How to version control a record in a database

Let's say you have a FOO table that admins and users can update. Most of the time you can write queries against the FOO table. Happy days.

Then, I would create a FOO_HISTORY table. This has all the columns of the FOO table. The primary key is the same as FOO plus a RevisionNumber column. There is a foreign key from FOO_HISTORY to FOO. You might also add columns related to the revision such as the UserId and RevisionDate. Populate the RevisionNumbers in an ever-increasing fashion across all the *_HISTORY tables (i.e. from an Oracle sequence or equivalent). Do not rely on there only being one change in a second (i.e. do not put RevisionDate into the primary key).

Now, every time you update FOO, just before you do the update you insert the old values into FOO_HISTORY. You do this at some fundamental level in your design so that programmers can't accidentally miss this step.

If you want to delete a row from FOO you have some choices. Either cascade and delete all the history, or perform a logical delete by flagging FOO as deleted.

This solution is good when you are largely interested in the current values and only occasionally in the history. If you always need the history then you can put effective start and end dates and keep all the records in FOO itself. Every query then needs to check those dates.

I think you are looking for versioning the content of database records (as StackOverflow does when someone edits a question/answer). A good starting point might be looking at some database model that uses revision tracking.

The best example that comes to mind is MediaWiki, the Wikipedia engine. Compare the database diagram here, particularly the revision table.

Depending on what technologies you're using, you'll have to find some good diff/merge algorithms.

Check this question if it's for .NET.

In the BI world, you could accomplish this by adding a startDate and endDate to the table you want to version. When you insert the first record into the table, the startDate is populated, but the endDate is null. When you insert the second record, you also update the endDate of the first record with the startDate of the second record.

When you want to view the current record, you select the one where endDate is null.

This is sometimes called a type 2 Slowly Changing Dimension. See also TupleVersioning

Upgrade to SQL 2008.

Try using SQL Change Tracking, in SQL 2008. Instead of timestamping and tombstone column hacks, you can use this new feature for tracking changes on data in your database.

MSDN SQL 2008 Change Tracking

Two options:

1. Have a history table - insert the old data into this history table whenever the original is updated.
2. Audit table - store the before and after values - just for the modified columns in an audit table along with other information like who updated and when.

You can perform auditing on a SQL table via SQL triggers. From a trigger you can access 2 special tables (inserted and deleted). These tables contain the exact rows that were inserted or deleted each time the table is updated. In the trigger SQL you can take these modified rows and insert them into the audit table. This approach means that your auditing is transparent to the programmer; requiring no effort from them or any implementational knowledge.

The added bonus of this approach is that the auditing will occur regardless of whether the sql operation took place via your data access DLLs, or via a manual SQL query; (as the auditing is performed on the server itself).

Alok suggested Audit table above, I would like explain it in my post.

I adopted this schema-less, single table design on my project.

Schema:

• id - INTEGER AUTO INCREMENT
• username - STRING
• tablename - STRING
• oldvalue - TEXT / JSON
• newvalue - TEXT / JSON
• createdon - DATETIME

This table can hold historical records for each table all in once place, with complete object history in one record. This table can populated using triggers / hooks where data changes, storing old and new value snapshot of the target row.

Pros with this design:

• Less number of tables to manage for history management.
• Stores full snapshot of each row old and new state.
• Easy to search on each table.
• Can create partition by table.
• Can define data retention policy per table.

Cons with this design:

• Data size can be large, if system has frequent changes.

You don't say what database, and I don't see it in the post tags. If it's for Oracle, I can recommend the approach that is built in in Designer: use journal tables. If it's for any other database, well, I basically recommend the same way, too...

The way it works, in case you want to replicate it in another DB, or maybe if you just want to understand it, is that for a table there is a shadow table created too, just a normal database table, with the same field specs, plus some extra fields: like what action was last taken (string, typical values "INS" for insert, "UPD" for update and "DEL" for delete), datetime for when the action took place, and user id for who did it.

Through triggers, every action to any row in the table inserts a new row in the journal table with the new values, what action was taken, when, and by what user. You don't ever delete any rows (at least not for the last few months). Yes it'll grow big, easily millions of rows, but you can easily track the value for any record at any point in time since the journaling started or the old journal rows got last purged, and who made the last change.

In Oracle everything you need is generated automatically as SQL code, all you have to do is to compile/run it; and it comes with a basic CRUD application (actually only "R") to inspect it.

As an additional step to the answers above me, I would suggest giving each generated change a unique ID, likely something with the date/time and a unique counter for each day (so that multiple updates a second don't overlap). I would include a action type code within this code, so "9129128213939REPLACE". This provides a robustness to allow sanity checking that your other system of history is working correctly.