Outer Join - Shifting Result Set.

I'm a quantitative securities analyst working with Compustat data
(company fiscal reports and pricing feeds).

My coworker came across a problem that we fixed, but I'd like to
understand 'why' it was happening and just don't get it yet.

Here's the starting query (reduced to simple prefixes):


----INITIAL-----

declare @mthDate datetime
set @mthDate = (select max(datadate) from t)
declare @wkDate datetime
set @wkDate = (select max(datadate) from z)

Select
...
from
z
left join a on a.idA = z.idA and a.idB = z.idB
and a.datadate = z.datadate
left join b on b.idA = z.idA and b.idB = z.idB
and b.datadate = @mthDate
left join c on c.idA = z.idA and c.idB = z.idB
and c.datadate = @mthDate
left join d on d.idA = z.idA and d.idB = z.idB
and d.datadate = z.datadate
left join e on e.idA = z.idA
and e.datadate = @mthDate
left join f on f.idA = e.idA and f.datadate=e.date2
left join g on g.idA = e.idA and g.datadate=e.date2
left join h on h.idA = z.idA
left join k on k.ticker = z.ticker
left join m on m.idA = z.idA and m.idB=z.idB
where
z.datadate = @wkDate
<..some other expression filters...>
and k.ticker is null

----END INITIAL-----------

As you can see 'z' is the main table that things are linked to via
outer joins (our security master). Table 'k' has a list of securities
that we wish not to have results for.

There are 77 entries in table k and 4933 in table z for that given
time. We'd expect 4856 to be in this, but no. it's 4400, and then the
next time you run it (no changes whatsover) it's 2312, and so on.
Every time you execute you get a different record count.

My thought/and fix was to move the (k.ticker) predicate out of the
where clause and get a differenced set from z using NOT EXISTS:


-----AMENDED---------------
from
(z where not exists(select * from k where k.ticker=y.ticker)) y
left join a on a.idA = y.idA and a.idB = y.idB
and a.datadate = y.datadate
left join b on b.idA = y.idA and b.idB = y.idB
and b.datadate = @mthDate
left join c on c.idA = y.idA and c.idB = y.idB
and c.datadate = @mthDate
left join d on d.idA = y.idA and d.idB = y.idB
and d.datadate = y.datadate
left join e on e.idA = y.idA
and e.datadate = @mthDate
left join f on f.idA = e.idA and f.datadate=e.date2
left join g on g.idA = e.idA and g.datadate=e.date2
left join h on h.idA = y.idA
left join k on k.ticker = y.ticker
left join m on m.idA = y.idA and m.idB=y.idB
where
y.datadate = @wkDate
<..some other expression filters...>

------------------------

And this works. It's stable now.

I'm hoping someone here can help me up the wisdom curve by explaining
to me 'why' the recordset kept changing before.

My guess is that the cost-based optimizer was resorting the outer joins
and handing back different sets as a result, but i want to understand,
and thought i'd come to this group for help.

I appreciate your time and look forward to replies.

Greg McIntire