PG 7.2b4 bug?

Don Baccus <dhogaza@pacifier.com> writes:

Apparently there's been a change in the way views are handled within
PostreSQL. The following program works fine in earlier versions.

AFAICT, it was just pure, unadulterated luck that it "works" in prior
versions.

In 7.1 I get:

regression=# select test_seq.nextval from multiple_rows;
NOTICE: Adding missing FROM-clause entry for table "test_seq"
nextval
---------
3
4
(2 rows)

regression=# explain select test_seq.nextval from multiple_rows;
NOTICE: Adding missing FROM-clause entry for table "test_seq"
NOTICE: QUERY PLAN:

Nested Loop (cost=0.00..30.00 rows=1000 width=4)
-> Seq Scan on multiple_rows (cost=0.00..20.00 rows=1000 width=0)
-> Subquery Scan test_seq (cost=0.00..0.00 rows=0 width=0)
-> Result (cost=0.00..0.00 rows=0 width=0)

EXPLAIN

In 7.2 I get:

regression=# select test_seq.nextval from multiple_rows;
NOTICE: Adding missing FROM-clause entry for table "test_seq"
nextval
---------
4
4
(2 rows)

regression=# explain select test_seq.nextval from multiple_rows;
NOTICE: Adding missing FROM-clause entry for table "test_seq"
NOTICE: QUERY PLAN:

Nested Loop (cost=0.00..30.01 rows=1000 width=8)
-> Subquery Scan test_seq (cost=0.00..0.01 rows=1 width=0)
-> Result (cost=0.00..0.01 rows=1 width=0)
-> Seq Scan on multiple_rows (cost=0.00..20.00 rows=1000 width=0)

EXPLAIN

The reason it "works" in 7.1 is that the view is the inside of the
nested loop, and so is re-evaluated for each tuple from the outer query.
(The Result node is where the nextval call is actually being evaluated.)
In 7.2 the view has been placed on the outside of the nested loop, so
it's only evaluated once. The reason for the change is that the 7.2
planner makes the (much more realistic) assumption that evaluating the
Result node isn't free, and so it considers that evaluating the view
multiple times is more expensive than doing it only once. This can be
demonstrated to be the cause by setting the Result cost to zero; then
the behavior matches 7.1:

regression=# show cpu_tuple_cost ;
NOTICE: cpu_tuple_cost is 0.01
SHOW VARIABLE
regression=# set cpu_tuple_cost to 0;
SET VARIABLE
regression=# explain select test_seq.nextval from multiple_rows;
NOTICE: Adding missing FROM-clause entry for table "test_seq"
NOTICE: QUERY PLAN:

Nested Loop (cost=0.00..10.00 rows=1000 width=8)
-> Seq Scan on multiple_rows (cost=0.00..10.00 rows=1000 width=0)
-> Subquery Scan test_seq (cost=0.00..0.00 rows=1 width=0)
-> Result (cost=0.00..0.00 rows=1 width=0)

EXPLAIN
regression=# select test_seq.nextval from multiple_rows;
NOTICE: Adding missing FROM-clause entry for table "test_seq"
nextval
---------
5
6
(2 rows)

However, it's pure luck that you get the nested loop expressed this way
and not the other way when the costs come out the same. I'm surprised
that you consistently got the behavior you wanted in queries more
complex than this test case.

I'd have to say that I consider the code as given to be broken; it's not
a bug for the planner to rearrange this query in any way it sees fit.

It would be nice to accept the Oracle syntax for nextval, but I'm
afraid this hack doesn't get the job done :-(

regards, tom lane

On Mon, 17 Dec 2001, Don Baccus wrote:

insert into data
select test_seq.nextval
from multiple_rows;

I'm not sure that's wrong though with that example. test_seq.nextval in
the select list means to PostgreSQL a join with test_seq which is a view
with one row and I'd expect it to only evaluate that one row once, if
it did it more than once in the past, I'd say it was buggy.

However, I'd think:
"select (select nextval from test_seq) from multiple_rows;"
should give you different values and doesn't, although
"select (select nextval from test_seq where i IS NULL or i IS NOT NULL)
from multiple_rows;" does give you different values.

Stephan Szabo <sszabo@megazone23.bigpanda.com> writes:

However, I'd think:
"select (select nextval from test_seq) from multiple_rows;"
should give you different values and doesn't, although
"select (select nextval from test_seq where i IS NULL or i IS NOT NULL)
from multiple_rows;" does give you different values.

In the first case, the subselect is visibly not dependent on the outer
query, so it's evaluated only once; in the second case it has to be
re-evaluated for each row using that row's value of i. You can see the
difference (InitPlan vs. SubPlan) in the query's EXPLAIN output.

regards, tom lane

Tom Lane wrote:

Don Baccus <dhogaza@pacifier.com> writes:

Apparently there's been a change in the way views are handled within
PostreSQL. The following program works fine in earlier versions.

AFAICT, it was just pure, unadulterated luck that it "works" in prior
versions.

In 7.1 I get:

regression=# select test_seq.nextval from multiple_rows;
NOTICE: Adding missing FROM-clause entry for table "test_seq"
nextval
---------
3
4
(2 rows)

Normally one expects a statement's sematics to depend only upon the
source code and to be consistent, not to vary depending on the mood du
jour of the processor ... this also fails (it's the same statement with
manual substitution):

test=# select (select nextval('test_seq_x') as nextval) as test_seq from
multiple_rows;
test_seq
----------
2
2
(2 rows)

test=#

In other words the function's only called once (as I expected).

I've looked at Date and Darwin's appendix on SQL3's PSMs but it's no
help that I can see, it doesn't get into nitpicking semantic details
regarding their use in queries, just their definition.

Maybe the behavior's implementation defined ... if not, I'd presume SQL3
states that a function in the above context is called either once per
row or once per query, not sometimes one or sometimes the other.

So I think it's too early to write this off as not being a bug ...

--
Don Baccus
Portland, OR
http://donb.photo.net, http://birdnotes.net, http://openacs.org

On Mon, 17 Dec 2001, Tom Lane wrote:

Stephan Szabo <sszabo@megazone23.bigpanda.com> writes:

However, I'd think:
"select (select nextval from test_seq) from multiple_rows;"
should give you different values and doesn't, although
"select (select nextval from test_seq where i IS NULL or i IS NOT NULL)
from multiple_rows;" does give you different values.

In the first case, the subselect is visibly not dependent on the outer
query, so it's evaluated only once; in the second case it has to be
re-evaluated for each row using that row's value of i. You can see the
difference (InitPlan vs. SubPlan) in the query's EXPLAIN output.

I figured that, but I'm not sure whether or not that's a bug of
over-optimizing since I think that it probably *should* give you
different results since it is different each time it's evaluated.
Without checking spec, I'd expect that conceptually the select list
entries are evaluated per row, even if we can avoid that when the
value is certain to be the same, which would mean the result is
incorrect since if it was evaluated per row it would give different
results each time.

Don Baccus <dhogaza@pacifier.com> writes:

Maybe the behavior's implementation defined ... if not, I'd presume SQL3
states that a function in the above context is called either once per
row or once per query, not sometimes one or sometimes the other.

AFAICT, the relevant concept in SQL99 is whether a function is
"deterministic" or not:

An SQL-invoked routine is either deterministic or possibly non-
deterministic. An SQL-invoked function that is deterministic always
returns the same return value for a given list of SQL argument
values. An SQL-invoked procedure that is deterministic always
returns the same values in its output and inout SQL parameters
for a given list of SQL argument values. An SQL-invoked routine
is possibly non-deterministic if, during invocation of that SQL-
invoked routine, an SQL-implementation might, at two different
times when the state of the SQL-data is the same, produce unequal
results due to General Rules that specify implementation-dependent
behavior.

It looks to me like the spec does NOT attempt to nail down the behavior
of non-deterministic functions; in the places where they talk about
non-deterministic functions at all, it's mostly to forbid their use in
contexts where nondeterminism would affect the final result. Otherwise
the results are implementation-defined.

regards, tom lane

Tom Lane wrote:

Don Baccus <dhogaza@pacifier.com> writes:

Maybe the behavior's implementation defined ... if not, I'd presume SQL3
states that a function in the above context is called either once per
row or once per query, not sometimes one or sometimes the other.

It looks to me like the spec does NOT attempt to nail down the behavior
of non-deterministic functions; in the places where they talk about
non-deterministic functions at all, it's mostly to forbid their use in
contexts where nondeterminism would affect the final result. Otherwise
the results are implementation-defined.

Thanks ... I wasn't trying to lobby for a change, I just wanted to make
sure that the standard stated that the behavior is implementation
defined or otherwise punted on the issue before my example was written
off as a non-bug.

At some point the non-deterministic behavior of non-deterministic
functions called in subselects in the target list should probably be
documented, no? Most language standards - at least the ones I've worked
on - require compliant implementations to define and document
implementation-defined behavior ...

Maybe a warning would be appropriate, too?

I realize both of the above would rank pretty low in priority on the
todo list ...

--
Don Baccus
Portland, OR
http://donb.photo.net, http://birdnotes.net, http://openacs.org

Don Baccus <dhogaza@pacifier.com> writes:

Most language standards - at least the ones I've worked
on - require compliant implementations to define and document
implementation-defined behavior ...

SQL99 saith:

g) implementation-defined: Possibly differing between SQL-
implementations, but specified by the implementor for each
particular SQL-implementation.

h) implementation-dependent: Possibly differing between SQL-
implementations, but not specified by ISO/IEC 9075, and not
required to be specified by the implementor for any particular
SQL-implementations.

Behavior of nondeterministic functions falls in the second category ...

regards, tom lane

Tom Lane wrote:

Don Baccus <dhogaza@pacifier.com> writes:

Most language standards - at least the ones I've worked
on - require compliant implementations to define and document
implementation-defined behavior ...

SQL99 saith:

g) implementation-defined: Possibly differing between SQL-
implementations, but specified by the implementor for each
particular SQL-implementation.

h) implementation-dependent: Possibly differing between SQL-
implementations, but not specified by ISO/IEC 9075, and not
required to be specified by the implementor for any particular
SQL-implementations.

Behavior of nondeterministic functions falls in the second category ...

Yep, those are the definitions I'm used to. OK, then, since this is
implementation-dependent, not implementation-defined, PG's off the hook
entirely!

--
Don Baccus
Portland, OR
http://donb.photo.net, http://birdnotes.net, http://openacs.org

Tom Lane wrote:

Don Baccus <dhogaza@pacifier.com> writes:

Maybe the behavior's implementation defined ... if not, I'd presume SQL3
states that a function in the above context is called either once per
row or once per query, not sometimes one or sometimes the other.

This is still bothering me so I decided to plunge into the standard
myself. First of all...

AFAICT, the relevant concept in SQL99 is whether a function is
"deterministic" or not:

Actually this argument may well apply to the function all within the
subselect or view, but I fail to see any language in the standard that
suggests that this trumps the following declaration about the execution
of a <query specification> (what many of us informally refer to as a
"SELECT"):

(from section 7.12, Query Specification)

a) If T is not a grouped table, then

Case:

(I deleted Case i, which refers to standard aggregates like COUNT)

ii) If the <select list> does not include a <set function
specification> that contains a reference to T, then
each <value expression> is applied to each row of T

yielding a table of M rows, where M is the cardinality

of T ...

(FYI a <set function specification> is a standard aggregate like COUNT,
i.e. Case ii pertains to those queries that don't fall into Case i)

ISTM that this quite clearly states that a subselect in a target list
should be applied to each row to be returned in M. I don't see any
waffle-room here. I would have to dig more deeply into the standard's
view regarding VIEW semantics but I would assume it would knit together
in a consistent manner.

For instance, earlier we saw the following exchange between Stephen and Tom:

Stephan Szabo <sszabo@megazone23.bigpanda.com> writes:

However, I'd think:
"select (select nextval from test_seq) from multiple_rows;"
should give you different values and doesn't, although
"select (select nextval from test_seq where i IS NULL or i IS NOT NULL)
from multiple_rows;" does give you different values.

In the first case, the subselect is visibly not dependent on the outer
query, so it's evaluated only once; in the second case it has to be
re-evaluated for each row using that row's value of i.

Note that the standard does not give you this freedom. It says the
<value expression> (in this case the subselect, and yes subselects are
valid <value expressions> in SQL3, at least in my draft) must be applied
to each row.

IMO this means that the optimizer can choose to evaluate the <value
expression> once only if it knows for certain that multiple calls will
return the same value. For example "my_cachable_function()", not
"my_uncachable_function()" or "nextval()".

Or IMO a view built using a non-cachable function.

In other words it can only suppress evaluation if it can be certain that
doing so doesn't change the result.

Another nit-pick, the claim's not even strictly true. "i IS NULL OR i
IS NOT NULL" can be folded to true, so the subselect's not "visibly
dependent on i". In fact, it is quite visibly *not* dependent on the
outer query. PG just isn't smart enough to fold the boolean expression
into the known value "true".

It's this kind of uncertainty that makes the current behavior so ...
ugly. You get different answers depending on various optimization
values, the complexity of the query, etc.

ISTM that the standard is quite clearly worded to avoid this unpleasantness.

...

It looks to me like the spec does NOT attempt to nail down the behavior
of non-deterministic functions; in the places where they talk about
non-deterministic functions at all, it's mostly to forbid their use in
contexts where nondeterminism would affect the final result. Otherwise
the results are implementation-dependent.

I've been looking at a few of the "non-deterministic" clauses in the
General Rules, out of curiousity.

They generally aren't involved with the execution or non-execution of
expressions. Ordering of execution is in many cases non-deterministic
and implementation-dependent. There are plenty of General Rules of this
sort.

We also have this big gaping black hole of non-determinism due to
character set collation.

In other words:

select foo
from bar
order by foo;

is non-deterministic (we don't know the order in which the rows will be
returned) if foo is a character type. This can even be true within
implementations, for instance in PG it changes with when you change
locales (and have locale support enabled).

However, it seems clear that:

select foo, my_function()
from bar
order by foo;

requires my_function() to be called for every row - we just can't depend
on the order in which it will be applied to those rows in the case where
foo is a character type. Of course, iscachable tells the optimizer that
it's OK to just call it once but that's an extension outside SQL3's
domain.

Obviously if you run this query over and over again with the same
collation order the "order by" is deterministic. The non-determinism is
in respect to the portability of the query to implementations built on
differing character sets.

I'm just not seeing justification for claiming that Section 7.12 can be
ignored if the subselect or view happens to contain a function that's
not cachable.

--
Don Baccus
Portland, OR
http://donb.photo.net, http://birdnotes.net, http://openacs.org

Some more bug-or-not-bug thoughts ...

I thought I'd add a quote from Date that furthers my belief that the
subselect example I posted does indeed expose a bug:

(T1 is the table conceptually created by the various joins, etc)

"[if] the select-item takes the form "scalar-expression [[AS] column]"

...

For each row in T1 the scalar-expression is evaluated .."

(page 151 Date & Darwin)

SQL92 didn't support subselects in the select-item-list. SQL3 extends
the expression to include one-row selects that return a single scalar
value. It does NOT however add any wording that allows the subselect to
be yanked and evaluated once rather than evaluated for each row. The
standard uses the word "applied" not "evaluated". I interpret this to
mean "evaluated" and it appears that Date does, too.

On the other hand the view example is giving the proper result in PG
7.2, though only by luck, as Tom pointed out earler. For (given the
view "create view foo as select nextval('foo_sequence') as nextval;")

select foo.nextval
from multiple_rows;

isn't actually legal SQL. It must be stated as:

select foo.nextval
from foo, multiple_rows;

(all PG does is add "foo" to the from clause for me if I leave it out).

The semantics of this are obvious when you think about it - materialize
"foo" then cross-join the resulting table with multiple_rows. Since
"foo" returns a single row computed by "nextval('foo_sequence')"
obviously the result seen with PG 7.2 is correct. Date is quite clear
on the semantics of this and it makes tons of sense since views are
meant to be treated like tables.

So:

1. If an explicit scalar subselect appears in the target list, it should
be executed for every row in the result set.

2. A view referenced in the target list is actually supposed to be
materialized in the FROM clause (even if implictly added to it for
you) then joined to the other tables in the query, if any. Meaning
it should always be executed once and only once. The standard
doesn't have PG-style rules, of course, but such tables are also
should be in the FROM clause, evaluated and joined afterwards
IMO.

At least that's my reading and I've spent quite a bit of time on this now.

Unfortunately PG currently doesn't use the form of the query to decide
whether or not to execute the subselect or view once or for each row,
but rather does so depending on the estimated cost of each approach.

That's the real bug it seems. The form of the query, not the whim of
the optimizer, is the determinant.

Neither of these cases is likely to arise frequently in practice, so if
I ruled Middle Earth I'd decree that:

1. It be filed as a bug

2. It not be assigned a high priority.

However it's not merely of academic interest. The semantics of the view
example is such that you should be able to force single-evaluation of a
function by simply wrapping it in a view, regardless of whether or not
it has side-effects.

Meanwhile I get to go off and inspect the roughly 750 queries that use
this particular style view and determine which ones incorrectly assume
that the view's evaluated more than once per query! :)

--
Don Baccus
Portland, OR
http://donb.photo.net, http://birdnotes.net, http://openacs.org