> I't may seem easy to replace a database table with "something else" for
> collecting the changes which have happened during the transaction, but
> you have to answer the following questions:
>
> 1) do I need persistence, what about 2PC ?
>
> 2) does the "something else" work well for all situations an event table
> would work (say, for example, a load of 500GB of data in one
> transaction)
Those are good questions, and a generic system would need to work for
all three of those requirements.
> 3) what would I gain in return for all the work needed to implement the
> "something else" ?
Speed. In my test case, which was replicating view snapshots between
PostgreSQL and Redis, the difference between using an event table and
perverting the constrainttriggers to do an after-insert trigger directly
to redis was a speed difference of around 400%, not counting vacuum
overhead.
>>>> (3) A method of marking DDL changes in the data modification stream.
>
> Yes, DDL triggers or somesuch would be highly desirable.
>
>>> Hmm..can you expand on what you have in mind here? Something more than
>>> just treating the DDL as another item in the (txn ordered) queue?
>> Yeah, that would be one way to handle it. Alternately, you could have
>> the ability to mark rows with a DDL "version".
>
> But the actual DDL would still need to be transferred, no ?
Yes. It may be that having a ddl change simply inserted into the
replication stream is the way to go. Alternatively, DDL versioning
makes a certain amount of sense except that it's pretty hard to make
generic, and would require additional catalog tables.
--Josh Berkus