shibpid serializable transactions

[Cantor, Scott]

On 3/6/16, 3:32 PM, "users on behalf of Charles R. Tompkins" <users-bounces at shibboleth.net on behalf of crt at ufl.edu> wrote:


>
>What is the reasoning behind making the transaction isolation level
>serializable for these transactions?

It is the basic contract the code requires, or there would be no point to using a relational database because the application would have to do all the locking and error handling anyway. Lower isolation levels violate the basic properties of a transaction. 

Oracle, and most databases, refuse to actually implement serializable transactions anymore. At best they require proprietary SQL or a number of configuration changes to approximate it.

The only allowance I'm willing to make for that insanity is the retry logic, and I'm not happy about that.

>It's not easily agreeing with my multimode environment.

It is a *requirement* of a multinode system because the locking can't be done in the application without using a distributed lock manager. If the code weren't assuming multiple nodes, then it wouldn't be a big burden to implement the locking internally and lower the isolation level to begin with.

-- Scott