Roman Rokytskyy wrote:

>> That fact that transactions on different nodes will see
>> different database states is no different than a pair of concurrent
>> transactions on a single system.
>>
>
> Without giving much though I tend to agree with this for the
> share-nothing architecture...
>
> But what about situations where the data sets which are located on
> different nodes intersect or are "identical"? In other words, what if
> the same transaction executed on both nodes see different versions of
> the same record?
>
>

If a transaction executes on a single node, there is no issue. But even
relaxing this, it boils down to the same principle that governs Firebird
transactions, which is that any given transaction sees a stable set of
committed transactions. Records that were committed by a transaction
that was visibly committed when another transaction starts are visible,
otherwise a back version (or nothing) is required.

--
Jim Starkey
President, NimbusDB, Inc.
978 526-1376