1

Background

I'm in a process of designing a database (using a STAR schema).

There are three tables to model: products, tests, states.

The database will be used to store results of tests conducted on products (in a great simplification). There can be many tests pointing out to a single product but each test is unique (they are not shared among products). Besides, I need to record the current state of the product, at the time when the test was conducted. Let's assume that the state of a product describes its current location and owner, which are changing very often. That will most likely involve SCD lvl 2 - to track the history of state changes and to be able to locate a product with all its tests as well as the states it had during these tests.

Problem

I'm not entirely sure how to model this problem. It seems obvious to store every test in a FACT table. This table would then consist of thousands of transactions. On the other hand, there will also be hundreds (and later thousands) of products, so I should probably keep them in a second FACT table. Then, there will also be thousands of state changes, so in order to record their entire history, I would need to keep them in a ... FACT table as well? I've been told that FACT tables are typically used to store multiple-rows data but on the other hand where are the DIMs in this model?

I also don't know how to model the relationships between these tables. Products - states is a 1:* relationship. Products - tests is a 1:* as well. Finally, states - tests is a 1:* too. I would then go with linking products to states and then states to tests (products 1<-* states 1<-* tests), what would allow me to find all states for a particular product and all tests (in all states or in a selected state). What do you think about that? The problem here is that, as I keep adding states, I have two options: either keep duplicating products in the products table (with added "recorded_timestamp" column) or use a SCD lvl 2 in states table, pointing out to the products table with a FK, but this would effectively make the product table a DIM!

Any help here would be very appreciated.

0

3 Answers 3

0

Firstly, neither "star" nor "fact" are abbreviations, as isn't "dimension", so they should not be in all caps. "Fact tables" are called such because they contain facts. A "star schema" resembles (to some people) a star when represented graphically, in case there are many dimensions describing a single fact.

A product cannot be a fact, so it is a dimension. A state may or may not be a fact, so you should look harder at this entity to determine what it really is. A test is certainly a fact.

Depending on what you decide a state to be, your model would be either a two-point "star", with product and state as dimensions, or a one-legged snowflake. You could also find that the combination of state and test is indeed a fact, then you'll end up with a one-point "star".

1
  • I very much like how you cleared this out: "A product cannot be a fact, so it is a dimension. A state may or may not be a fact, so you should look harder at this entity to determine what it really is. A test is certainly a fact.".
    – Maciejo95
    Jul 17, 2020 at 8:23
1

So the first question is: Are you absolutely sure you need to use a star schema? Having worked in a number of warehouses built to varying levels of quality, I have found star schema to be: 1. Inflexible 2. Difficult to use (tons of joins) 3. Really bad for time series.

The way you have described things (Product, Product State, Test) along with the need to track points in time would generate something like this (slight embellishment for party/test):

enter image description here

If you need to get the state as of a certain test this is attainable by a simple join, which can be encapsulated in a view so users don't have to worry about getting the logic correct:

SELECT
  ProductTest.ProductId
 ,ProductTest.TestDtm
 ,ProductTest.TestCd
 ,<Other relevant columns>
 ,ProductState.OwnerId
 ,ProductState.<Location info> --Don't know if this is address or lat/long, use whatever works best for you
FROM
  ProductTest ProductTest
LEFT JOIN
  ProductState ProductState
    ON ProductState.ProductId = ProductTest.ProductId
         AND ProductState.UpdateDtm =
               (
                 SELECT
                   MAX(UpdateDtm)
                 FROM
                   ProductState
                 WHERE
                   ProductId = ProductTest.ProductId
                     AND UpdateDtm <= ProductTest.TestDtm
               )
2
  • For this to work well you will need to be use a database engine that allows/maintains clustered indexes/index organized tables. That limits you (according to my limited memory) to SQL Server, MySQL (recent versions with InnoDb), Oracle, Sybase, and DB2. Which really limits you to SQL Server and MySQL unless you (or the person paying) have very, very deep pockets.
    – bbaird
    Jul 16, 2020 at 19:58
  • Thanks for your extensive answer. I understand the limitations of the star schema. The solution you proposed would indeed work for this problem. However, to keep the question simple, I didn't mention other dimensions and relations involved, which are definitely easier to solve with a star schema - that's why I will stick with it. Nevertheless, thanks for help (due to low reputation, I can't upvote your answer sadly).
    – Maciejo95
    Jul 17, 2020 at 8:20
0

In case anyone ends up having a similar problem to mine - here is how I solved it.

Firstly, following the suggestions found in other answers, I keep my products in a dim table and my tests (results of tests) in a fact. The problem now is how to store the state (history of state changes of a product).

Let's assume a state of a product is defined by attributes A, B and C, which are changing in time. We want to record the history of changes and relate the test results to any particular state in time.

I used mini-dimensions for that. Three mini-dim tables (for A, B and C) are all connected to the main fact table (tests). In this way we always know to which state a particular test relates.

In order to keep track of the history of state changes, I also connect these mini-dims to a factless fact table (which also has attributes such as effective, expired, current ).

This seems to solve both problems. Will now try to build it and update this answer if it doesn't really do the job.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.