I don't know why I'm down-voted for the comment above. Relations have certain semantics that are very hard to preserve without RDBMS controlling the data. For example in a relational system a query like this:
`select p0.name, p1.name from processes as p0 join processes as p1 on p1.id = 1 where p0.id = 1;`
Should never produce two different values. This is impossible to guarantee in all cases if data is updated asynchronously by OS.
Haven't read the code, but this transactional behaviour should be possible to implement. Just ensure that all bound tables in a query reference the same snapshot.
I get what you are alluding to, and you are right that this would require kernel cooperation. But I still think that consistency needs to be defined, before you can make such claim. For instance let's say we're joining a table of processes and open network connections. Now we might have network connections that belong to a processes not listed -is this consistent? What if the implementation filters out the connections that do not belong to a process?
I guess my point is that a reasonable definition of consistency could be many things, including a definition that says referenced tables are read only once in the order they are encountered (this should be equivalent to a series of eg. ps and netstat commands stored in variables and then manipulated)
