Re: Oracle TPC Benchmarks and "discrete transactions"

From: bob devine <devine_at_olympus.CS.Berkeley.EDU>
Date: Tue, 1 Sep 1992 21:19:55 GMT
Message-ID: <1992Sep1.211955.10539_at_pasteur.Berkeley.EDU>

levine_charles_at_tandem.com (Charles Levine) writes:

: ---------------------------------------

: - Are discrete transactions a good idea?
: - Is this an innovation that other database vendors should pursue?
: - Given their limited nature and the user sophistication required,
: do discrete transactions make sense in the relational model?

It is obvious that Oracle has invented discrete transactions as a way to generate higher TPC numbers. While these probably don't break the letter of the TPC rules, they do break its spirit!

The bigger question is if a special, slimmed-down transaction semantics are required for real-world cases rather than for the semi-artificial TPC example. Are regular transactions too "heavy" or too general purpose for a class of problems that could more efficiently be solved with a "lite" transaction?

My gut feeling is that: Yes, "lite" transactions might be a valuable addition to the bag of tricks used in database. The optimizer could use them as a hint to avoid some locking or logging that would normally be done for regular transactions.

What needs to be done is a tight analysis of what benefits could result from adding more complexity to the db programmer's life (and to the folks who implement a db's optimizer!). If only a few percentage improvement is possible with a general "lite" transaction then it is not worth the bother. However, the Oracle-7 numbers show that some hefty benefit is possible with their TPC-oriented transaction so a more general "lite" transaction might produce a > 10% gain. In this case it is probably worth the bother when tweaking the last bit of performance out of an large application.

Bob Devine Received on Tue Sep 01 1992 - 23:19:55 CEST