Teradata Surrogate Keys Guidelines

Roland Wenzlofsky

November 19, 2019

minutes reading time


What are Teradata Surrogate Keys?

There are several definitions of what a surrogate key is, but on the whole, they all say the same thing. A Teradata Surrogate Key maps a natural key to an artificial key which is usually of the data type INTEGER or BIGINT. While the natural key can consist of several columns, the surrogate key is one column. Typically, the surrogate key is generated automatically and is represented by an ascending number.

In Teradata, we can use the ROW_NUM() function to programmatically generate the surrogate key or an IDENTITY column for which Teradata automatically generates numbers. What possibilities we have to generate numbers using IDENTITY columns is described in this article:

The IDENTITY column has the advantage over the programmatic method in that the numbers are generated by the system and are consistent. With the programmatic method, it can happen that errors creep in which can be fatal, especially with surrogate keys.

But also with the IDENTITY column, it is important to define it correctly to prevent the creation of duplicates. For this, the two parameters GENERATED ALWAYS and NO CYCLE are necessary. GENERATED ALWAYS prevents that own values can be inserted into the IDENTITY column. NO CYCLE prevents the counter from starting again at the lowest value when the largest allowed number has been generated. Instead, Teradata generates an error when we want to insert more rows.

Detailed information about currently used IDENTITY columns is available in table DBC.IdCol

Disaster Reload of Key Tables with IDENTITY columns

If it is necessary to reload a key table, take care of the following to prevent duplicates from being created:

  • The key table needs to be re-created with new START and MINVALUE definitions to exclude already existing numbers to avoid the same number being generated again.
  • The IDENTITY column definition needs to be changed from GENERATED ALWAYS to GENERATED BY DEFAULT to be able to insert the keys that Teradata generated earlier.
  • New surrogate keys will be generated only when NULL values are passed to the IDENTITY column.

Here is an example where the highest value of the IDENTITY column is 200,000:

CREATE MULTISET TABLE TheKeys 
(
   NATURAL_KEY VARCHAR(500) CHARACTER SET LATIN NOT CASESPECIFIC NOT NULL,
   SURROGATE_KEY BIGINT GENERATED ALWAYS AS IDENTITY
   (START WITH 1 INCREMENT BY 1 
    MINVALUE -999999999999999 
    MAXVALUE 999999999999999 
    NO CYCLE),
)
UNIQUE PRIMARY INDEX (NATURAL_KEY);

If we need to reload the key table, the following DDL is required:

CREATE MULTISET TABLE TheKeys 
(
   NATURAL_KEY VARCHAR(500) CHARACTER SET LATIN NOT CASESPECIFIC NOT NULL,
   SURROGATE_KEY BIGINT GENERATED BY DEFAULT AS IDENTITY
   (START WITH 1 INCREMENT BY 1 
    MINVALUE 200000 
    MAXVALUE 999999999999999 
    NO CYCLE),
)
UNIQUE PRIMARY INDEX (NATURAL_KEY);

Why should we use Surrogate Keys?

There are several reasons why replacing the natural keys with Teradata surrogate keys is suitable.

Different source systems deliver the same natural key for different information.

We may have more than one source system delivering the same value in the natural key for the same target table but with a different meaning. This situation requires using surrogate keys because keeping the natural key from both sources would not allow us to distinguish the objects.

Different source systems deliver a different natural key for the same information.

It can happen that the same information comes from different source systems with different natural keys. For example, the customer information of a bank may be kept in various source systems. Surrogate keys are ideally suited if it is necessary to represent these as a single customer in the data warehouse.

Unfortunately, many Teradata data warehouse implementations decide against surrogate keys. This only shifts the effort of clean integration to the back (data marts, reports). Such an approach will ultimately cause more costs than having the right surrogate key design from the beginning (don’t forget, such an approach requires repeating the logic of combining object information into every data mart or report).

Exchange of Source System leads to the Delivery of new Natural Keys.

If the source system’s natural keys are used, and we undergo a replacement of the original system, we end up in a nasty situation.

If we have surrogate keys available, we must adjust only the mapping table (Natural Keys -> Surrogate Keys). All surrogate keys have to be assigned once to the new natural keys.

If we only have the natural keys available in our data warehouse. The old natural keys must be replaced everywhere by the new natural keys!

We should never tightly couple your data model to the source system structures. An invoice table should be called Invoice and not something like SAP-Invoice. Data Warehousing is not the 1:1 storage of operational data.

At the beginning of a project, it is always more agile to avoid surrogate keys and integrate source system tables 1:1. However, this only shifts costs to the end of the development chain. Changes that have to be made afterward are many times more expensive!

Teradata Surrogates Keys and Performance

If properly used, surrogate keys may slightly increase performance as we can replace several natural key columns with one integer column, but only if there was a character set conversion in joins happening, otherwise, the hashing algorithm’s benefit will be minuscule. Still, there might be an enormous performance advantage when switching to surrogate keys from character column natural keys as described here:

  • Avatar
    Aleksey Svitin says:

    Hi,

    It’s another excellent article. Thank you.

    I would also add
    – Surrogate ids are important elements for Data vault modeling
    – Surrogate ids are must have for tables with complex natural key or for combining different sources(like Original_Id + Source_Id) as Roland says.
    – For many cases, the performance is better with surrogate id because they could be smaller in size (e.g. one INT column vs a couple of varchar columns + source_id), thus joining/grouping is faster.
    – The most important performance benefit is related to”attribute” or dimension tables which is used by “fact” tables. It’s not obvious. Let’s assume we have “small” Country table. Every country has unique string code. We can use this code as a primary key.
    Let’s assume that we have a “big” fact table (translations, mobile calls or anything else) which should have reference to Country table.
    If we use Country_Code as PK when we will use it in dependent fact table as well. But string Country_Code will be always bigger in size than alternative numeric surrogate key. So this difference will be multiplied by number of rows in fact table ( which could be in billions). So it’s very important to have as small PK as possible for dimensions and surrogate keys help a lot here.
    – Personally, I like the approach, when a system should use own “small” internal entity identifiers (generated internally) everywhere inside and GUIDs or natural keys for External communication. It gives both flexibility, strong performance and solid data model.

    It’s just my experience.

    Have a good day!

    Best regards,
    Aleksey Svitin.

    • Hi Aleksey,

      Thanks for your comment.

      Right now I am “eyewitnessing” a total misdesign of a data mart where not using surrogate keys leads to significant problems.

      The data is both delivered from the source system and sent back to it and re-sent a second time. But the format is changed when the data is sent a second time. This is really about little things like spaces etc.

      Since the information from the source system has not been normalized (more than one piece of information is stored in a column), the result is that every join in every SQL must be done with string functions. From a performance point of view a disaster. The optimizer cannot use statistics and fights with full table scans. An elegant solution with surrogate keys at the beginning of the ETL chain could have prevented this.

      – Roland

  • Terdata and few td consultants have fooled organisations with this concept and complicated simple things. Simple surrogate key is to maintain slowly changing dimensions and not for integration of different sources as said in this post. If integration was purpose like example of customer lifecycle then its a separate topic of master data management, dont try to mix all master data , metadata(gcfr) and all things in one and complicate things. I have seen organisations sufferinf becoz of such design and majorly some pakistani consultants have been ruling these confusions and securing their jobs and sucking clients money.
    Honestly teradata is amazing product but design it right way and keep it simple

    • I give you an example, and would kindly ask you to tell me how you would design the model without surrogate keys: You have two billing systems, invoices have to be stored in the same entity (“Invoice”). The natural keys of both billing systems are overlapping. How would you design the “Invoice” table without surrogate keys?

      • Avatar
        Boris Mogilevsky says:

        Conceptually, surrogate keys have a say but using Identity columns for this purpose is a big problem. Their definitions are tied to Hash map which will change in case if the Box needs to be re-configured, for example when there is an upgrade. I personally was in a project where everything has been built using Surrogate Keys based on Identity and there were lots of issues.

        • Avatar
          Алексей Свитин says:

          Hi Boris,

          As a solution, I can suggest to generate Ids during ETL stage instead of relying on Identity column ( which has many side effects) . We use Informatica for that purpose. It works pretty well.

          With best regards,
          Aleksei Svitin

  • can you please let me know about bkey and bmap genearation? and waht they a re

  • {"email":"Email address invalid","url":"Website address invalid","required":"Required field missing"}

    You might also like

    >