Comprehensive testing of a data warehouse at every point throughout the ETL (extract, transform, and load) process is becoming increasingly important as more data is being collected and used for strategic decision-making. Data warehouse or ETL testing is often initiated as a result of mergers and acquisitions, compliance and regulations, data consolidation, and the increased reliance on data-driven decision making (use of Business Intelligence tools, etc.). ETL testing is commonly implemented either manually or with the help of a tool (functional testing tool, ETL tool, proprietary utilities). Let us understand some of the basic ETL concepts.
BI / Data Warehousing testing projects can be conjectured to be divided into ETL (Extract – Transform – Load) testing and henceforth the report testing.
Extract Transform Load is the process to enable businesses to consolidate their data while moving it from place to place (i.e.) moving data from source systems into the data warehouse. The data can arrive from any source:
Extract – It can be defined as extracting the data from numerous heterogeneous systems.
Transform – Applying the business logics as specified b y the business on the data derived from sources.
Load – Pumping the data into the final warehouse after completing the above two process. The ETL part of the testing mainly deals with how, when, from, where and what data we carry in our data warehouse from which the final reports are supposed to be generated. Thus, ETL testing spreads across all and each stage of data flow in the warehouse starting from the source databases to the final target warehouse.
The star schema is perhaps the simplest data warehouse schema. It is called a star schema because the entity-relationship diagram of this schema resembles a star, with points radiating from a central table. The center of the star consists of a large fact table and the points of the star are the dimension tables.
A star schema is characterized by one OR more of very large fact tables that contain the primary information in the data warehouse, and a number of much smaller dimension tables (OR lookup tables), each of which contains information about the entries for a particular attribute in the fact table.
A star query is a join between a fact table and a number of dimension tables. Each dimension table is joined to the fact table using a primary key to foreign key join, but the dimension tables are not joined to each other. The cost-based optimizer recognizes star queries and generates efficient execution plans for them. A typical fact table contains keys and measures. For example, in the sample schema, the fact table sales, contain the measures, quantity sold, amount, average, the keys time key, item-key, branch key, and location key. The dimension tables are time, branch, item and location.
The snowflake schema is a more complex data warehouse model than a star schema, and is a type of star schema. It is called a snowflake schema because the diagram of the schema resembles a snowflake. Snowflake schemas normalize dimensions to eliminate redundancy. That is, the dimension data has been grouped into multiple tables instead of one large table.
For example, a location dimension table in a star schema might be normalized into a location table and city table in a snowflake schema. While this saves space, it increases the number of dimension tables and requires more foreign key joins. The result is more complex queries and reduced query performance. Figure above presents a graphical representation of a snowflake schema.
When to use star schema and snowflake schema?
When we refer to Star and Snowflake Schemas, we are talking about a dimensional model for a Data Warehouse or a Datamart. The Star schema model gets it name from the design appearance because there is one central fact table surrounded by many dimension tables. The relationship between the fact and dimension tables is created by PK -> FK relationship and the keys are generally surrogate to the natural or business key of the dimension tables. All data for any given dimension is stored in the one dimension table. Thus, the design of the model could potentially look like a STAR. On the other hand, the Snowflake schema model breaks the dimension data into multiple tables for the purpose of making the data more easily understood or for reducing the width of the dimension table. An example of this type of schema might be a dimension with Product data of multiple levels. Each level in the Product Hierarchy might have multiple attributes that are meaningful only to that level. Thus, one would break the single dimension table into multiple tables in a hierarchical fashion with the highest level tied to the fact table. Each table in the dimension hierarchy would be tied to the level above by natural or business key where the highest level would be tied to the fact table by a surrogate key. As you can imagine the appearance of this schema design could resemble the appearance of a snowflake.
Types of Dimensions Tables
Type 1: This is straightforward r e f r e s h . The fields are constantly overwritten and history is not kept for the column. For example should a description change for a Product number,the old value will be over written by the new value.
Type 2: This is known as a slowly changing dimension, as history can be kept. The column(s) where the history is captured has to be defined. In our example of the Product description changing for a product number, if the slowly changing attribute captured is the product description, a new row of data will be created showing the new product description. The old description will still be contained in the old.
Type 3: This is also a slowly changing dimension. However, instead of a new row, in the example, the old product description will be moved to an “old value” column in the dimension, while the new description will overwrite the existing column. In addition, a date stamp column exists to say when the value was updated. Although there will be no full history here, the previous value prior to the update is captured. No new rows will be created for history as the attribute is measured for the slowly changing value.
Types of fact tables:
Transactional: Most facts will fall into this category. The transactional fact will capture transactional data such as sales lines or stock movement lines. The measures for these facts can be summed together.
Snapshot: A snapshot fact will capture the current data for point for a day. For example, all the current stock positions, where items are, in which branch, at the end of a working day can be captured.
Snapshot fact measures can be summed for this day, but cannot be summed across more than 2 snapshot days as this data will be incorrect.
Accumulative: An accumulative snapshot will sum data up for an attribute, and is not based on time. For example, to get the accumulative sales quantity for a sale of a particular product, the row of data will be calculated for this row each night – giving an “accumulative” value.
Key hit-points in ETL testing are:There are several levels of testing that can be performed during data warehouse testing and they should be defined as part of the testing strategy in different phases (Component Assembly, Product) of testing. Some examples include:
1. Constraint Testing: During constraint testing, the objective is to validate unique constraints, primary keys, foreign keys, indexes, and relationships. The test script should include these validation points. Some ETL processes can be developed to validate constraints during the loading of the warehouse. If the decision is made to add constraint validation to the ETL process, the ETL code must validate all business rules and relational data requirements. In Automation, it should be ensured that the setup is done correctly and maintained throughout the ever-changing requirements process for effective testing. An alternative to automation is to use manual queries. Queries are written to cover all test scenarios and executed manually.
2. Source to Target Counts: The objective of the count test scripts is to determine if the record counts in the source match the record counts in the target. Some ETL processes are capable of capturing record count information such as records read, records written, records in error, etc. If the ETL process used can capture that level of detail and create a list of the counts, allow it to do so. This will save time during the validation process. It is always a good practice to use queries to double check the source to target counts.
3. Source to Target Data Validation: No ETL process is smart enough to perform source to target field-to-field validation. This piece of the testing cycle is the most labor intensive and requires the most thorough analysis of the data. There are a variety of tests that can be performed during source to target validation. Below is a list of tests that are best practices:
4. Transformation and Business Rules: Tests to verify all possible outcomes of the transformation rules, default values, straight moves and as specified in the Business Specification document. As a special mention, Boundary conditions must be tested on the business rules.
5. Batch Sequence & Dependency Testing: ETL’s in DW are essentially a sequence of processes that execute in a particular sequence. Dependencies do exist among various processes and the same is critical to maintain the integrity of the data. Executing the sequences in a wrong order might result in inaccurate data in the warehouse. The testing process must include at least 2 iterations of the end–end execution of the whole batch sequence. Data must be checked for its integrity during this testing. The most common type of errors caused because of incorrect sequence is the referential integrity failures, incorrect end-dating (if applicable) etc, reject
6. Job restart Testing: In a real production environment, the ETL jobs/processes fail because of number of reasons (say for ex: database related failures, connectivity failures etc). The jobs can fail half/partly executed. A good design always allows for a restart ability of the jobs from the failure point. Although this is more of a design suggestion/approach, it is suggested that every ETL job is built and tested for restart capability.
7. Error Handling: Understanding a script might fail during data validation, may confirm the ETL process is working through process validation. During process validation the testing team will work to identify additional data cleansing needs, as well as identify consistent error patterns that could possibly be diverted by modifying the ETL code. It is the responsibility of the validation team to identify any and all records that seem suspect. Once a record has been both data and process validated and the script has passed, the ETL process is functioning correctly. Conversely, if suspect records have been identified and documented during data validation those are not supported through process validation, the ETL process is not functioning correctly.
8. Views: Views created on the tables should be tested to ensure the attributes mentioned in the views are correct and the data loaded in the target table matches what is being reflected in the views.
9. Sampling: Sampling will involve creating predictions out of a representative portion of the data that is to be loaded into the target table; these predictions will be matched with the actual results obtained from the data loaded for business Analyst Testing. Comparison will be verified to ensure that the predictions match the data loaded into the target table.
10. Process Testing: The testing of intermediate files and processes to ensure the final outcome is valid and that performance meets the system/business need.
11. Duplicate Testing: Duplicate Testing must be performed at each stage of the ETL process and in the final target table. This testing involves checks for duplicates rows and also checks for multiple rows with same primary key, both of which cannot be allowed.
12. Performance: It is the most important aspect after data validation. Performance testing should check if the ETL process is completing within the load window.
13. Volume: Verify that the system can process the maximum expected quantity of data for a given cycle in the time expected.
14.Connectivity Tests: As the name suggests, this involves testing the upstream, downstream interfaces and intra DW connectivity. It is suggested that the testing represents the exact transactions between these interfaces. For ex: If the design approach is to extract the files from source system, we should actually test extracting a file out of the system and not just the
15. Negative Testing: Negative Testing checks whether the application fails and where it should fail with invalid inputs and out of boundary scenarios and to check the behavior of the application.
16. Operational Readiness Testing (ORT): This is the final phase of testing which focuses on verifying the deployment of software and the operational readiness of the application. The main areas of testing in this phase include:
1. Tests the deployment of the solution
2. Tests overall technical deployment “checklist” and timeframes
3. Tests the security aspects of the system including user authentication and
authorization, and user-access levels.
Evolving needs of the business and changes in the source systems will drive continuous change in the data warehouse schema and the data being loaded. Hence, it is necessary that development and testing processes are clearly defined, followed by impact-analysis and strong alignment between development, operations and the business.