Related
I have a problem that can be summarized as follow:
Assume that I am implementing an employee database. For each person depends on his position, different fields should be filled. So for example if the employee is a software engineer, I have the following columns:
Name
Family
Language
Technology
CanDevelopWeb
And if the employee is a business manager I have the following columns:
Name
Family
FieldOfExpertise
MaximumContractValue
BonusRate
And if the employee is a salesperson then some other columns and so on.
How can I implement this in database schema?
One way that I thought is to have some related tables:
CoreTable:
Name
Family
Type
And if type is one then the employee is a software developer and hence the remaining information should be in table SoftwareDeveloper:
Language
Technology
CanDevelopWeb
For business Managers I have another table with columns:
FieldOfExpertise
MaximumContractValue
BonusRate
The problem with this structure is that I am not sure how to make relationship between tables, as one table has relationship with several tables on one column.
How to enforce relational integrity?
There are a few schools of thought here.
(1) store nullable columns in a single table and only populate the relevant ones (check constraints can enforce integrity here). Some people don't like this because they are afraid of NULLs.
(2) your multi-table design where each type gets its own table. Tougher to enforce with DRI but probably trivial with application or trigger logic.
The only problem with either of those, is as soon as you add a new property (like CanReadUpsideDown), you have to make schema changes to accommodate for that - in (1) you need to add a new column and a new constraint, in (2) you need to add a new table if that represents a new "type" of employee.
(3) EAV, where you have a single table that stores property name and value pairs. You have less control over data integrity here, but you can certainly constraint the property names to certain strings. I wrote about this here:
What is so bad about EAV, anyway?
You are describing one ("class per table") of the 3 possible strategies for implementing the category (aka. inheritance, generalization, subclass) hierarchy.
The correct "propagation" of PK from the parent to child tables is naturally enforced by straightforward foreign keys between them, but ensuring both presence and the exclusivity of the child rows is another matter. It can be done (as noted in the link above), but the added complexity is probably not worth it and I'd generally recommend handling it at the application level.
I would add a field called EmployeeId in the EmployeeTable
I'd get rid of Type
For BusinessManager table and SoftwareDeveloper for example, I'll add EmployeeId
From here, you can then proceed to create Foreign Keys from BusinessManager, SoftwareDeveloper table to Employee
To further expand on your one way with the core table is to create a surrogate key based off an identity column. This will create a unique employee id for each employee (this will help you distinguish between employees with the same name as well).
The foreign keys preserve your referential integrity. You wouldn't necessarily need EmployeeTypeId as someone else mentioned as you could filter on existence in the SoftwareDeveloper or BusinessManagers tables. The column would instead act as a cached data point for easier querying.
You have to fill in the types in the below sample code and rename the foreign keys.
create table EmployeeType(
EmployeeTypeId
, EmployeeTypeName
, constraint PK_EmployeeType primary key (EmployeeTypeId)
)
create table Employees(
EmployeeId int identity(1,1)
, Name
, Family
, EmployeeTypeId
, constraint PK_Employees primary key (EmployeeId)
, constraint FK_blahblah foreign key (EmployeeTypeId) references EmployeeType(EmployeeTypeId)
)
create table SoftwareDeveloper(
EmployeeId
, Language
, Technology
, CanDevelopWeb
, constraint FK_blahblah foreign key (EmployeeId) references Employees(EmployeeId)
)
create table BusinessManagers(
EmployeeId
, FieldOfExpertise
, MaximumContractValue
, BonusRate
, constraint FK_blahblah foreign key (EmployeeId) references Employees(EmployeeId)
)
No existing SQL engine has solutions that make life easy on you in this situation.
Your problem is discussed at fairly large in "Practical Issues in Database Management", in the chapter on "entity subtyping". Commendable reading, not only for this particular chapter.
The proper solution, from a logical design perspective, would be similar to yours, but for the "type" column in the core table. You don't need that, since you can derive the 'type' from which non-core table the employee appears in.
What you need to look at is the business rules, aka data constraints, that will ensure the overall integrity (aka consistency) of the data (of course whether any of these actually apply is something your business users, not me, should tell you) :
Each named employee must have exactly one job, and thus some job detail somewhere. iow : (1) no named employees without any job detail whatsoever and (2) no named employees with >1 job detail.
(3) All job details must be for a named employee.
Of these, (3) is the only one you can implement declaratively if you are using an SQL engine. It's just a regular FK from the non-core tables to the core table.
(1) and (2) could be defined declaratively in standard SQL, using either CREATE ASSERTION or a CHECK CONSTRAINT involving references to other tables than the one the CHECK CONSTRAINT is defined on, but neither of those constructs are supported by any SQL engine I know.
One more thing about why [including] the 'type' column is a rather poor choice to make : it changes how constraint (3) must be formulated. For example, you can no longer say "all business managers must be named employees", but instead you'd have to say "all business managers are named employees whose type is <type here>". Iow, the "regular FK" to your core table has now become a reference to a VIEW on your core table, something you might want to declare as, say,
CREATE TABLE BUSMANS ... REFERENCES (SELECT ... FROM CORE WHERE TYPE='BM');
or
CREATE VIEW BM AS (SELECT ... FROM CORE WHERE TYPE='BM');
CREATE TABLE BUSMANS ... REFERENCES BM;
Once again something SQL doesn't allow you to do.
You can use all fields in the same table, but you'll need an extra table named Employee_Type (for example) and here you have to put Developer, Business Manager, ... of course with an unique ID. So your relation will be employee_type_id in Employee table.
Using PHP or ASP you can control what field you want to show depending the employee_type_id (or text) in a drop-down menu.
You are on the right track. You can set up PK/FK relationships from the general person table to each of the specialized tables. You should add a personID to all the tables to use for the relationship as you do not want to set up a relationship on name because it cannot be a PK as it is not unique. Also names change, they are a very poor choice for an FK relationship as a name change could cause many records to need to change. It is important to use separate tables rather than one because some of those things are in a one to many relationship. A Developer for instnce may have many differnt technologies and that sort of thing should NEVER be stored in a comma delimted list.
You could also set up trigger to enforce that records can only be added to a specialty table if the main record has a particular personType. However, be wary of doing this as you wil have peopl who change roles over time. Do you want to lose the history of wha the person knew when he was a developer when he gets promoted to a manager. Then if he decides to step back down to development (A frequent occurance) you would have to recreate his old record.
I have three basic types of entities: People, Businesses, and Assets. Each Asset can be owned by one and only one Person or Business. Each Person and Business can own from 0 to many Assets. What would be the best practice for storing this type of conditional relationship in Microsoft SQL Server?
My initial plan is to have two nullable foreign keys in the Assets table, one for People and one for Businesses. One of these values will be null, while the other will point to the owner. The problem I see with this setup is that it requires application logic in order to be interpreted and enforced. Is this really the best possible solution or are there other options?
Introducing SuperTypes and SubTypes
I suggest that you use supertypes and subtypes. First, create PartyType and Party tables:
CREATE TABLE dbo.PartyType (
PartyTypeID int NOT NULL identity(1,1) CONSTRAINT PK_PartyType PRIMARY KEY CLUSTERED
Name varchar(32) CONSTRAINT UQ_PartyType_Name UNIQUE
);
INSERT dbo.PartyType VALUES ('Person'), ('Business');
SuperType
CREATE TABLE dbo.Party (
PartyID int identity(1,1) NOT NULL CONSTRAINT PK_Party PRIMARY KEY CLUSTERED,
FullName varchar(64) NOT NULL,
BeginDate smalldatetime, -- DOB for people or creation date for business
PartyTypeID int NOT NULL
CONSTRAINT FK_Party_PartyTypeID FOREIGN KEY REFERENCES dbo.PartyType (PartyTypeID)
);
SubTypes
Then, if there are columns that are unique to a Person, create a Person table with just those:
CREATE TABLE dbo.Person (
PersonPartyID int NOT NULL
CONSTRAINT PK_Person PRIMARY KEY CLUSTERED
CONSTRAINT FK_Person_PersonPartyID FOREIGN KEY REFERENCES dbo.Party (PartyID)
ON DELETE CASCADE,
-- add columns unique to people
);
And if there are columns that are unique to Businesses, create a Business table with just those:
CREATE TABLE dbo.Business (
BusinessPartyID int NOT NULL
CONSTRAINT PK_Business PRIMARY KEY CLUSTERED
CONSTRAINT FK_Business_BusinessPartyID FOREIGN KEY REFERENCES dbo.Party (PartyID)
ON DELETE CASCADE,
-- add columns unique to businesses
);
Usage and Notes
Finally, your Asset table will look something like this:
CREATE TABLE dbo.Asset (
AssetID int NOT NULL identity(1,1) CONSTRAINT PK_Asset PRIMARY KEY CLUSTERED,
PartyID int NOT NULL
CONSTRAINT FK_Asset_PartyID FOREIGN KEY REFERENCES dbo.Party (PartyID),
AssetTag varchar(64) CONSTRAINT UQ_Asset_AssetTag UNIQUE
);
The relationship the supertype Party table shares with the subtype tables Business and Person is "one to zero-or-one". Now, while the subtypes generally have no corresponding row in the other table, there is the possibility in this design of having a Party that ends up in both tables. However, you may actually like this: sometimes a person and a business are nearly interchangeable. If not useful, while a trigger to enforce this will be fairly easily done, the best solution is probably to add the PartyTypeID column to the subtype tables, making it part of the PK & FK, and put a CHECK constraint on the PartyTypeID.
The beauty of this model is that when you want to create a column that has a constraint to a business or a person, then you make the constraint to the appropriate table instead of the party table.
Also, if desired, turning on cascade delete on the constraints can be useful, as well as an INSTEAD OF DELETE trigger on the subtype tables that instead delete the corresponding IDs from the supertype table (this guarantees no supertype rows that have no subtype rows present). These queries are very simple and work at the entire-row-exists-or-doesn't-exist level, which in my opinion is a gigantic improvement over any design that requires checking column value consistency.
Also, please notice that in many cases columns that you would think should go in one of the subtype tables really can be combined in the supertype table, such as social security number. Call it TIN (taxpayer identification number) and it works for both businesses and people.
ID Column Naming
The question of whether or not to call the column in the Person table PartyID, PersonID, or PersonPartyID is your own preference, but I think it's best to call these PersonPartyID or BusinessPartyID—tolerating the cost of the longer name, this avoids two types of confusion. E.g., someone unfamiliar with the database sees BusinessID and doesn't know this is a PartyID, or sees PartyID and doesn't know it is restricted by foreign key to just those in the Business table.
If you want to create views for the Party and Business tables, they can even be materialized views since it's a simple inner join, and there you could rename the PersonPartyID column to PersonID if you were truly so inclined (though I wouldn't). If it's of great value to you, you can even make INSTEAD OF INSERT and INSTEAD OF UPDATE triggers on these views to handle the inserts to the two tables for you, making the views appear completely like their own tables to many application programs.
Making Your Proposed Design Work As-Is
Also, I hate to mention it, but if you want to have a constraint in your proposed design that enforces only one column being filled in, here is code for that:
ALTER TABLE dbo.Assets
ADD CONSTRAINT CK_Asset_PersonOrBusiness CHECK (
CASE WHEN PersonID IS NULL THEN 0 ELSE 1 END
+ CASE WHEN BusinessID IS NULL THEN 0 ELSE 1 END = 1
);
However, I don't recommend this solution.
Final Thoughts
A natural third subtype to add is organization, in the sense of something that people and businesses can have membership in. Supertype and subtype also elegantly solve customer/employee, customer/vendor, and other problems similar to the one you presented.
Be careful not to confuse "Is-A" with "Acts-As-A". You can tell a party is a customer by looking in your order table or viewing the order count, and may not need a Customer table at all. Also don't confuse identity with life cycle: a rental car may eventually be sold, but this is a progression in life cycle and should be handled with column data, not table presence--the car doesn't start out as a RentalCar and get turned into a ForSaleCar later, it's a Car the whole time. Or perhaps a RentalItem, maybe the business will rent other things too. You get the idea.
It may not even be necessary to have a PartyType table. The party type can be determined by the presence of a row in the corresponding subtype table. This would also avoid the potential problem of the PartyTypeID not matching the subtype table presence. One possible implementation is to keep the PartyType table, but remove PartyTypeID from the Party table, then in a view on the Party table return the correct PartyTypeID based on which subtype table has the corresponding row. This won't work if you choose to allow parties to be both subtypes. Then you would just stick with the subtype views and know that the same value of BusinessID and PersonID refer to the same party.
Further Reading On This Pattern
Please see A Universal Person and Organization Data Model for a more complete and theoretical treatment.
I recently found the following articles to be useful for describing some alternate approaches for modeling inheritance in a database. Though specific to Microsoft's Entity Framework ORM tool, there's no reason you couldn't implement these yourself in any DB development:
Table Per Hierarchy
Table Per Type (this is what I advocate above as the only fully normalized method of implementing inheritance in a database)
Table Per Concrete Class
Or a more brief overview of these three ways: How to choose an inheritance strategy
P.S. I have switched, more than once, my opinion on column naming of IDs in subtype tables, due to having more experience under my belt.
You don't need application logic to enforce this. The easiest way is with a check constraint:
(PeopleID is null and BusinessID is not null) or (PeopleID is not null and BusinessID is null)
You can have another entity from which Person and Business "extend". We call this entity Party in our current project. Both Person and Business have a FK to Party (is-a relationship). And Asset may have also a FK to Party (belongs to relationship).
With that said, if in the future an Asset can be shared by multiple instances, is better to create m:n relationships, it gives flexibility but complicates the application logic and the queries a bit more.
ErikE's answer gives a good explanation on how to go about the supertype / subtype relationship in tables and is likely what I'd go for in your situation, however, it doesn't really address the question(s) you've posed which are also interesting, namely:
What would be the best practice for storing this type of conditional relationship in Microsoft SQL Server?
...are there other options?
For those I recommend this blog entry on TechTarget which has an excerpt from excerpt from "A Developer's Guide to Data Modeling for SQL Server, Covering SQL Server 2005 and 2008" by Eric Johnson and Joshua Jones which addresses 3 possible options.
In summary they are:
Supertype Table - Almost matches what you've proposed, have a table with some fields that will always be null when others are filled. Good when only a couple of fields aren't shared. So depending on how different Business and People are you could possibly combine them into one table, Owners perhaps, and then just have OwnerID in your Asset table.
Subtype Tables - Basically the opposite of what Supertype tables are and is what you have just now. Here we have lots of unique fields and maybe one or two the same so we just have the repeated fields appear in each table. As you are finding this isn't really suitable for your situation.
Supertype and Subtype Tables - A combination of both of the above where the matching fields are placed in a single table and the unique ones in separate tables and matching IDs are used to join the record from one table to the other. This matches ErikE's proposed solution and, as mentioned, is the one I would favour as well.
Sadly it doesn't go on to explain which, if any, are best practice but it is certainly a good read to get an idea of the options that are out there.
YOu can enforce the logic with a trigger instead. Then no matter how the record is changed, only one of the fileds will be filled in.
You could also have a PeopleAsset table and a BusinessAsset table, but stillwould have the problem of enforcing that only one of them has a record.
An asset would have a foreign key to the owning person, and you should setup an association table to link assets and businesses. As said in other comments, you can use triggers and/or constraints to ensure that the data stays in a consistent state. ie. when you delete a business, delete the lines in your association table.
Table People, Businesses both can use UUID as primary key, and union both to a view for sql join purpose.
so you can simply use one foreign key column in Assets relation to both People and Businesses, because UUID is nearly unique. And you can simply query like:
select * from Assets
join view_People_Businesses as v on v.id = Assets.fk
For example, I have a table that stores classes, and a table that stores class_attributes. class_attributes has a class_attribute_id and a class_id, while classes has a class_id.
I'd guess if a dataset is "a solely child of" or "belongs solely to" or "is solely owned by", then I need a FK to identify the parent. Without class_id in the class_attributes table I could never find out to which class this attribute belongs to.
Maybe there's an helpful answer matrix for this?
Wikipedia is helpful.
In the context of relational
databases, a foreign key is a
referential constraint between two
tables.1 The foreign key identifies
a column or a set of columns in one
(referencing) table that refers to a
column or set of columns in another
(referenced) table. The columns in the
referencing table must be the primary
key or other candidate key in the
referenced table.
(and it goes on into more and more detail)
If you want to enforce the constraint that each row in class_attributes applies to exactly one row of classes, you need a foreign key. If you don't care about enforcing this (ie, you're fine to have attributes for non-existent classes), you don't need an FK.
I don't have an answer matrix, but just for clarification purposes, we're talking about Database Normalization:
http://en.wikipedia.org/wiki/Database_normalization
And to a certain extent Denormalization:
http://en.wikipedia.org/wiki/Denormalization
I would say, it's the other way around. First, you design what kind of objects you need to have. For those will create a table.
Part of this phase is designing the keys, that is the combinations of attributes (columns) that uniquely identify the object. You may or may not add an artificial key or surrogate key for convenience or performance reasons. From these keys, you typically elect one canonical key, the primary key, which you try to use consistently to identify objects in that table (you keep the other keys too, they serve to ensure unicity as a business rule, not so much for identificattion purposes.)
Then, you think what relationships exist between the objects. An object that is 'owned' by another object, or an object that refers to another object needs some way to identify its related object. In the corresponding table (child table) you add columns to make a foreign key to point to the primary key of the referenced table.
This takes care of all one to many relationships.
Sometimes, an object can be related multiple times to another object. For example, an order can be used to order multiple products, but a product can appear on multiple orders as well. For those relationships, you design a separate table (intersection table - in this example, order_items). This table will have a unique key created from two foreign keys: one pointing to the one parent (orders), one to the other parent (products). And again, you add the columns to the intersection table that you need to create those foreign keys.
So in short, you first design keys and foreign keys, only then you start adding columns to implement them.
Don't be concerned with the type of relationship -- it has more to do with the cardinality of the relationship.
If you have a one-to-many relationship, then you'd want to assign a Primary Key to the smaller of the tables, and store it as a Foreign Key in the larger table.
You'd also do it with one-to-one relationships, but some people argue that you should avoid them.
In the case of a many-to-many relationship, you'd want to make a join table, and then have each of the original tables have a foreign key to the join table.
Should I always have a primary key in my database tables?
Let's take the SO tagging. You can see the tag in any revision, its likely to be in a tag_rev table with the postID and revision number. Would I need a PK for that?
Also since it is in a rev table and not currently use the tags should be a blob of tagIDs instead of multiple entries of multiple post_id tagid pair?
A table should have a primary key so that you could identify each row uniquely with it.
Technically, you can have tables without a primary key, but you'll be breaking good database design rules.
You should strive to have a primary key in any non-trivial table where you're likely to want to access (or update or delete) individual records by that key. Primary keys can consist of multiple columns, and formally speaking, will be the shortest available superkey; that is, the shortest available group of columns which, together, uniquely identify any row.
I don't know what the Stack Overflow database schema looks like (and from some of the things I've read on Jeff's blog, I don't want to), but in the situation you describe, it's entirely possible there is a primary key across the post identifier, revision number and tag value; certainly, that would be the shortest (and only) superkey available.
With regards to your second point, while it may be reasonable to argue in favour of aggregating values in archive tables, it does go against the principle that each row/column intersection in a table ought to contain one single value. While it may slightly simplify development, there is no reason you can't keep to a normalised table with versioned metadata, even for something as trivial as tags.
I tend to agree that most tables should have a primary key. I can only think of two times where it doesn't make sense to do it.
If you have a table that relates keys to other keys. For example, to relate a user_id to an answer_id, that table wouldn't need a primary key.
A logging table, whose only real purpose is to create an audit trail.
Basically, if you are writing a table that may ever need to be referenced in a foreign key relationship then a primary key is important, and if you can't be positive it won't be, then just add the PK. :)
See this related question about whether an integer primary key is required. One of the answers uses tagging as an example:
Are there any good reasons to have a database table without an integer primary key
For more discussion of tagging and keys, see this question:
Id for tags in tag systems
From MySQL 5.5 Reference Manual section 13.1.17:
If you do not have a PRIMARY KEY and an application asks for the PRIMARY KEY in your tables, MySQL returns the first UNIQUE index that has no NULL columns as the PRIMARY KEY.
So, technically, the answer is no. However, as others have stated, in most cases it is quite useful.
I firmly believe every table should have a way to uniquely identify a record. For 99% of the tables, this is a primary key. For the rest you may get away with a unique index (I'm thinking one column look up type tables here). Any time I have a had to work with a table without a way to uniquely identify records, there has been trouble.
I also believe if you are using surrogate keys as your PK, you should, where at all possible, have a separate unique index on whatever combination of fields make up the natural key. I realize there are all too many times when you don't have a true natural key (names are not unique or what makes something unique might be spread across several parentchild tables), but if you do have one, please please please make sure it has a unique index or is created as the PK.
If there is no PK, how will you update or delete a single row ? It would be impossible ! To be honest I have used a few times tables without PK, for instance to store activity logs, but even in this case it is advisable to have one because the timestamps could not be granular enough. Temporary tables is another example. But according to relational theory the PK is mandatory.
it is good to have keys and relationships . Helps a lot. however if your app is good enough to handle the relationships then you could possibly skip the keys ( although i recommend that you have them )
Since I use Subsonic, I always create a primary key for all of my tables. Many DB Abstraction libraries require a primary key to work.
Note: that doesn't answer the "Grand Unified Theory" tone of your question, but I'm just saying that in practice, sometimes you MUST make a primary key for every table.
If it's a join table then I wouldn't say that you need a primary key. Suppose, for example, that you have tables PERSONS, SICKPEOPLE, and ILLNESSES. The ILLNESSES table has things like flu, cold, etc., each with a primary key. PERSONS has the usual stuff about people, each also with a primary key. The SICKPEOPLE table only has people in it who are sick, and it has two columns, PERSONID and ILLNESSID, foreign keys back to their respective tables, and no primary key. The PERSONS and ILLNESSES tables contain entities and entities get primary keys. The entries in the SICKPEOPLE table aren't entities and don't get primary keys.
Databases don't have keys, per se, but their constituent tables might. I assume you mean that, but just in case...
Anyway, tables with a large number of rows should absolutely have primary keys; tables with only a few rows don't need them, necessarily, though they don't hurt. It depends upon the usage and the size of the table. Purists will put primary keys in every table. This is not wrong; and neither is omitting PKs in small tables.
Edited to add a link to my blog entry on this question, in which I discuss a case in which database administration staff did not consider it necessary to include a primary key in a particular table. I think this illustrates my point adequately.
Cyberherbalist's Blog Post on Primary Keys
I have the following tables in my database that have a many-to-many relationship, which is expressed by a connecting table that has foreign keys to the primary keys of each of the main tables:
Widget: WidgetID (PK), Title, Price
User: UserID (PK), FirstName, LastName
Assume that each User-Widget combination is unique. I can see two options for how to structure the connecting table that defines the data relationship:
UserWidgets1: UserWidgetID (PK), WidgetID (FK), UserID (FK)
UserWidgets2: WidgetID (PK, FK), UserID (PK, FK)
Option 1 has a single column for the Primary Key. However, this seems unnecessary since the only data being stored in the table is the relationship between the two primary tables, and this relationship itself can form a unique key. Thus leading to option 2, which has a two-column primary key, but loses the one-column unique identifier that option 1 has. I could also optionally add a two-column unique index (WidgetID, UserID) to the first table.
Is there any real difference between the two performance-wise, or any reason to prefer one approach over the other for structuring the UserWidgets many-to-many table?
You only have one primary key in either case. The second one is what's called a compound key. There's no good reason for introducing a new column. In practise, you will have to keep a unique index on all candidate keys. Adding a new column buys you nothing but maintenance overhead.
Go with option 2.
Personally, I would have the synthetic/surrogate key column in many-to-many tables for the following reasons:
If you've used numeric synthetic keys in your entity tables then having the same on the relationship tables maintains consistency in design and naming convention.
It may be the case in the future that the many-to-many table itself becomes a parent entity to a subordinate entity that needs a unique reference to an individual row.
It's not really going to use that much additional disk space.
The synthetic key is not a replacement to the natural/compound key nor becomes the PRIMARY KEY for that table just because it's the first column in the table, so I partially agree with the Josh Berkus article. However, I don't agree that natural keys are always good candidates for PRIMARY KEY's and certainly should not be used if they are to be used as foreign keys in other tables.
Option 2 uses a simple compund key, option 1 uses a surrogate key. Option 2 is preferred in most scenarios and is close to the relational model in that it is a good candidate key.
There are situations where you may want to use a surrogate key (Option 1)
You are not certain that the compound key is a good candidate key over time. Particularly with temporal data (data that changes over time). What if you wanted to add another row to the UserWidget table with the same UserId and WidgetId? Think of Employment(EmployeeId,EmployeeId) - it would work in most cases except if someone went back to work for the same employer at a later date
If you are creating messages/business transactions or something similar that requires an easier key to use for integration. Replication maybe?
If you want to create your own auditing mechanisms (or similar) and don't want keys to get too long.
As a rule of thumb, when modeling data you will find that most associative entities (many to many) are the result of an event. Person takes up employment, item is added to basket etc. Most events have a temporal dependency on the event, where the date or time is relevant - in which case a surrogate key may be the best alternative.
So, take option 2, but make sure that you have the complete model.
I agree with the previous answers but I have one remark to add.
If you want to add more information to the relation and allow more relations between the same two entities you need option one.
For example if you want to track all the times user 1 has used widget 664 in the userwidget table the userid and widgetid isn't unique anymore.
What is the benefit of a primary key in this scenario? Consider the option of no primary key:
UserWidgets3: WidgetID (FK), UserID (FK)
If you want uniqueness then use either the compound key (UserWidgets2) or a uniqueness constraint.
The usual performance advantage of having a primary key is that you often query the table by the primary key, which is fast. In the case of many-to-many tables you don't usually query by the primary key so there is no performance benefit. Many-to-many tables are queried by their foreign keys, so you should consider adding indexes on WidgetID and UserID.
Option 2 is the correct answer, unless you have a really good reason to add a surrogate numeric key (which you have done in option 1).
Surrogate numeric key columns are not 'primary keys'. Primary keys are technically one of the combination of columns that uniquely identify a record within a table.
Anyone building a database should read this article http://it.toolbox.com/blogs/database-soup/primary-keyvil-part-i-7327 by Josh Berkus to understand the difference between surrogate numeric key columns and primary keys.
In my experience the only real reason to add a surrogate numeric key to your table is if your primary key is a compound key and needs to be used as a foreign key reference in another table. Only then should you even think to add an extra column to the table.
Whenever I see a database structure where every table has an 'id' column the chances are it has been designed by someone who doesn't appreciate the relational model and it will invariably display one or more of the problems identified in Josh's article.
I would go with both.
Hear me out:
The compound key is obviously the nice, correct way to go in so far as reflecting the meaning of your data goes. No question.
However: I have had all sorts of trouble making hibernate work properly unless you use a single generated primary key - a surrogate key.
So I would use a logical and physical data model. The logical one has the compound key. The physical model - which implements the logical model - has the surrogate key and foreign keys.
Since each User-Widget combination is unique, you should represent that in your table by making the combination unique. In other words, go with option 2. Otherwise you may have two entries with the same widget and user IDs but different user-widget IDs.
The userwidgetid in the first table is not needed, as like you said the uniqueness comes from the combination of the widgetid and the userid.
I would use the second table, keep the foriegn keys and add a unique index on widgetid and userid.
So:
userwidgets( widgetid(fk), userid(fk),
unique_index(widgetid, userid)
)
There is some preformance gain in not having the extra primary key, as the database would not need to calculate the index for the key. In the above model though this index (through the unique_index) is still calculated, but I believe that this is easier to understand.