: BB-
  6. Industrials: BBB-
  7. Information Technology: AA-
  8. Materials: BBB+
  9. Telecommunication Services: AA
  10. Utilities: BB+
  11. Real Estate: B+
  12. Government: AAA

Above is the average rating for each sector, for a given day, for a particular portfolio.

Design Choices for Rating Dimension

To satisfy the above requirement (fact tables which analyse rating at a level higher than instrument, e.g. at sector level or portfolio level), we can design the rating dimension in several ways:

  1. Rating schemes and hierarchies are created as attributes
  2. Rating schemes are created as rows, with hierarchies as attributes
  3. Rating schemes and hierarchies are created as rows

Before we dive into the details I’d like emphasize that we should not try to find out which approach is the best one, because different projects will require different approaches. So instead of asking “What is the best approach” we should ask “What is the most suitable approach for my project?”

And secondly I’d like to remind us that if the rating data is always used at instrument level than it should reside in the instrument dimension, not in a rating dimension on its own.

Approach 1. Rating schemes and hierarchies are created as attributes

approach1

Of course we should always have RatingKey 0 with description = Unknown. This row is distinctly difference to NR (No Rating). NR means S&P gives the bond a rating of NR. RatingKey 0 means that we don’t have any info about the rating of this bond.

The distinct advantage of this approach is that we have a mapping between different rating schemes. Moody’s Baa3 for example corresponds to S&P’s BBB-.

The Grade column is either Investment Grade (AAA down to BBB-) and High Yield (HY, from BB+ down to D).

Approach 2. Rating schemes are created as rows, with hierarchies as attributes

approach2

The distinct advantage of approach 2 over approach 1 is RatingKey 33 means Aa3, we don’t have to tell it that we meant Moody’s, like in approach 1. In Approach 1, RatingKey 4 can means Aa3 or AA-, depending whether we meant Moody’s or S&P.

The distinct disadvantage of approach 2 compared to approach 1 is we lost the mapping between S&P and Moody’s. In Approach 1 we know that Aa3 in Moody’s corresponds to AA- in S&P, but in Approach 2 we don’t know that. In the majority of the circumstances, this is not an issue, because an instrument with Moody’s rating of Aa3, may not have an S&P rating of AA- any way. Instead, its S&P rating could be AA or BBB+. So the mapping between S&P and Moody’s are the book standard, academic assumption, not the reality in the real world. In the above example, nobody cares whether the corresponding S&P rating for Aa3 is AA-. What the fund managers and analysts want to know is what Moody’s really assigned to that bond.

Approach 3. Rating schemes and hierarchies are created as rows

approach3

The distinct advantage of Approach 3 over Approach 1 and 2 is that RatingKey 31 means AA tier, it is clear and not ambiguous. In Approach 1 and 2, RatingKey 2,3,4 all means AA tier.

The other advantage of Approach 3 is the ability to create short term rating and cash rating, because they are simply different schemes.

Why do we need to keep the Tier column in Approach 3? Because we still need to convert AA+, AA and AA- to AA.

Rating Value (aka Rating Number)

In some projects we need to convert the rating letters to numbers. For this we can create an attribute in the above rating dimension (any of the 3 approaches) called Rating Value. It is usually AAA = 1, AA+ = 2, AA = 3, AA- = 4, and so on. NR (No Rating) and WR (Withdrawn Rating) got 0 rather than 23.

rating-value

Note that unlike in S&P and Moody’s schemes, in Fitch scheme the D rating is differentiated into 3: DDD, DD, and D. So DDD = 22, DD = 23, D = 24.

Miscellaneous

A common mistake is that people assume there are CC+ and CC- rating. There aren’t. Only CCC is split into 3 (CCC+, CCC, CCC-), but CC and C are not.

Sometimes the relationship between a rating in one scheme and another scheme is not a one-to-one mapping. For example, the Moody’s rating of Ca corresponds to 2 ratings in S&P scheme: CC and C. This is why approach 1 is tricky, because the rating does not have one to one relationship between 1 scheme and another. Another example is as I mentioned above, the D rating in S&P corresponds to 3 ratings in Fitch: DDD, DD and D.

20 June 2016

Data Lake vs Data Warehouse

Filed under: Data Architecture,Data Warehousing — Vincent Rainardi @ 6:22 pm

A Data Lake is a storage and analytic system which stores structured and unstructured data from all source systems in the company in its raw form. The data is queried, combined and analysed to get useful patterns and insights. It is built on a cost effetive Hadoop storage infrastructure, and can be dynamically scaled.

A Data Warehouse is a system that retrieves and consolidates data periodically from source systems into a dimensional or normalised data store. It usually keeps years of history and queried for business intelligence or other analytical activities. It is typically updated in batches, not every time a transaction happens in the source system.

Data Lake Advantages

Because data from all source systems are there in the data lake, the main advantage of data lake is it enables the analyst to get analytics and insight from multiple systems. Yes the data is still in their raw form, so would requires some processing, but it is a lot quicker (1-2 week) than building a data warehouse (6-12 months). And that is the second advantage: time to build.

The third advantage of data lake is its ability to store unstructured data such as documents and social media, which we can query and combine with structured data from databases.

The fourth and probably the most powerful business case for a data lake is the cost efficiency. In investment banking world for example, we can store market data (e.g. prices, yields, spreads, ratings) not only for the securities that we hold but all securities in the market, cheaper than if we store it in Oracle or SQL Server.

The fifth advantage of a data lake is the flexibility. Unlike Oracle or SQL Server, data lakes dynamically scales up. If this morning it is 2 TB, in the afternoon it can be 3 TB. As new data arrives, and we need new capacity, we can add storage easily. As we requires more computing power, we can get more computing power there and then, instantly. There is no need to wait a few days for adding a new node in a SQL always-on availability groups, or adding storage to the SAN, or extending the computing power of an Oracle Grid. This means that we do not need to spend a lot of money upfront paying for 3 TB data and 16 processors. We can just start with 300 GB and 4 processors, and expand when required.

Data Lake Disadvantages

The first issue is that the data lake technology is immature. The language which can query across database and unstructure files only has very limited features (each vendor has different language, for Microsoft Azure it is U-SQL). It is probably only 10% of what PL/SQL or T-SQL can do. We can solve this by putting QlikView or Tableau on top. We use U-SQL only to query individual table and we join the data in QlikView/Tableau and do further processing there.

The second issue is cost (but I don’t find this argument holds water). The issue is: it actually costs a lot of money to store data from all source systems in the company, let alone storing external market data which requires a lot of licence cost.

Let’s take Microsoft Azure Data Lake for pricing example. Being in the cloud, the price is only US$ 0.04 per GB and US$0.07/1 million transactions (ref: link). Let’s say that “data from ALL systems in the company” is 500 GB. And every day we store this data in the data lake. So 500 GB per day x 365 days = 182,500 GB per year x $0.04/GB = $7300 per year. Let’s say we have 10 million transactions per day, which is 3650 million transactions per year x $0.07/million transaction = $256 per year. So it is only cost us about $7500 per year. This is a very reasonable price to pay, to have “all the data in the company in one place”. Even 5 years later, when the volume grows to 5 times, it is only about $35,000 per year. Still very reasonable price to pay.

The third issue is performance. Because data lakes stores the data in its raw format, the query which joins the different data could be running like a dog. Luckily Data Lake is running on HDFS, a distributed file system, which is very fast. So yes it is slower than a data warehouse, but it is not too bad. We are not talking about 30 minutes to run a query, but something like 20-30 seconds (compared to DW query which is say 1-3 seconds).

Data Lake being “exploration tool”, I can tolerate a little bit of slowness. After we proof that (for example) the risk data we query is making money, then it can pay for creating a Risk data mart specifically for that risk analysis purpose.

The forth issue is skill. Data Lake requires a superb business knowledge. Because the analyst needs to join tables from different source system. What is the difference between Real Modified Duration and Effective Real Duration and Modified Duration? But this issue is the same whether we are building ad Data Warehouse or a Data Lake. Both requires good business knowledge.

I don’t find U-SQL is difficult to understand for people with SQL knowledge, which is a common skill. But how about Machine Learning? That is difficult to master, right? Yes, that is true, but it is worth paying an expert data scientist to discover the insight in the data, because this insight can be used to save cost or boost our revenue. The potential benefit is 10-100 times the salary of the data scientist.

Conclusion

So considering all the pros and cons above, I am in favour for creating a Data Lake. In addition to having a Data Warehouse of course.

 

1 April 2016

Data Sourcing

Filed under: Data Architecture,Data Warehousing — Vincent Rainardi @ 8:16 pm
Tags: ,

One of the critical activity in a data warehouse project (and in any data project) is data sourcing. The business users have a few reports which need to be produced from the data warehouse. There is no data for that yet in the data warehouse, so we look at the report and ask ourselves: “Where can I get the data from to produce this report?” The plan is to find the data source, and then bring the data into the data warehouse.

There are 6 steps of data sourcing:

  1. Find the source
  2. Reproduce the numbers
  3. Verify the extraction
  4. Check the coverage
  5. Check the timing
  6. Check the cost

Step 1. Finding the source

We find out how the report was created, who created it, and from where. The report could be a PDF which is created from PowerPoint, which was created from an Excel spreadsheet, which was created from a file, which was exported from a system. It is this system that we need to check. Find the tables, the columns and the rows where the data is coming from.

Step 2. Reproduce the numbers

Once we have located the tables, we then try to reproduce the numbers we see on the report. For example, the report could be like this:

Asset Allocation

Try to reproduce the first number (24.3%) by querying the tables which you think the data is sourced from. In doing this you will need to look at the data processing happening in the Excel spreadsheet. Do they exclude anything? Is it a straight forward sum of the holdings in government bonds? Do they include supranational? Government agencies? Regional? Municipal? Emerging Market Debt?

If you can match 24.3% then great. If not, investigate if the data is manually overridden before it was summed up. For example, in Excel, there might be a particular bond which was manually classified as Government even though the asset class is Corporate, and this is because it was 80% owned by the government, or because it was backed by the government.

We need to particularly careful with regards to the totals. For example, on the “Allocation by Country of Risk”, if the total portfolio value is $200m, but they exclude FX forwards/swaps, or bond futures, or certain derivatives, then the total portfolio value could decrease to $198m, and all the percentages would be incorrect (they are slightly higher).

Understanding the logic behind the report is critical in reproducing the numbers. In-depth industry knowledge will be helpful to understand the logic. For example:

  1. The credit rating for each debt security is categorised into IG (Investment Grade) if it is BBB- or above, and into HY (High Yield) if it is BB+ or below, or NR.
  2. The credit rating used for point 1 is the average between S&P, Moody’s an Fitch, except a) for own fund use look-through, b) for outside fund use IMMFA
  3. The “Allocation by Country of Risk” table excludes cash and FX forwards/swaps.
  4. Each country is categorised into Developed, Emerging or Frontier market.
  5. When determining the country of risk in point 4, for derivative use the underlying.

In the above case, if we are not familiar with the investment banking industry, it would take a long time for us to understand the logic. So, yes, when doing Data Sourcing, it is best if it is done by a Business Analyst with good knowledge & experience in that industry sector.

Step 3. Verify the extraction

Once we can reproduce the numbers, we need to verify if we can get the data out. A few gotchas are:

  1. The numbers are calculated on-the-fly by the system, and not stored anywhere in the database. If this is the case, find out from the vendor if they have an export utility which produces a file after the numbers have been calculated.
  2. Are we allowed to connect to that database and query it? Do not assume that we can, because I’ve encountered a few cases that we are not allowed to do that. It could be because of the system work load / performance (it is a critical transaction system and they don’t want any big query ruining the front end users), or it could be because they have provided daily extract files which all downstream systems must use (instead of querying the database directly). From the system admin point of view, it makes sense not to allow any external query runs on the database, because we don’t know what kind of damage those external queries can cause, it could block the front end queries and causing a lock.
  3. Loosing precision, i.e. the data must be exported from the database but during the export process the precision decreases from 8 decimal places to 2 decimal places.
  4. There is a security restriction because the it is against the “chinese wall” compliance rules (in investment banking, the public-facing departments must not get data from the M&A department)
  5. The system is being migrated, or rewritten, so it is still in a state of flux and we need to wait a few months.
  6. The system is not a solid “production quality”, but only a “thrown-away”, which means that within a few months they could be dropping those tables.

Step 4. Check the coverage

This step is often missed by many business analysts. We need to check if all the products that we need is available in that system. If the report we are trying to reproduce is reporting 4000 products from 100 branches, but the source system/tables only covers 3000 products from 70 stores, than we’ve got to find out where the other 1000 products and 30 stores are sourced from. Are they produced from a different system.

Not only product and stores. We need to check the coverage in terms of: customer/clients, portfolios, securities, line of business, underwriting classes, asset classes, data providers. And the most important coverage check is on dates, e.g. does the source system have data from 2011 to present? It is possible that the source system only have data from 2014.

Step 5. Check the timing

After checking the coverage, we need to check if the data is available when we need it. We need to check these 3 things: data is too late, available extraction window, the data is overwritten.

  1. Data is too late: If our DW load starts at 2.15 am, will the data be available before that? If not, could the business user live with a bit of stale data (data from 2 days ago, i.e. if today is Wednesday, the latest data in the DW would be Monday data).
  2. Available extraction window: In the production environment, when can we extract the data from that source system? If from 11pm to 6am there is an overnight batch running, and we can’t run our extract during that time, then the ealierst we can run is 7am. If the DW load takes 3 hours, DW users can access it at 10am. Is that too late for the users or not?
  3. The data is overwritten: the data from the source system can be updated many times during the day and when we extract it at night, we have no trace of these changes. Is that ok? Do we need intraday, push-driven data load into the DW? Or would 10 minutes data extraction frequency (pull-driven) be enough?

Step 6. Check the cost

There is no free lunch. We need to check how much it would cost us to use that source data.

  1. If the data is valuable (such as prices, yield and rating from Bloomberg, Reuters and Markit) we would have to pay the data providers. We need to check the cost. The cost could be per $5 security, per call, so it could easily be $20-30k per day. The cost is usually shared with other department.
  2. Check with the data provider, if you use the data only as an input to your calculation, and you don’t publish it / send it on to any external parties (clients, etc.), would it still cost you a lot? Even if you don’t have to pay the data providers, your DW project might still have to share the cost with other departments. Say the data provider is Markit and your company pays $300k/year for prices data and it is currently shared by 5 departments ($60k each). Your project may have to bear the cost of $50k/year ($300/6).
  3. The cost could be a killer to the whole thing, i.e. even if #1 to #5 above are all ok, if the cost of the data is $50k, it could force you to cancel the whole project.
  4. Sometimes other department has to create the data for you. Let’s day yield calculation, or risk engine, or OTC pricing engine, and the requirement from the BI/DW is specific so they have to develop it. It could take them 3 months x 3 people and they could cross charge your project $50k (one off). And that could also be a killer to the DW project.
  5. Developing interface: some systems do not allow external system to pull the data out. They insist to develop an export, and charge the cost to your DW project.
  6. Standard data interface system: some large companies (such as multinational banks) have standard interface (real time, end of day, etc.), and the central middle ware team might charge your DW project some low amount (say $2000 one off + $50/month) to use that standard data interface system. Say you need FX rate data from FX system, and there is already a standard message queue for FX rates with publication time of 11am, 1pm and 3pm. So you “subscribe” to this publication MQ and pay the cost (project cross charge).

16 March 2016

Different Measures for Different Product Types

Filed under: Data Architecture,Data Warehousing,Investment Banking — Vincent Rainardi @ 8:33 am

What I mean by a measure here is a time-variant, numerical property of an entity. It is best to explain this by example. In the investment industry, we have different asset classes: equities, bonds, funds, ETFs, etc. Each asset class has different measures. Equities have opening and closing prices, daily volume, market capitalisation, daily high and low prices, as well as annual and quarterly measures such as turnover, pretax profit and EPS. Bonds have different daily measures: clean and dirty prices, accrued interest, yield and duration. Funds have different daily measures: NAV, alpha, sharpe ratio, and volatility, as well as monthly measures such as 3M return, 1Y return, historic yield, fund size and number of holdings. ETFs have daily bid, mid and offer prices, year high and low, and volume; as well as monthly measures such as performance. The question is: what is an appropriate data model for this situation?

We have three choices:

  1. Put all measures from different product types into a single table.
  2. Separate measures from each product types into different tables.
  3. Put the common measures into one table, and put the uncommon measures into separate tables.

My preference is approach 2, because we don’t need to join across table for each product type. Yes we will need to union across different tables to sum up across product types, but union is much more performant than join operation. The main weakness of approach a is column sparsity.

On top of this of course we will need to separate the daily measures and monthly measures into two different tables. Annual and quarterly measures for equities (such as financial statement numbers) can be combined into one table. We need to remember that measures with different time granularity usually are from different groups. For example, the prices are daily but the performance are monthly.

Static Properties

In addition to different time-variant properties (usually numerical), each asset class also different static properties (can be textual, date or numeric). For example, equities have listed exchanges, industry sectors, country of domicile, and dividend dates. Bonds have issuers, call and maturity dates, and credit ratings. Funds have benchmark, trustee, legal structure and inception date. Examples of numerical properties are minimum initial investment and annual charges for funds; outstanding shares and denomination for equities; par and coupon for bonds. Some static properties are common across asset classes, such as ISIN, country of risk, currency.

Static properties from different asset classes are best stored in separate tables. So we have equity table, bond table, fund table and ETF table. Common properties such as ISIN, country of risk, etc. are best stored in a common table (usually named security table or instrument table).

Why not store all static properties in a common table? Because the properties are different for each asset class so it is like forcing a square peg into a round hole.

Historical Data

For time variant properties it is clear that the table already stores historical data in the rows. Different dates are stored as different rows. What we are discussing here is the historical data of the static attributes. Here we have two choices:

  1. Using SCD approach: store the historical values on different rows (called versions), and each row is only valid for certain time period. SCD stands for slowly changing dimension, a Kimball approach in data warehousing.
  2. Using Audit Table approach: store the historical rows in an audit table (also called history table). This is the traditional approach in normalised modelling. The main advantage is that the main table is light weight and performant.

When to use them? Approach a is suitable for situations where the historical versions are accessed a lot, whereas approach b is suitable for situations where the historical versions are very rarely accessed.

The main issue with approach a is that we need to use “between” on the validity date columns. In data warehousing we have a surrogate key to resolve this issue, but in normalised modelling we don’t. Well, we could and we should. Regardless we are using appraoch a or b, in the time-variant tables we need to store the ID of the historical row for that date. This will make getting historical data a lot faster.

29 December 2015

DimMonth, DimQuarter and DimYear

Filed under: Data Architecture,Data Warehousing — Vincent Rainardi @ 6:21 am
Tags: ,

Sometimes the grains of our fact tables are monthly, quarterly, or yearly. In such cases, how do we create DimMonth, DimQuarter and DimYear? Some of the questions in these cases are:

  1. Why do we need to create Month as a dimension? We can’t Month column in the fact table remain as Month, not as a dimension key?
  2. What does DimMonth look like? What are the attributes? How about DimQuarter?
  3. Should we create DimMonth as a physical table, or as a view on top of DimDate?
  4. What is the surrogate key of DimMonth and DimQuarter? Do we create a new SK, or do we use the SK from DimDate?
  5. Do we need to create a dimension for year? It seems weird because it only has 1 column (so that would be against the “degenerate dimension” concept)

For example, in the source system we have a table which stores the monthly targets for each store:

Or quarterly target like this:

How do we create a fact table for this? Should we create it like this?

Question 1 above: Why do we need to create Month as a dimension? Why can’t we leave it as Month in the fact table, not as a dim key, like this?

Question 2 above: if we decided to keep the column as MonthKey, how should DimMonth look like? What are the attributes? Should DimMonth be like this?

What attributes should we put there?

  • Do we need quarter and year?
  • Do we need half year? e.g. 2015 H1 and 2015 H2
  • For month name, should we use the short name or long name? e.g. Oct 2015 or October 2015?
  • Do we need an attribute for “Month Name without Year”, e.g. October?
  • Do we need “Month End Date” column? e.g. 2015-10-30
  • Do we need “Month is a Quarter End” indicator column for March, June, Sep and Dec?
  • Do we need “Number of days in a month” column? e.g. 30 or 31, or 28 or 29.

Or should we use “intelligent key” like this?

Question 3 above: Should we create DimMonth as a view of DimDate like this?

create view DimMonth as

select min(DateKey) as MonthKey, MonthNumber, MonthName, Quarter, [Year]

from DimDate

group by MonthNumber, MonthName, Quarter, [Year]

 

What are the advantages of creating DimMonth and DimQuarter as a view of DimDate? (compared to creating it as a physical table) What are the disadvantages?

 

I think with the above questions and examples we are now clear about what the issue is. Now let’s answer those questions.

 

Q1. Do we need to create Month as dimension? We can’t Month column in the fact table remain as Month, not as a dimension key, like this?

 

We need the Month column in the fact table to be a Dim Key to a month dimension because we need to access Year and other attributes such as Quarter.

 

Bringing Year into the Sales Target fact table like below is not a good idea, because it makes it inflexible. For example, if we want to add Quarter column we have to alter the fact table structure.

 

Using a Dimension Key to link the fact table to a Month dimension makes it a flexible structure:

 

There is an exception to this: Snapshot Month column. In a monthly periodic snapshot fact table, the first column is Snapshot Month. In this case, we do not need to create this column as a dimension key, linking it to DimMonth. In this case, we do not need a DimMonth. Because we do not need other attributes (like Year or Quarter). A monthly periodic snapshot fact table stores the measures as of the last day of every month, or within that month. For example: number of customers, number of products, number of orders, number of orders for each customer, the highest price and lowest price within that month for every product, the number of new customers for that month, etc.

 

Q2. What does DimMonth look like? What are the attributes?

 

Obviously, the grain of DimMonth is 1 row per month. So we are clear about what the rows are. But what are the columns? Well it depends on what we need.

I usually put MonthNumber, MonthName, Quarter and Year in DimMonth, because they are frequently used.

I don’t find “Month Name without the Year” as a useful attribute. I rarely come across the need for “Half Year” attribute.

“Month is a Quarter End” column is also rarely used. Instead, we usually use “Quarter” column.

“Month End Date” and “Number of days in a month” are also rarely used. Instead, we usually use “IsMonthEnd” indicator column in the DimDate.

For month name, should we use the short name (Oct 2015) or the long name (October 2015)? I found that the short name is more used that the long name. But the month number (2015-10) is even more frequently used that the short name

Q3. Should we create DimMonth as a physical table, or as a view on top of DimDate?

This is really the core of this article. A view on top of DimDate is better in my opinion, because we avoid maintaining two physical tables. And it makes the dimensions less cluttered.

If we make DimMonth and DimQuarter as a physical dimensions, in SSMS Object Explorer, when we open the table section we would see these:

DimDate

DimMonth

DimQuarter

DimYear

But if we create DimMonth and DimQuarter as views, then we will only see DimDate in the Object Explorer’s table section. The DimMonth and DimQuarter will be in the view section.

The main disadvantage of creating DimMonth as a view from DimDate is that it is less flexible. The attribute column that we want to appear in DimMonth should exist in DimDate. But I found that DimMonth usually only need 2 or 3 attributes i.e. Month, Quarter, Year; and all of them are available in the DimDate table. So this is not an issue.

Avoiding maintaining 2 physical tables is quite important because when we extend the date dimension (adding more years i.e. more rows) and we forget to extend DimMonth and DimQuarter, then we will cause an error.

The other consideration is of course the performance. I do not find the performance of DimMonth and DimQuarter to be an issue. This is because DimDate is not too large, and more importantly because the monthly and quarterly fact tables are small, less than 1 million rows. They are much smaller than daily fact tables which have millions or billions of rows.

 

Q4. What is the surrogate key of DimMonth and DimQuarter? Do we create a new SK, or do we use the SK from DimDate?

If we create DimMonth and DimQuarter as physical tables, then the surrogate key can either be pure surrogate (1, 2, 3, …) or intelligent key (201510, 201511, 201512, etc.)

But if we create them as a view of DimDate, then the surrogate key can be either the first day of the month (20151001, 20151101, 20151201, etc.) or the month itself (201510, 201511, 201512, etc.). I prefer the latter than the former because it is more intuitive (intelligent key) and there is no ambiguity like the former.

The script to create the view for DimMonth with SK = 201510, 201511, 201512, etc. is like this:

create view DimMonth as

select distinct convert(int,left(convert(varchar,SK_Date),6)) as MonthKey,

[MonthName] , Quarter, [Year]

from DimDate

 

Q5. Do we need to create a dimension for year?

 

No we don’t need to create DimYear, because it would only have 1 column.

What should we call the dim key column in the fact table then? Is it YearKey or Year? We should call it YearKey, to be consistent with the other dim key columns.

A dimension which only has 1 column, and therefore be kept in the fact table is called a Degenerate Dimension. A Degenerate Dimension is usually used to store identifier of the source table, such as Transaction ID and Order ID. But it is also perfectly valid for dimensions which naturally only have one attribute/column, like Year dimension. See my article about “A dimension with only one attribute” here: link.

23 December 2015

Data Dictionary

Filed under: Data Architecture,Data Warehousing — Vincent Rainardi @ 8:15 am
Tags: ,

During my 10 years or so doing data warehousing projects, I have seen several initiatives doing data dictionary. A data dictionary is a list of definitions used in the system. Each definition is about usually about 10 to 50 words long. And there are about 50 to 200 definitions. What being defined is mostly technical terms, such as the meaning of each field in the database. This is the origin of data dictionary, grew from the need to explain the fields. For example, delivery date, product category, client reference number, tax amount, and so on. I saw this in 1998 all the way to today (2015).

This data dictionary was created by the business analyst in the project, who was trying to explain each field to clear any ambiguity about the terms used in the system. I found that the business people view it differently. They found that it is not too useful for them. The data dictionary is very useful to the new starter in the IT development project, but not too useful for the business. I’ll illustrate with several examples below.

In the database there are terms like: CDS, obligor, PD, yield. The BA defines these terms by copying a definition from internet searches, so they ended up with something like this:

  • Credit Default Swap (CDS): a financial contract on which the seller will pay the buyer if the bond defaults
  • Obligor: a person or entity who is legally obliged to provide some payment to another
  • Probability of Default (PD): likelihood of a default over a particular time frame
  • Yield: the interest received from a bond

The CDS definition is trying to explain a big concept in one sentence. Of course it fails miserably: what bond? What do you mean by “defaults”? After reading the whole sentence, the readers are none of the wiser. The readers will get much better understanding about CDS if they read the Wikipedia page about CDS.

The Obligor definition is too generic. It is too high level so that it is useless because it doesn’t provide the context to the project/system. In a credit portfolio project, obligor should be defined as the borrower of the loan, whereas in a fixed income portfolio project, obligor should be defined as the issuer of the bond.

The PD definition contains a common mistake in data dictionary: the definition contains the word being defined, because the author doesn’t have a good understanding about the topic. What is a default? It is not explained. Which PD is used in this field, Through The Cycle or Point In Time? Stressed or Unstressed? Is it calculated using linear regression or discriminant analysis? That is what the business wants to know.

The Yield definition does not explain the data source. What the business users need is whether it is Yield To Maturity, Current Yield, Simple Yield, Yield to Call or Yield to Worse. There are different definitions of yield depending on the context: fixed income, equity, property or fund. The author has firmly frame it for fixed income, which is good, but within fixed income there are 5 different definitions so the business want to know which one.

Delivery mechanism: Linking and Web Pages

To make the life of the business users easier, sometimes the system (the BI system, reporting system, trade system, finance system, etc.) is equipped with hyperlinks to the data dictionary, directly from the screen. So from the screen we can see these words: CDS, obligor, PD, yield are all hyperlinks and clickable. If we click them, the data dictionary page opens, highlighting the definition. Or it can also be delivered as a “balloon text”.

But mostly, data dictionaries are delivered as static web pages. It is part of the system documentation such as architecture diagram, content of the warehouse, support pages, and troubleshooting guide.

It goes without saying that as web pages, it should have links between definitions.

I would say that the mechanism of delivery only doesn’t contribute much to the value. It is the content which adds a lot of value.

Accessibility

A data dictionary should be accessible by the business, as well as by IT. Therefore it is better if it is delivered as web pages rather than links direct from the system’s screens. This is because web pages can be accessed by everybody in the company, and the access can be easily controlled using AD groups.

Data Source

The most important thing in data dictionary is in-context, concise definition of the field. The second most important thing is the data source, i.e. where is this field coming from. System A’s definition of Duration could be different from system B’s. They may be coming from different sources. In System A it might be sourced from Modified Duration, whereas in System B it is sourced from Effective Duration.

Because of this a data dictionary is per system. We should not attempt building a company-wide data dictionary (like this: link) before we completed the system-specific data dictionaries.

How deep? If the field is sourced from system A, and system A is sourced from system B, which one should we put as the source, system A or system B? We should put both. This is a common issue. The data dictionary says that the source of this field is the data warehouse. Well of course! A lot of data in the company ended up being in the data warehouse! But where does the data warehouse get it from? That’s what the business what to know.

Not Just the Fields

A good data dictionary should also define the values in the field. For example, if we have a field called Region which has 3 possible values: EMEA, APAC, Americas, we should explain what APAC means, what EMEA means and what Americas means (does it include the Caribbean countries?)

This doesn’t mean that if we have a field called currency we then have to define USD, EUR, GBP and 100 other currencies.  If the value of a field is self-explanatory, we leave it. But if it is ambiguous, we explain it.

If the field is a classification field, we should explain why the values are classified that way. For example: the value of Asset Class field could be: equity, fixed income, money market instruments (MMI), CDS, IRS. Many people would argue that CDS and MMI are included in fixed income, so why having separate category for CDS and MMI? Perhaps because MMI has short durations and the business would like to see its numbers separately. Perhaps because the business views CDS as hedging mechanism rather than investment vehicle so they would like to see its numbers separately.

Summary

So in summary, a data dictionary should:

  1. Contain in-context, concise definition of every field
  2. Contain where the field is sourced from
  3. Contain the definition of the values in the field
  4. It should be system specific
  5. It should be delivered as web pages
  6. It should be accessible by both the business and IT

A good data dictionary is part of every data warehouse. I would say that a data warehouse project is not finished until we produce a good data dictionary.

18 December 2015

Measure or Attribute

Filed under: Data Architecture,Data Warehousing — Vincent Rainardi @ 10:00 pm

Last week two people asked me about this subject so I thought it would be beneficial for others if I write it. It is about whether a number should be a measure (a fact) or an attribute (of a dimension). Apologies in advance I can’t reveal the real case as it is a client’s case. But I will present several similar cases to illustrate this issue. As always, it is better to learn by example (cases) than by theory.

First case: worker rate

The first case is the timesheet system in the construction industry. The daily rate of the worker is the number in question. The hourly rate is recorded as an attribute of the worker. The fact table recorded the hours each worker worked in the project, every day. The data mart is used to analyse the cost the project (by work area), and to plan for the next 6-12 months. The cost for each worker for each day is simply the rate multiplied by the hours worked.

There are 2 issues here. One, when the rate entered last month is incorrect, how do we correct it. Second, how to plan for next year using the current rate + 5%, and current work hours -5%.

In this case, it is better to treat the rate as a measure, located in the fact table for actual costs as well as in the fact table for planning. This way we can correct the past rates and recalculate last month’s costs. We can also change the future rates (in the planning fact table) and recalculate the future costs. Yes we will occupy more disk space because we are repeating the rate on every row, but that is a very good price to pay for the advantage we get: more flexibility, ability to support the required functionality and higher performance.

The need to put the rate in the worker dimension arises when we need to know the historical rates for each worker. But this need can also be satisfied if we store the rate as a measure. Granted it would be less performant as we have to plough through lots of fact rows to distil the data to get distinct historical rate per worker. So if this is a mandatory requirement, it is a good idea if we put rate an type 2 attribute, in addition to putting the rate as a measure.

Second case: product price

This is a classic example from the retail industry. But it is also applicable to banking and insurance, pharma, energy and telecom sectors. In a shop, each product has a price. The same product can have different price in different stores. The price can change at any time. The price reviews can happen weekly or monthly, or ad-hoc. It can be done at the store level, city level or regional level. The actual price paid depends on discount or promotion applied for each transaction, at the time the customer purchase the product.

The issue here is, whether we should store the price of the product in a sales fact table, or as an attribute in the product dimension. I maintain an opinion that the price must be store in the sales fact table, for each transaction. In addition, the price should also be stored in the price list fact table. The price on the price list fact table is the official price for each store, for that day. Whereas the price on the sales fact table is the actual price paid, which could be different from price list, because of the offer or promotion applied to that transaction. For example, buy one get one free, or three for the price of two. Or, because the product is in a bundle such as cheese and wine, pyjamas and sleeper, a dictionary and thesaurus, etc. Or, because the product is simply discounted on that day, perhaps due to expiry date, or because the manager want to get rid of the stock quickly.

Having a price in the product dimension, in my opinion is not right. Product attributes such as size and colour, model and optional extras, are true product properties and therefore should be in the product dimension as attributes. These attributes are the same across stores, they are truly the property of a product. But prices can be different across stores, so they are not sole property of a product.

The price may be attached not to an individual product code, but to several product codes. For example, regardless of colour, the prices of the products are the same. Prices can be dynamic, e.g. prices of tickets (entertainment industry or transport industry) of the same show or flight can be different from one seat to the next, from one booking to the next, from one time slot to the next. Prices can be in bands, for example in utility industry (electricity and gas) and in telecommunication industry (mobile calls) the first 100 kWh can have different price to the second 100 kWh, Friday price can be different to Sunday calls.

The concept of price list fact table is more accommodative to satisfy the above demanding requirements, compared to putting price as a type 2 attribute in the product dimension.

11 November 2015

Indexing Fact Tables

A bit on Primary Key

Yesterday a friend asked me why there was no primary key on a fact table. I explained that we did have a fact table surrogate key (FSK) on that fact table, but I made it as a unique non-clustered index because we needed the physical ordering (the clustered index) to be on the snapshot date as it was a daily periodic snapshot fact table, queried mostly by using the snapshot date.

The purpose of having a primary key (PK) is to enforce uniqueness in one column of the table. We can achieve the same thing, by creating a unique key, hence we do not need a PK in that fact table.

We need to pause and think, if in the fact table we need a unique identifier of the rows. If we need to refer a fact row from another fact row (self referencing), then we do need a PK, which is usually a single column bigint FSK. But this unique identifier single column bigint FSK does not have to be an FK, it can be a non-clustered unique index.

The second purpose of having a PK is to enforce not null. This is really not the function of the PK, but more of a requirement of a PK. A PK requires that the column must be not-nullable. So not-nullable is a property of the column itself, not a property of the PK. And we implement this non-nullability when declaring the column on the table DDL.

We need to bear in mind that a PK has nothing to do with clustered or non-clustered indexes. SQL Server will automatically implement a PK as either a unique clustered index (UCI) or a unique non-clustered index (UNCI), depending on whether a clustered index already exists.

The Clustered Index

A clustered index does two things:

  1. Sorting the table physically
  2. Act as the row locator in non-clustered indexes

Point a) is for the performance of the queries. If we don’t partition a periodic snapshot fact table on the snapshot date, the next best thing is to cluster the table on the snapshot date.

But point a) is also for the performance of the update and delete. It is rare, but in some cases we need to update a periodic snapshot fact table (PSFT). So far I only found 1 case where I need to update a PSFT, out of about 150 PFSTs that I have created over the last 10 years. When updating fact table, it is absolutely crucial that the partitioning key, or the clustered fact table if you don’t have it partitioned, to be on the business date, plus the columns used as the joining criteria between the fact staging table and the PSFT. The clustered index should not be on the query criteria columns. It is the job of the non-clustered index to support the query.

Point b) means that the narrower the clustered key, the smaller the non-clustered indexes. Some people think that the narrow clustered key means that the non-clustered index will also have better performance but in my opinion and observation this performance increase is negligible.

So that’s the clustered index. For an insert-only PSFT we put the clustered index on the snapshot date plus the query criteria column to support the query performance. For an insert-and-update PSFT we put the clustered index on the joining criteria of the update statement.

For example, if the joining criteria of the update is snapshot date key + order ID (a degenerate dimension, the PK in the source table), whereas the query criteria is snapshot date key + account key, then for insert-only PSFT the clustered index should be snapshot date key + account key whereas for update PSFT the clustered index should be on snapshot date key + order ID.

The join SQL engine takes when updating the fact table depends on not only the clustered index of the PSFT, but also on the clustered index of the fact staging table (the source of the merge). If we do use the Merge command, we should convert it to update & insert commands. See my article here (link) about the Merge command’s inefficiency.

Partitioning

We can replace the physical ordering functionality above with partitioning. It is common and it make sense to partition a PSFT, I agree. But I would recommend not to partition the fact table when we create it, but later on. This is because of these reasons:

  1. We should not spend the time unless it is required, and when we create the fact table we don’t yet know if the partitioning is required.
  2. When the table is populated and queried we can learn about its condition and behaviour. Much, much better than imagining. For example, is the partitioning required to support query performance, or loading performance?
  3. We may have purging on the fact table, limiting the volume so that we don’t need to partition it because the performance is OK.
  4. We may need to create a new column for the partitioning key.
  5. When we create the fact table, we don’t yet know how the data will be distributed and queried. Purging can change the data distribution. We don’t know the query pattern, for example, 90% of the queries might be on today’s data so we should put it into a separate partition.
  6. Point e above dictates the partition maintenance strategy, i.e. whether we have partitioning functions splitting the table into 10,000 partitions or to 100 partitions with a “sliding window” maintenance. At the creation time, we have very limited knowledge of this.

Non Clustered Indexes

Each surrogate key is ideally indexed. Not combined as 1 index, but as separate indexes. All as non-clustered indexes (NCI). Say we have order date key, customer key, product key, store key, ship date key. Then we create 1 NCI on order date key, 1 NCI on customer key, 1 NCI on product key, 1 NCI on store key and 1 NCI on ship date key.

We should not combine these NCIs into 1 NCI because the second, third, and forth column of the combined NCI will not be used. Because their ordering is not sequential.

See also two of my articles which are related:

  • Indexing Fact Tables in SQL Server (link)
  • Primary Key and Clustered Index in the Fact Table (link)
Next Page »

Create a free website or blog at WordPress.com.