Re: computational model of transactions

> A slight correction here, Marshall.
>
> I believe what Erwin is describing is a situation where there are multiple
> transactions running concurrently, and some of them are read-write while
> others are read-only. MVCC can be used to isolate the read only
> transactions from the updaters, while some other mechanism (locking?) can
> be used to protect the updaters from each other.
>
> I may be reading the above into what Erwin wrote, because that's what I'm
> familiar with: a DB server where the reader can be fed consistent data from
> superceded versions of data, while updaters queue up for certain resources.

That's indeed what I meant. The MVCC read philosophy is good for the purpose of guaranteeing that any read-only transaction can always complete without having to wait for some pending update to commit. It becomes problematic when several update transactions come at the stage of constraint checking : then certain reads in other parts of the database must be done for evaluating the constraint expressions, but it is unsafe to use the MVCC read philiosophy for those reads.

I posted the following example already to Marshall in private, thinking it was not all that groundbreaking or even interesting, but here goes again in public :

Transaction T1 updates database object O1 from value O1V1 to value O1V2.
Transaction T2 updates database object O2 from value O2V1 to value O2V2.
There is some database constraint DC involving both O1 and O2. It happens to be the case that all of the value combinations (O1V1,O2V1), (O1V2,O2V1) and (O1V1,O2V2) are valid, but (O1V2,O2V2) is invalid.
When any of these transactions come at the point of constraint checking, they must do some read to evaluate (the expression defining) DC. What value does T1 get when reading O2, if the MVCC read philosophy is applied ? What value does T2 get when reading O1, if the MVCC read philosophy is applied ? What value does the database have if both T1 and T2 pass the constraint checking, believing everything is ok ? Received on Mon Aug 07 2006 - 08:39:18 CEST