search_path for PL/pgSQL functions partially cached?

On Friday, December 27, 2024, Jan Behrens <jbe-mlist@magnetkern.de> wrote:

It seems that it matters *both* how the search_path was set during the
*first* invocation of the function within a session *and* how it is set
during the actual call of the function. So even if there are just two
schemas involved, there are 4 possible outcomes for the "run" function's
result ('2.4', '2', '5', and '5.4'). To me, this behavior seems to be
somewhat dangerous. Maybe it is even considered a bug?

It is what it is - and if one is not careful one can end up writing
hard-to-understand and possibly buggy code due to the various execution
environments and caches involved.

I’ve never really understood why “%TYPE’ exists…

Or is it documented somewhere?

https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING

Can someone explain to me what's going on, and what is the best practice to

deal with it? Is there a way to avoid fully qualifying every type and
expression? Which parts do I have to qualify or is this something that
could be fixed in a future version of PostgreSQL?

Add qualification or attach a “set search_path” clause to “create
function”. Code stored in the server should not rely on the session
search_path.

David J.

Hi

pá 27. 12. 2024 v 21:26 odesílatel David G. Johnston <
david.g.johnston@gmail.com> napsal:

On Friday, December 27, 2024, Jan Behrens <jbe-mlist@magnetkern.de> wrote:

It seems that it matters *both* how the search_path was set during the
*first* invocation of the function within a session *and* how it is set
during the actual call of the function. So even if there are just two
schemas involved, there are 4 possible outcomes for the "run" function's
result ('2.4', '2', '5', and '5.4'). To me, this behavior seems to be
somewhat dangerous. Maybe it is even considered a bug?

It is what it is - and if one is not careful one can end up writing
hard-to-understand and possibly buggy code due to the various execution
environments and caches involved.

I think plan cache should be invalidated when search_path is different, but
maybe there is some bug - there are some optimizations related to faster
execution of simple expressions.

I’ve never really understood why “%TYPE’ exists…

referenced types should increase readability - it ensures type
compatibility - minimally on oracle, where the change of schema requires
recompilation. In Postgres it is working on 99% - plpgsql functions don't
hold dependency on types.

Or is it documented somewhere?

https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING

Can someone explain to me what's going on, and what is the best practice

to deal with it? Is there a way to avoid fully qualifying every type and
expression? Which parts do I have to qualify or is this something that
could be fixed in a future version of PostgreSQL?

Add qualification or attach a “set search_path” clause to “create
function”. Code stored in the server should not rely on the session
search_path.

a lot of functionality in Postgres depends on the search path - and then
all should be consistent. Sure, writing procedures that depend on the
current search path can be a short way to hell.

I cannot to reproduce it

CREATE OR REPLACE FUNCTION s1.fx1()
RETURNS integer
LANGUAGE plpgsql
AS $function$
begin
return 100;
end
$function$

CREATE OR REPLACE FUNCTION s2.fx1()
RETURNS integer
LANGUAGE plpgsql
AS $function$
begin
return 200;
end
$function$

CREATE OR REPLACE FUNCTION public.foo()
RETURNS void
LANGUAGE plpgsql
AS $function$
declare v int;
begin v := fx1();
raise notice '%', v;
end;
$function$

(2024-12-27 21:53:13) postgres=# set search_path to s1;
SET
(2024-12-27 21:53:34) postgres=# select public.foo();
NOTICE: 100
┌─────┐
│ foo │
╞═════╡
│ │
└─────┘
(1 row)

(2024-12-27 21:53:44) postgres=# set search_path to s2;
SET
(2024-12-27 21:53:47) postgres=# select public.foo();
NOTICE: 200
┌─────┐
│ foo │
╞═════╡
│ │
└─────┘
(1 row)

(2024-12-27 21:53:48) postgres=# set search_path to s1;
SET
(2024-12-27 21:53:51) postgres=# select public.foo();
NOTICE: 100
┌─────┐
│ foo │
╞═════╡
│ │
└─────┘
(1 row)

so from my perspective is pg ok, tested on pg16 and pg18

On 12/27/24 12:26, David G. Johnston wrote:

On Friday, December 27, 2024, Jan Behrens <jbe-mlist@magnetkern.de
<mailto:jbe-mlist@magnetkern.de>> wrote:

It seems that it matters *both* how the search_path was set during
the *first* invocation of the function within a session *and* how it
is set during the actual call of the function. So even if there are
just two schemas involved, there are 4 possible outcomes for the
"run" function's result ('2.4', '2', '5', and '5.4'). To me, this
behavior seems to be somewhat dangerous. Maybe it is even considered
a bug?

It is what it is - and if one is not careful one can end up writing
hard-to-understand and possibly buggy code due to the various execution
environments and caches involved.

I’ve never really understood why “%TYPE’ exists…

Per:

https://www.postgresql.org/docs/current/plpgsql-declarations.html#PLPGSQL-DECLARATION-TYPE

"By using %TYPE you don't need to know the data type of the structure
you are referencing, and most importantly, if the data type of the
referenced item changes in the future (for instance: you change the type
of user_id from integer to real), you might not need to change your
function definition.

%TYPE is particularly valuable in polymorphic functions, since the data
types needed for internal variables can change from one call to the
next. Appropriate variables can be created by applying %TYPE to the
function's arguments or result placeholders."

The second case I can buy, the first I am not so sure of. It seems to me
the first case it can be 'solved' by the second case.

Or is it documented somewhere?

https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING <https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING&gt;

Can someone explain to me what's going on, and what is the best
practice to deal with it? Is there a way to avoid fully qualifying
every type and expression? Which parts do I have to qualify or is
this something that could be fixed in a future version of PostgreSQL?

Add qualification or attach a “set search_path” clause to “create
function”.  Code stored in the server should not rely on the session
search_path.

David J.

--
Adrian Klaver
adrian.klaver@aklaver.com

"David G. Johnston" <david.g.johnston@gmail.com> writes:

It is what it is - and if one is not careful one can end up writing
hard-to-understand and possibly buggy code due to the various execution
environments and caches involved.

Yeah, I don't see this changing. The actual answer is that we have
search_path-aware caching of expressions and query plans within a
plpgsql function, which is why the call to foo() reacts to the current
search path. But the types of plpgsql variables are only looked up
on the first use (within a session). Perhaps we ought to work harder
on that, but it seems like a lot of overhead to add for something that
will benefit next to nobody.

I’ve never really understood why “%TYPE’ exists…

Compatibility with Oracle, I imagine. I agree it's a bizarre feature.
But you could get the same behavior without %TYPE, just by referencing
some other type that has different declarations in different schemas.

Add qualification or attach a “set search_path” clause to “create
function”. Code stored in the server should not rely on the session
search_path.

Yeah, adding "set search_path" is recommendable if you don't want to
think hard about this stuff.

regards, tom lane

pá 27. 12. 2024 v 22:03 odesílatel Tom Lane <tgl@sss.pgh.pa.us> napsal:

"David G. Johnston" <david.g.johnston@gmail.com> writes:

It is what it is - and if one is not careful one can end up writing
hard-to-understand and possibly buggy code due to the various execution
environments and caches involved.

Yeah, I don't see this changing. The actual answer is that we have
search_path-aware caching of expressions and query plans within a
plpgsql function, which is why the call to foo() reacts to the current
search path. But the types of plpgsql variables are only looked up
on the first use (within a session). Perhaps we ought to work harder
on that, but it seems like a lot of overhead to add for something that
will benefit next to nobody.

I’ve never really understood why “%TYPE’ exists…

Compatibility with Oracle, I imagine. I agree it's a bizarre feature.
But you could get the same behavior without %TYPE, just by referencing
some other type that has different declarations in different schemas.

This feature is not bizarre - just the implementation in Postgres is not
fully complete (and I am not sure if it is fixable). PLpgSQL uses plan
cache, but there is nothing similar for types.
It is designed for Oracle where search_path doesn't exist, and where change
of schema invalidates code, and requires recompilation. PL/pgSQL and
Postgres are much more dynamic systems than Oracle. Maybe PL/pgSQL
functions can holds dependency on types, and when any related custom type
is changed, then the cached function can be invalidated. Unfortunately, the
frequent change of search path can kill the performance.

On Fri, 27 Dec 2024 13:26:28 -0700
"David G. Johnston" <david.g.johnston@gmail.com> wrote:

Or is it documented somewhere?

https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING

I can't find any notes regarding functions and schemas in that section.

Can someone explain to me what's going on, and what is the best practice to

deal with it? Is there a way to avoid fully qualifying every type and
expression? Which parts do I have to qualify or is this something that
could be fixed in a future version of PostgreSQL?

Add qualification or attach a “set search_path” clause to “create
function”. Code stored in the server should not rely on the session
search_path.

David J.

In my (real world) case, I was unable to use "SET search_path FROM
CURRENT" because it isn't possible to use "SET" in procedures that use
transactions, due to this documented limitation:

"If a SET clause is attached to a procedure, then that procedure cannot
execute transaction control statements (for example, COMMIT and
ROLLBACK, depending on the language)."

https://www.postgresql.org/docs/17/sql-createprocedure.html

My procedure looks more or less like this:

CREATE PROCEDURE "myfunc"()
LANGUAGE plpgsql AS
$$
DECLARE
"old_search_path" TEXT;
-- some more variables
BEGIN
SELECT current_setting('search_path') INTO "old_search_path";
SET search_path TO 'myschema';
-- some code that uses COMMIT and SET TRANSACTION ISOLATION LEVEL
PERFORM set_config('search_path', "old_search_path", FALSE);
END;
$$;

My question is: Am I safe if I use fully-qualified types in the DECLARE
section only? Or do I need to provide full qualification also in the
code below (after SET search_path TO 'myschema')?

And bonus question: Is it documented somewhere?

Maybe not many people run into these issues because schemas and
functions aren't used as often in combination?

Kind Regards
Jan Behrens

Hi

Maybe not many people run into these issues because schemas and
functions aren't used as often in combination?

I think schema and functions are common combinations. But when people have
objects with the same name, then they are careful to be sure, so objects
have really identical structure.
Using different types in these objects is very rare. And because Postgres
doesn't support it well, experienced developers don't use it. Similar
issues can do some issues after an stored procedures update, because can
require session reset. Or when you need it, you can use a much more dynamic
type like record.

On Sat, 28 Dec 2024 00:40:09 +0100
Jan Behrens <jbe-mlist@magnetkern.de> wrote:

On Fri, 27 Dec 2024 13:26:28 -0700
"David G. Johnston" <david.g.johnston@gmail.com> wrote:

Or is it documented somewhere?

https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING

I can't find any notes regarding functions and schemas in that section.

Actually, I found another note in the documentation. But it doesn't
explain things correctly. In the documentation for PostgreSQL 17,
section 36.17.6.1. (Security Considerations for Extension Functions)
says:

"SQL-language and PL-language functions provided by extensions are at
risk of search-path-based attacks when they are executed, since parsing
of these functions occurs at execution time not creation time."

https://www.postgresql.org/docs/17/extend-extensions.html#EXTEND-EXTENSIONS-SECURITY

So here, the manual explicity states that functions are parsed at
execution, not creation time. As seen in my original example in this
thread, this isn't (fully) true. Moreover, it isn't true for all
SQL-language functions, as can be demonstrated with the following code:

============

CREATE SCHEMA s1;
CREATE SCHEMA s2;

CREATE VIEW s1.v AS SELECT 'creation' AS col;
CREATE VIEW s2.v AS SELECT 'runtime' AS col;

SET search_path TO 'public', 's1';

CREATE FUNCTION use_sql_atomic() RETURNS TEXT LANGUAGE sql BEGIN ATOMIC
SELECT 'use_sql_atomic = ' || col FROM v;
END;

CREATE FUNCTION use_sql_string() RETURNS TEXT LANGUAGE sql AS $$
SELECT 'use_sql_string = ' || col FROM v;
$$;

CREATE FUNCTION use_plpgsql() RETURNS TEXT LANGUAGE plpgsql AS $$ BEGIN
RETURN (SELECT 'use_plpgsql = ' || col FROM v);
END; $$;

SET search_path TO 'public', 's2';

SELECT use_sql_atomic() AS "output" UNION ALL
SELECT use_sql_string() AS "output" UNION ALL
SELECT use_plpgsql() AS "output";

============

This generates the following output:

output
---------------------------
use_sql_atomic = creation
use_sql_string = runtime
use_plpgsql = runtime
(3 rows)

Overall, PostgreSQL doesn't behave consistent, and to me it seems that
the documentation isn't describing its behavior correctly either.

I understand if fixing this is too much work (even though I would
really like to see this fixed). But given that the current behavior is
highly surprising and inconsistent - and keeping in mind that this is a
subject that may affect security - I think the documentation should
reflect the current behavior at least. I thus see this as a
documentation issue.

Kind regards,
Jan Behrens

On 1/1/25 09:55, Jan Behrens wrote:

On Sat, 28 Dec 2024 00:40:09 +0100
Jan Behrens <jbe-mlist@magnetkern.de> wrote:

On Fri, 27 Dec 2024 13:26:28 -0700
"David G. Johnston" <david.g.johnston@gmail.com> wrote:

Or is it documented somewhere?

https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING

I can't find any notes regarding functions and schemas in that section.

Actually, I found another note in the documentation. But it doesn't
explain things correctly. In the documentation for PostgreSQL 17,
section 36.17.6.1. (Security Considerations for Extension Functions)
says:

"SQL-language and PL-language functions provided by extensions are at
risk of search-path-based attacks when they are executed, since parsing
of these functions occurs at execution time not creation time."

https://www.postgresql.org/docs/17/extend-extensions.html#EXTEND-EXTENSIONS-SECURITY

So here, the manual explicity states that functions are parsed at
execution, not creation time. As seen in my original example in this
thread, this isn't (fully) true. Moreover, it isn't true for all
SQL-language functions, as can be demonstrated with the following code:

============

CREATE SCHEMA s1;
CREATE SCHEMA s2;

CREATE VIEW s1.v AS SELECT 'creation' AS col;
CREATE VIEW s2.v AS SELECT 'runtime' AS col;

SET search_path TO 'public', 's1';

CREATE FUNCTION use_sql_atomic() RETURNS TEXT LANGUAGE sql BEGIN ATOMIC
SELECT 'use_sql_atomic = ' || col FROM v;
END;

CREATE FUNCTION use_sql_string() RETURNS TEXT LANGUAGE sql AS $$
SELECT 'use_sql_string = ' || col FROM v;
$$;

CREATE FUNCTION use_plpgsql() RETURNS TEXT LANGUAGE plpgsql AS $$ BEGIN
RETURN (SELECT 'use_plpgsql = ' || col FROM v);
END; $$;

SET search_path TO 'public', 's2';

SELECT use_sql_atomic() AS "output" UNION ALL
SELECT use_sql_string() AS "output" UNION ALL
SELECT use_plpgsql() AS "output";

============

This generates the following output:

output
---------------------------
use_sql_atomic = creation
use_sql_string = runtime
use_plpgsql = runtime
(3 rows)

Overall, PostgreSQL doesn't behave consistent, and to me it seems that
the documentation isn't describing its behavior correctly either.

https://www.postgresql.org/docs/current/sql-createfunction.html

"sql_body

The body of a LANGUAGE SQL function. This can either be a single
statement

RETURN expression

or a block

BEGIN ATOMIC
statement;
statement;
...
statement;
END

This is similar to writing the text of the function body as a
string constant (see definition above), but there are some differences:
This form only works for LANGUAGE SQL, the string constant form works
for all languages. This form is parsed at function definition time, the
string constant form is parsed at execution time; therefore this form
cannot support polymorphic argument types and other constructs that are
not resolvable at function definition time. This form tracks
dependencies between the function and objects used in the function body,
so DROP ... CASCADE will work correctly, whereas the form using string
literals may leave dangling functions. Finally, this form is more
compatible with the SQL standard and other SQL implementations.
"

I understand if fixing this is too much work (even though I would
really like to see this fixed). But given that the current behavior is
highly surprising and inconsistent - and keeping in mind that this is a
subject that may affect security - I think the documentation should
reflect the current behavior at least. I thus see this as a
documentation issue.

Kind regards,
Jan Behrens

--
Adrian Klaver
adrian.klaver@aklaver.com

On Wed, Jan 1, 2025 at 10:55 AM Jan Behrens <jbe-mlist@magnetkern.de> wrote:

On Sat, 28 Dec 2024 00:40:09 +0100
Jan Behrens <jbe-mlist@magnetkern.de> wrote:

On Fri, 27 Dec 2024 13:26:28 -0700
"David G. Johnston" <david.g.johnston@gmail.com> wrote:

Or is it documented somewhere?

https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING

I can't find any notes regarding functions and schemas in that section.

"Because PL/pgSQL saves prepared statements and sometimes execution plans
in this way, SQL commands that appear directly in a PL/pgSQL function must
refer to the same tables and columns on every execution; that is, you
cannot use a parameter as the name of a table or column in an SQL command."

Changing search_path is just one possible way to change out which object a
name tries to refer to so it is not called out explicitly.

"SQL-language and PL-language functions provided by extensions are at
risk of search-path-based attacks when they are executed, since parsing
of these functions occurs at execution time not creation time."

Moreover, it isn't true for all
SQL-language functions, as can be demonstrated with the following code:

Yeah, when we added a second method to write an SQL-language function, one
that doesn't simply accept a string body, we didn't update that section to
point out that is the string input variant of create function that is
affected in this manner, the non-string (atomic) variant stores the result
of parsing the inline code as opposed to storing the raw text.

David J.

On Wed, 1 Jan 2025 11:19:32 -0700
"David G. Johnston" <david.g.johnston@gmail.com> wrote:

On Wed, Jan 1, 2025 at 10:55 AM Jan Behrens <jbe-mlist@magnetkern.de> wrote:

On Sat, 28 Dec 2024 00:40:09 +0100
Jan Behrens <jbe-mlist@magnetkern.de> wrote:

On Fri, 27 Dec 2024 13:26:28 -0700
"David G. Johnston" <david.g.johnston@gmail.com> wrote:

Or is it documented somewhere?

https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING

I can't find any notes regarding functions and schemas in that section.

"Because PL/pgSQL saves prepared statements and sometimes execution plans
in this way, SQL commands that appear directly in a PL/pgSQL function must
refer to the same tables and columns on every execution; that is, you
cannot use a parameter as the name of a table or column in an SQL command."

Changing search_path is just one possible way to change out which object a
name tries to refer to so it is not called out explicitly.

The first part of the cited sentence seems helpful ("you must always
refer to the same tables and columns on every execution"). I would thus
conclude that using a dynamic search_path when running functions or
procedures is *always* considered errorneous (even though not reported
by the database as an error), except when using EXECUTE.

I wonder if the database could/should generate an error (or at least a
warning?) when a function or procedure without a "SET search_path"
statement uses a non-qualified name? According to the documentation
using a dynamic search_path to refer to different entities in the
database is a case that "must" not happen.

But following through, this might lead to more warnings one might
expect, e.g. when using simple operators such as "=" or the "IN" or
"CASE expression WHEN" statements, as these rely on the search_path as
well. Should such code be considered non-idiomatic, dangerous, or even
errorneous if a "SET search_path" option is missing in the
function's/procedure's definition?

Maybe I'm overthinking this. But in practice, I've been running into
surprising issues whenever functions and schemas are involved, and I'm
not sure if every programmer will be aware of how important it is to
properly set a search_path in the function's defintion after reading
the documentation. (Besides, it's not always possible in procedures.)

"SQL-language and PL-language functions provided by extensions are at
risk of search-path-based attacks when they are executed, since parsing
of these functions occurs at execution time not creation time."

Moreover, it isn't true for all
SQL-language functions, as can be demonstrated with the following code:

Yeah, when we added a second method to write an SQL-language function, one
that doesn't simply accept a string body, we didn't update that section to
point out that is the string input variant of create function that is
affected in this manner, the non-string (atomic) variant stores the result
of parsing the inline code as opposed to storing the raw text.

David J.

I missed that other part in the manual (which is in a totally different
section). Should I report the missing update in section 36.17.6.1. of
the documentation as a documentation issue, or is it not necessary?

Kind regards,
Jan Behrens

Hi

čt 2. 1. 2025 v 11:37 odesílatel Jan Behrens <jbe-mlist@magnetkern.de>
napsal:

On Wed, 1 Jan 2025 11:19:32 -0700
"David G. Johnston" <david.g.johnston@gmail.com> wrote:

On Wed, Jan 1, 2025 at 10:55 AM Jan Behrens <jbe-mlist@magnetkern.de>

wrote:

On Sat, 28 Dec 2024 00:40:09 +0100
Jan Behrens <jbe-mlist@magnetkern.de> wrote:

On Fri, 27 Dec 2024 13:26:28 -0700
"David G. Johnston" <david.g.johnston@gmail.com> wrote:

Or is it documented somewhere?

https://www.postgresql.org/docs/current/plpgsql-implementation.html#PLPGSQL-PLAN-CACHING

I can't find any notes regarding functions and schemas in that

section.

"Because PL/pgSQL saves prepared statements and sometimes execution plans
in this way, SQL commands that appear directly in a PL/pgSQL function

must

refer to the same tables and columns on every execution; that is, you
cannot use a parameter as the name of a table or column in an SQL

command."

Changing search_path is just one possible way to change out which object

a

name tries to refer to so it is not called out explicitly.

The first part of the cited sentence seems helpful ("you must always
refer to the same tables and columns on every execution"). I would thus
conclude that using a dynamic search_path when running functions or
procedures is *always* considered errorneous (even though not reported
by the database as an error), except when using EXECUTE.

I wonder if the database could/should generate an error (or at least a
warning?) when a function or procedure without a "SET search_path"
statement uses a non-qualified name? According to the documentation
using a dynamic search_path to refer to different entities in the
database is a case that "must" not happen.

But following through, this might lead to more warnings one might
expect, e.g. when using simple operators such as "=" or the "IN" or
"CASE expression WHEN" statements, as these rely on the search_path as
well. Should such code be considered non-idiomatic, dangerous, or even
errorneous if a "SET search_path" option is missing in the
function's/procedure's definition?

Maybe I'm overthinking this. But in practice, I've been running into
surprising issues whenever functions and schemas are involved, and I'm
not sure if every programmer will be aware of how important it is to
properly set a search_path in the function's defintion after reading
the documentation. (Besides, it's not always possible in procedures.)

How can you identify unwanted usage of non qualified identifiers from
wanted usage of non qualified identifiers? It is a common pattern for
sharding. Using not qualified identifiers of operators, functions is common
when you are using orafce extensions, etc.

Using qualified identifiers everywhere strongly reduces readability. There
are no aliases to the schema, so aliases cannot help.

you can identify the functions where search_path is not explicitly assigned

select oid::regprocedure
from pg_proc
where pronamespace::regnamespace not in ('pg_catalog',
'information_schema')
and not exists(select 1 from unnest(proconfig) g(v) where v ~
'^search_path');

Regards

Pavel

On Thu, 2 Jan 2025 12:40:59 +0100
Pavel Stehule <pavel.stehule@gmail.com> wrote:

How can you identify unwanted usage of non qualified identifiers from
wanted usage of non qualified identifiers? It is a common pattern for
sharding. Using not qualified identifiers of operators, functions is common
when you are using orafce extensions, etc.

I don't fully understand the use-case. Could you elaborate?

As I understand, even if identifiers are not fully-qualified, it is
forbidden to use the search_path to refer to different database
entities at run-time (as David pointed out).

So I don't understand how a dynamic "search_path" could be used in any
scenario within functions except when EXECUTE is involved.

Using qualified identifiers everywhere strongly reduces readability. There
are no aliases to the schema, so aliases cannot help.

Yes, I agree on that. Using "SET search_path" in the function's
definition fixes that problem, but it's easy to miss how important this
is from reading the documentation:

The manual regarding "CREATE FUNCTION" refers to "search_path" only
within the "Writing SECURITY DEFINER Functions Safely" section. It's
easy to skip that part unless you use that feature. Moreover, that
section alone doesn't explain the weird behavior of four different
outcomes of a function with only two schemas involved which I brought
up in the beginning of this thread.

The part on "SET configuration_parameter" part in the "CREATE FUNCTION"
documentation doesn't mention the search_path or schemas. And I don't
think you can expect every programmer will read the "Plan Caching"
subsection in the "PL/pgSQL under the Hood" section. But even then, the
information is just provided indirectly.

Searching for "schema" in "CREATE FUNCTION"'s documentation doesn't
give any hint either.

I think (assuming that the behavior isn't fixed) that some slighly more
prominent warning would be reasonable.

you can identify the functions where search_path is not explicitly assigned

select oid::regprocedure
from pg_proc
where pronamespace::regnamespace not in ('pg_catalog',
'information_schema')
and not exists(select 1 from unnest(proconfig) g(v) where v ~
'^search_path');

Regards

Pavel

Kind regards,
Jan

čt 2. 1. 2025 v 13:15 odesílatel Jan Behrens <jbe-mlist@magnetkern.de>
napsal:

On Thu, 2 Jan 2025 12:40:59 +0100
Pavel Stehule <pavel.stehule@gmail.com> wrote:

How can you identify unwanted usage of non qualified identifiers from
wanted usage of non qualified identifiers? It is a common pattern for
sharding. Using not qualified identifiers of operators, functions is

common

when you are using orafce extensions, etc.

I don't fully understand the use-case. Could you elaborate?

As I understand, even if identifiers are not fully-qualified, it is
forbidden to use the search_path to refer to different database
entities at run-time (as David pointed out).

So I don't understand how a dynamic "search_path" could be used in any
scenario within functions except when EXECUTE is involved.

you don't need more databases

schema one - customer x
schema two - customer y

create table one.t1(..); create table one.t2(..);
create table two.t1(..); create table two.t2(..);

set search_path to one;
-- work with data set of customer x

set search_path to two;
-- work wit data set of customer y

some times can be pretty ineffective to have database per customer - more
connect, disconnect in postgres is much more expensive than SET search_path
TO .. and maybe RESET plans;

Using qualified identifiers everywhere strongly reduces readability.

There

are no aliases to the schema, so aliases cannot help.

Yes, I agree on that. Using "SET search_path" in the function's
definition fixes that problem, but it's easy to miss how important this
is from reading the documentation:

The manual regarding "CREATE FUNCTION" refers to "search_path" only
within the "Writing SECURITY DEFINER Functions Safely" section. It's
easy to skip that part unless you use that feature. Moreover, that
section alone doesn't explain the weird behavior of four different
outcomes of a function with only two schemas involved which I brought
up in the beginning of this thread.

The part on "SET configuration_parameter" part in the "CREATE FUNCTION"
documentation doesn't mention the search_path or schemas. And I don't
think you can expect every programmer will read the "Plan Caching"
subsection in the "PL/pgSQL under the Hood" section. But even then, the
information is just provided indirectly.

yes, probably nobody reads the plan caching doc. And if they read it, then
because they have performance problems.

Searching for "schema" in "CREATE FUNCTION"'s documentation doesn't
give any hint either.

This is a question - this is a generic feature in Postgres. Every query
can be impacted by setting of search_path.

From my perspective, there can be a note in the documentation related to
copy types and row types.

https://www.postgresql.org/docs/current/plpgsql-declarations.html#PLPGSQL-DECLARATION-TYPE

The problem that you found is not just about the change of search_path.
Same problem can be found after altering the table.

Regards

Pavel

On Thu, 2 Jan 2025 13:48:29 +0100
Pavel Stehule <pavel.stehule@gmail.com> wrote:

čt 2. 1. 2025 v 13:15 odesílatel Jan Behrens <jbe-mlist@magnetkern.de>
napsal:

On Thu, 2 Jan 2025 12:40:59 +0100
Pavel Stehule <pavel.stehule@gmail.com> wrote:

How can you identify unwanted usage of non qualified identifiers from
wanted usage of non qualified identifiers? It is a common pattern for
sharding. Using not qualified identifiers of operators, functions is

common

when you are using orafce extensions, etc.

I don't fully understand the use-case. Could you elaborate?

As I understand, even if identifiers are not fully-qualified, it is
forbidden to use the search_path to refer to different database
entities at run-time (as David pointed out).

So I don't understand how a dynamic "search_path" could be used in any
scenario within functions except when EXECUTE is involved.

you don't need more databases

schema one - customer x
schema two - customer y

create table one.t1(..); create table one.t2(..);
create table two.t1(..); create table two.t2(..);

set search_path to one;
-- work with data set of customer x

set search_path to two;
-- work wit data set of customer y

some times can be pretty ineffective to have database per customer - more
connect, disconnect in postgres is much more expensive than SET search_path
TO .. and maybe RESET plans;

I guess that means there is a practical application where search_path
MAY change at runtime IF done in different sessions or if the cache is
reset using the DISCARD command:

https://www.postgresql.org/docs/17/sql-discard.html

I assume DISCARD PLANS would be the right command?

This seems to be a very special case though. I think there should be a
warning in the documentation of CREATE FUNCTION with regard to schemas
anyway, though.

Regards,
Jan

Hi

some times can be pretty ineffective to have database per customer - more
connect, disconnect in postgres is much more expensive than SET

search_path

TO .. and maybe RESET plans;

I guess that means there is a practical application where search_path
MAY change at runtime IF done in different sessions or if the cache is
reset using the DISCARD command:

https://www.postgresql.org/docs/17/sql-discard.html

I assume DISCARD PLANS would be the right command?

that depends. plan inside plan cache is invalidated when search_path is
different. You use RESET plans because you want to release all plans
quickly.

Unfortunately, the types assigned to plpgsql variables are not invalidated.
This is the source of problems. It is a classical problem - it is hard to
say when you should invalidate cache.
Current design is not ideal - but it is almost a good enough compromise
between correctness and performance. It is true, so nobody did some work to
fix it. So maybe the impact to performance should not be too bad, but it is
not an easy issue. plans are isolated - and the impact of one plan to the
second plan is zero. For variables it is exactly opposite.

This seems to be a very special case though. I think there should be a
warning in the documentation of CREATE FUNCTION with regard to schemas
anyway, though.

I am not sure. If you want to use this warning, then it should be
everywhere where any non-qualified identifier can be used. Maybe in plpgsql
can be more accented so almost everything in plpgsql depends on the current
setting of search_path. Lot of people don't understand, so every expression
in plpgsql is SQL and every expression is executed like part of a query.
And unfortunately there are some different caches - plpgsql cache and plan
cache and both caches are invalidated at different times (I think so
plpgsql cache is not resetted by RESET PLANS). Maybe it is better to
explain how plpgsql works. It is a little bit different from well known
interpreted languages.

On Sat, 28 Dec 2024 00:40:09 +0100
Jan Behrens <jbe-mlist@magnetkern.de> wrote:

Add qualification or attach a “set search_path” clause to “create
function”. Code stored in the server should not rely on the session
search_path.

David J.

I have been trying to adjust some of my code, and I still have cases
where I have to rely on the session's search_path. I'll provide an
example below.

[...]

My question is: Am I safe if I use fully-qualified types in the DECLARE
section only? Or do I need to provide full qualification also in the
code below (after SET search_path TO 'myschema')?

And bonus question: Is it documented somewhere?

[...]

Kind Regards
Jan Behrens

The following code is taken from a project I'm currently working on:

============

-- Let's assume we don't know the name of the schema in which the
-- "pgratio" extension with the RATIONAL data type is installed.
CREATE SCHEMA "qwertyuiop";
CREATE EXTENSION "pgratio" WITH SCHEMA "qwertyuiop";

-- This installs schema "myschema" with some dynamic function:
BEGIN;

CREATE SCHEMA "myschema";
SET LOCAL search_path TO "myschema";

-- Append schema of "pgratio" extension, which provides the RATIONAL
-- data type, to search_path:
SELECT set_config(
'search_path',
current_setting('search_path') || ', ' || quote_ident(nspname),
TRUE
) FROM pg_namespace, pg_extension
WHERE pg_namespace.oid = extnamespace AND extname = 'pgratio';

CREATE DOMAIN "rational_wrapper" AS RATIONAL;

CREATE FUNCTION "some_function" ("query_p" TEXT) RETURNS RATIONAL

--------------------------------------------------------------------
-- I cannot use SET search_path FROM CURRENT here, because "query_p"
-- shall refer to tables in the search_path of the caller.
--------------------------------------------------------------------

LANGUAGE plpgsql AS $$
DECLARE
"old_search_path" TEXT;

----------------------------------------------------------------
-- I have to fully qualify the following type.
-- Moreover, I can't use RATIONAL as I don't know its schema.
----------------------------------------------------------------

"result" "myschema"."rational_wrapper";
BEGIN
SELECT current_setting('search_path') INTO "old_search_path";
PERFORM set_config(
'search_path',
'myschema, ' || quote_ident(nspname) || ', pg_temp, ' ||
"old_search_path",
TRUE
) FROM pg_namespace, pg_extension
WHERE pg_namespace.oid = extnamespace AND extname = 'pgratio';

----------------------------------------------------------------
-- Is it safe to not fully qualify type RATIONAL below?
-- And, if yes, where in the documentation is this explained?
----------------------------------------------------------------

CREATE TEMP TABLE "mytemptable" ("val" RATIONAL);
EXECUTE 'INSERT INTO "mytemptable" '
'SELECT "query"."a" * "query"."b" '
'FROM (' || "query_p" || ') AS "query"';
-- Do some things here.
SELECT sum("val") INTO "result" FROM "mytemptable";
PERFORM set_config('search_path', "old_search_path", TRUE);
RETURN "result";
END;
$$;

COMMIT;

CREATE TABLE "tbl" ("foo" INT8, "bar" INT8);
INSERT INTO "tbl" VALUES (5, 7), (1, 10);

SELECT "myschema"."some_function"(
'SELECT "foo" AS "a", "bar" AS "b" FROM "tbl"'
);
\c
SELECT "myschema"."some_function"(
'SELECT "foo" AS "a", "bar" AS "b" FROM "tbl"'
);

============

The code for the pgratio extension that provides the RATIONAL data type
is found here: https://www.public-software-group.org/pgratio

Running that code on my machine correctly gives:

some_function
---------------
45
(1 row)

You are now connected to database "jbe" as user "jbe".
some_function
---------------
45
(1 row)

Because extensions can only be installed in one schema, it may be a bad
idea to have a component requiring an extension to be installed in a
particular schema (because if different components have different
expectations on the schema name, e.g. some might expect "pgratio" to be
installed in "public" and others might expect it in "pgratio" or some
other schema such as "qwertyuiop", this would lead to an unresolvable
conflict).

I would like to know if the above example is correct. It seems overall
bulky, but I haven't found a better way, assuming that it can be
unknown where a particular extension has been installed to. In
particular I feel a bit insecure about where I have to fully qualify,
and where not. See the comments in the code above.

Note that I want the function to accept a query that makes sense in the
caller's search_path. Thus using "SET search_path FROM CURRENT" is not
an option for me, I believe.

Regards,
Jan Behrens

On Friday, January 3, 2025, Jan Behrens <jbe-mlist@magnetkern.de> wrote:

I would like to know if the above example is correct. It seems overall
bulky, but I haven't found a better way, assuming that it can be
unknown where a particular extension has been installed to. In
particular I feel a bit insecure about where I have to fully qualify,
and where not. See the comments in the code above.

Short answer, you cannot looking at a definition and know the answer -
whether the code is going to be executed in a sanitized search_path is what
matters. Anything that would be executed during pg_restore has to be made
safe. Therefore, code that is only ever executed by applications directly
can use swarch_path.

I’d probably modify the function signature to take search_path as a second
optional argument and then invoke a set search_path within the function.
At worse the caller can place current_setting(search_path) as the value of
that argument though being explicit would be recommended.

David J.

On Fri, 3 Jan 2025 08:34:57 -0700
"David G. Johnston" <david.g.johnston@gmail.com> wrote:

On Friday, January 3, 2025, Jan Behrens <jbe-mlist@magnetkern.de> wrote:

I would like to know if the above example is correct. It seems overall
bulky, but I haven't found a better way, assuming that it can be
unknown where a particular extension has been installed to. In
particular I feel a bit insecure about where I have to fully qualify,
and where not. See the comments in the code above.

Short answer, you cannot looking at a definition and know the answer -
whether the code is going to be executed in a sanitized search_path is what
matters.

I don't understand. Do you mean my last example is wrong / insecure?
If so, why?

Anything that would be executed during pg_restore has to be made
safe. Therefore, code that is only ever executed by applications directly
can use swarch_path.

Why should the function be executed during pg_restore?

I’d probably modify the function signature to take search_path as a second
optional argument and then invoke a set search_path within the function.
At worse the caller can place current_setting(search_path) as the value of
that argument though being explicit would be recommended.

David J.

I could do that, but I would like to understand if that is really
necessary as it makes the interface more complicated, and I would like
to avoid unnecessary complexity in my interface.

Is it really impossible to have functions without SET search_path in
the definition of a PL/pgSQL function if I fully-qualify all types in
the DECLARE section and if all other non-qualified identifiers occur
after set_config('search_path', ...)?

Kind regards,
Jan Behrens