Robert Haas wrote:
> It seems to me that the hard part of this problem is to describe
> the general mechanism by which conflicts will be detected, with
> specific references to the types of data structures that will be
> used to hold that information.
Well, the general approach to tracking SIREAD locks I had in mind is
to keep them in the existing lock data structures. I have in mind to
use multiple granularities, with automatic escalation to coarser
granularities at thresholds, to keep RAM usage reasonable. There are
clearly some tough problems with the pluggable indexes, types,
operators, and such that will take time to sort out an acceptable
implementation at any fine-grained level, so my intent it to punt
those to very coarse granularity in the first pass, with "XXX SIREAD
optimization opportunity" comments where that's not a production-
quality solution or it just seems likely that we can do better with
some work.
I didn't want to get too detailed before I checked that creating a
new lock method for this seemed sane, since the details of the
implementation depend on that choice. Lack of detail tends to draw
accusations of hand-waving, so I was trying to stay away from those
details until my intuition on the lock method was confirmed or shot
down, so I could solidify those details before presenting them.
There is a bit of a chicken and egg problem with moving this forward
-- I guess I was overly conservative on what I presented.
I do understand that this does mean that more RAM will need to be
allocated to the lock structures to support serializable mode. I
don't think that any other option is likely to provide acceptable
performance. I also realize that this means that in the final form,
optimized to where my shop considers it usable, there will still be
coarser granularity than theoretically possible and resulting false
positives causing serialization failures for which the cause is
obscure. We don't care, and anyone who does will probably not want
to use this isolation level. Traditional S2PL doesn't have that
fault, but it blocks so badly that performance is worse; I'll take
the transaction restarts over that any day. I know there are others
who won't.
Basically, the reasons given for having separate lock methods for
DEFAULT ("normal") locks and USER locks seem to apply with almost as
much force to SIREAD locks (no blocking between them, different
source of setting, different lifespans), so I was pretty sure this
was a sane choice, but I just wanted a quick reality check before
developing the level of detail that would move this past hand-waving.
Other than the SIREAD locks to cover predicate locking for
serializable transactions, there is no change to what locks are
acquired. There is no change to blocking or deadlock detection and
recovery. Other transaction isolation levels do not need to change,
except perhaps to fast-path a skip over blocking and deadlock
checking against SIREAD locks (one of those details I'm looking at).
Let me know if you need more information to firm up an opinion on the
sanity of my intuition regarding the new lock method; I'm eager to
move on to the next level of detail.
And thanks for the feedback. :-)
-Kevin