Question

I came across a question recently that was for “Generating primary key in a clustered environment of 5 App-Servers – [OAS Version 10] without using database”.

Usually we generate PK by a DB sequence, or storing the values in a database table and then using a SP to generate the new PK value…However current requirement is to generate primary key for my application without referencing the database using JDK 1.4.

Need expert’s help to arrive on better ways to handle this.

Thanks,

Answer 1

Use a UUID as your primary key and generate it client-side.

Edit:
Since your comment I felt I should expand on why this is a good way to do things.

Although sequential primary keys are the most common in databases, using a randomly generated primary key is frequently the best choice for distributed databases or (particularly) databases that support a “disconnected” user interface, i.e. a UI where the user is not continuously connected to the database at all times.

UUIDs are the best form of randomly generated key since they are guaranteed to be very unique; the likelyhood of the same UUID being generated twice is so extremely low as to be almost completely impossible. UUIDs are also ubiquitous; nearly every platform has support for the generation of them built in, and for those that don’t there’s almost always a third-party library to take up the slack.

The biggest benefit to using a randomly generated primary key is that you can build many complex data relationships (with primary and foreign keys) on the client side and (when you’re ready to save, for example) simply dump everything to the database in a single bulk insert without having to rely on post-insert steps to obtain the key for later relationship inserts.

On the con side, UUIDs are 16 bytes rather than a standard 4-byte int — 4 times the space. Is that really an issue these days? I’d say not, but I know some who would argue otherwise. The only real performance concern when it comes to UUIDs is indexing, specifically clustered indexing. I’m going to wander into the SQL Server world, since I don’t develop against Oracle all that often and that’s my current comfort zone, and talk about the fact that SQL Server will by default create a clustered index across all fields on the primary key of a table. This works fairly well in the auto-increment int world, and provides for some good performance for key-based lookups. Any DBA worth his salt, however, will cluster differently, but folks who don’t pay attention to that clustering and who also use UUIDs (GUIDs in the Microsoft world) tend to get some nasty slowdowns on insert-heavy databases, because the clustered index has to be recomputed every insert and if it’s clustered against a UUID, which could put the new key in the middle of the clustered sequence, a lot of data could potentially need to be rearranged to maintain the clustered index. This may or may not be an issue in the Oracle world — I just don’t know if Oracle PKs are clustered by default like they are in SQL Server.

If that run-on sentence was too hard to follow, just remember this: if you use a UUID as your primary key, do not cluster on that key!