TL;DR: If the database schema should hold all the business logic, how is it be possible to specify that an attribute type is a reference to a specific attribute, instead of a specific record (as is the case of a foreign key)?

To make an example, let’s suppose I have a table "Discounts" with a column "share" which holds the percentage to be applied to the value of column "cost", "price" or "shipping" of the table "Items".

"Discounts" also holds a foreign key to "item_id".

I need to add another column "base" to table "Discounts" where to store a reference to one of the column of table "Items", and calculate the percentage of the value of that column.
For example, given these values:

share    base                 item_id
50       (item's cost)        3
25       (item's price)       1
100      (item's shipping)    2

id    cost    price    shipping
1     10      40       20
2     55      60       30
3     50      85       10

I want to be able to calculate:

  • 50% of 50 (cost of item 3)
  • 25% of 40 (price of item 1)
  • 100% of 30 (shipping of item 2)

The column "base" should contain neither the number (e.g. 3) nor the name (e.g. "price") of the referenced column, because the name or the order of each table could change. In particular a database doesn't have any knowledge about the columns (attributes) order or the rows (records/tuplets) order, infact the RDB theory asserts that «the tuples of a relation have no specific order and that the tuples, in turn, impose no order on the attributes.»

Instead if we rely on the column names, we should enforce that each entry holds a valid attribute name, and whenever the attribute name changes, then we must change its records, constraints and app's validations. If the name is referred in multiple relations, maintaining the database integrity becomes very complex.

The problem here is that we are not writing a reference to the attribute name in the database schema (like when we add a foreign key), but into the data themselves, and this seems a very bad practice, since it threatens the referential integrity.

If there is no DB agnostic way to do this, then assume the database is PostgreSQL (v12+).

  • "The column "base" should contain neither the name (e.g. "price") nor the number (e.g. 3) of the referenced column" - I don't see how you could ever make a connection between the two tables, if you don't want to store any information about the linked item in your discounts table. I would store the "type" of the discount ("price", "cost" or "shipping") in there. How you link that to your apparently completely volatile model is another question though.
    – user1822
    Aug 10, 2020 at 17:20
  • @a_horse_with_no_name: «I don't see how you could ever make a connection between the two tables» | The connection between "Discounts" and "Items" is granted by a foreign key (item_id). A database is not a spreadsheet, so is pretty natural not to entrust the referential integrity to rows' order, columns' order or column's names. My question is if there is a lower level mechanism to reference a column without using its current name. Aug 10, 2020 at 17:41
  • You want to make a connection based on the pk (=classical foreign key relation ship) and the "discount type". The second part of that connection can't be done without storing something in the referencing table.
    – user1822
    Aug 10, 2020 at 17:49
  • 1
    > The column "base" should contain neither the name (e.g. "price") nor the number (e.g. 3) of the referenced column What should the base column then have? What is it that tells which discount this is? Aug 10, 2020 at 18:36
  • @VesaKarjalainen that is the question! Aug 11, 2020 at 9:24

4 Answers 4


I think your model incorrectly represents your use case. The item value element (Cost, Price, or Shipping) seems to be an entity in its own right. My model (without knowing the bigger picture) would probably looks something like this:

Id      | int      | PK
Name    | string   | 
... other attributes

Item ID | int      | PK, FK
Type    | enum     | PK       -- one of: 'cost', 'price', 'shipping'
Value   | currency |

-- (Item ID, Type) would be the primary key here.

Item ID | int      | FK
Type    | enum     | FK
Share   | decimal  |

In reality discounts do not normally apply to individual items, but to some classes of items, identified by different means, and your actual model would change accordingly.

If for you it's not the case, and you do want to have discount per item, you can simply conflate ItemValueElements and Discounts into one entity.

  • This is the best answer so far, and tbh it was the first thing that came up to my mind. I thought it added unnecessary complexity because the DB is indeed already complex, with 'cost' 'price' and 'shipping' belonging to separate models ('product', 'sale' and 'order' tables), along with other attributes that could be discounted (eg 'tax', 'fee', 'service'…). Moreover, those attributes are also variable, having a field for the tax's or service's name and one for its cost. I wanted a flexible way to reference any attribute, but your answer could be the only viable way. Aug 13, 2020 at 9:03

As a #database-design question (as you emphasize), you are going in wrong direction. The proper way to go in relational world is that you design your database so that column names don't change. Luckily, PostgreSQL has many noSQL features you can use.

Also, if number of attributes is not fixed, don't implement them as columns in one table, add one level of abstraction. This might be done by adding one or two more tables or using a JSONB to hold your attributes.

Any way you go, this will add to complexity to your design, making it more difficult to form ad hoq -queries and to implement it without errors. Please think one more time this is what you really need. It surely can be done, if you just can't go without it but it will take some time and more advance planning than you can give in this question.

If you decide to make it, make use of FUNCTIONs to calculate the discounts and make VIEWs so you don't have to type those long queries too often. If attribute names must change, map them to something fixed in a separate table.

  • The proper way to go in relational world is that you design your database so that column names don't change. You might be right. Could you add some authoritarian source for this statement to your answer? Aug 11, 2020 at 12:15
  • Hopefully I get enough up votes here, would that be enough for you? :) Aug 11, 2020 at 12:22
  • Sure! I trust the community. But it depends on the amount and quality of the upvotes ;) Aug 11, 2020 at 12:53

I'm not sure what base is supposed to contain, but I'll assume that it would be ok to store C for cost, P for price and S for shipping?

select id
     , cost - cost_rebate as cost
     , price - price_rebate as price
     , shipping - shipping_rebate as shipping
from (
    select i.id
         , case when d.base = 'C' 
                then d.share / 100.0 
                else 0 
           end * cost as cost_rebate
         , case when d.base = 'P' 
                then d.share / 100.0 
                else 0 
           end * price as price_rebate
         , case when d.base = 'S' 
                then d.share / 100.0 
                else 0 
           end * shipping as shipping_rebate
         , cost
         , price
         , shipping
    from items i
    join discounts d
        on i.id = d.item_id
) as tmp
  • The question is about #database-design. The problem is not how to write a query to calculate the discounts, but how to store a reference to any other column in another table, in order to be able to calculate the discounts. Aug 11, 2020 at 8:58

(Possible solutions, which I don't like by the way)


Following the suggestion by @a_horse_with_no_name to use a column type to describe the referenced column, a first attempt could be to model cost, price and shipping using two columns each, one holding a column type and the other holding the actual value.

We can design the database like this:

share    base_type    item_id
50       C            3
25       P            1
100      S            2

id    type_1    value_1    type_2    value_2    type_3    value_3
1     C         10         P         40         S         20
2     C         55         P         60         S         30
3     C         50         P         85         S         10

And then enforce the values of type_1, type_2, type_3 with a constraint, either in the application logic or in the business logic (ADD CONSTRAINT type_1 CHECK (type_1 IN ('C', 'P', 'S') ecc…)

The problem with this approach is that each type column still needs to know which is the associated value column (and viceversa): if value_1 is renamed "val_1" we would have the same drawbacks as using a column each.


A better approach could be to store cost, price and shipping in a single column holding a hash structure:

hstore(ARRAY['C','10'], ARRAY['P','40'],ARRAY['S','20'])
hstore(ARRAY['C','55'], ARRAY['P','60'],ARRAY['S','30'])
hstore(ARRAY['C','50'], ARRAY['P','85'],ARRAY['S','10'])

Resulting in

id    amounts
1     "C"=>"10","P"=>"40","S"=>"20"
2     "C"=>"55","P"=>"60","S"=>"30"
3     "C"=>"50","P"=>"85","S"=>"10"

In this way we could add data types, name and rename them as wish (either in the application or in another table with TYPE, NAME), and the application could raise an exception to handle the removal of any data type.

The main drawback with this approach is the performance penalty.

  • What advantage do you see with this approach? What would the query look like to produce the expected result in your question? Aug 10, 2020 at 20:01
  • @Lennart The advantage is that whenever column names are changed, you don’t have to search for and replace all occurrences and adjust all constraints and references for that name. If you wrote validation for the column names in the application logic, you have to change those too whenever the columns are renamed. Instead when using a type renaming has no impact, and a simple exception in the application logic is sufficient even if a column is removed. Consider that I’ve given an extremely simplified example, but the DB grows and could have more types and tables dealing with those columns. Aug 10, 2020 at 21:04
  • Unfortunately any code you write (like in Lennart's proposed answer) WILL have to know about the number and names of the columns. Say you add one more type of discount: you will need to change all that code. Aug 11, 2020 at 8:05
  • The only way to make the code totally flexible and make it automatically adapt to the data model is to use dynamic SQL. That introduces another level of complexity to your applications. Unless you hide that behind a layer of stored procedures and dynamic cursors. But even that will mean changing your applications: at the minimum, the host variables in which the results are written. Or the forms used by those applications. Or the web pages/reports updated by those applications ... Aug 11, 2020 at 8:08
  • 1
    Instead of hstore, I'd go to JSONB Aug 11, 2020 at 10:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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