Greg Smith <greg(at)2ndQuadrant(dot)com> wrote:
> I don't know what other issues Kevin is alluding to, besides this
> reading as if there are no good solutions around--which the above
> text should help with.
For me, that page reads as though it were written with the
assumption that transaction isolation is implemented through
"traditional" locks, with various mentions of MVCC bolted on here
and there in an awkward way, as an afterthought. We can throw in
one more "yeah, but..." clause with a reference out, and that would
help; but the whole thing seems rather badly in need of a rewrite.
I think maybe David Fetter's suggestion is on the mark -- I should
probably put together a SerialziableSI page, and reference that from
the "Snapshot isolation" page and from the "Isolation (database
systems)" page. Oh, and there's also the overlapping
"Serializability" page[1]. I still think a major reorganization of
the "Isolation (database systems)" page is needed, but that's likely
to require a lot more effort and be a more contentious sort of
change.
I'll probably do something this weekend along the lines of what Greg
is suggesting, then (when I find time) do what David is suggesting,
and then (when I feel really ambitious) try to tackle a rewrite of
the Isolation page.
Thanks for the suggestions.
-Kevin
[1] http://en.wikipedia.org/wiki/Serializability