Thread: Leading comments and client applications

Hi,
I'm trying to understand a behavior where, with our Postgres client, a leading comment in a SQL script causes the
CREATEFUNCTION statement following it to be not executed. I can't figure out if this is a bug somewhere or just a
misunderstandingon my part. I would appreciate some help understanding. 

Here's the contents of foo.sql --

-- this is a comment
CREATE FUNCTION foo(bar text) RETURNS text AS $$
    SELECT bar
$$
LANGUAGE sql IMMUTABLE PARALLEL SAFE
;


When I feed that to 'psql -f foo.sql', the function is created as I expect. In the Postgres log, the leading comment
*doesn't*appear. I see the same behavior if I just copy/paste the function into psql. 

Our test system uses Python 3.8, SQLAlchemy 1.3.6, and psycopg 2.8.5, and when our test harness reads foo.sql and
passesit to SQLAlchemy's execute(), I can see in the Postgres log that the leading comment is *not* stripped, and the
functionisn't created. 

The server is Postgres 11.

My naive interpretation is that one of the client layers (SQLAlchemy or psycopg2) should be stripping the leading
commentbut isn't, but that seems like a lot of responsibility to push onto a client application. I figured that would
bethe responsibility of the Postgres parser.  

I'd be grateful for any insights about what I'm missing.

Thanks
Philip

Philip Semanchuk <philip@americanefficient.com> writes:
> I'm trying to understand a behavior where, with our Postgres client, a leading comment in a SQL script causes the
CREATEFUNCTION statement following it to be not executed. I can't figure out if this is a bug somewhere or just a
misunderstandingon my part. I would appreciate some help understanding. 

Are you certain there's actually a newline after the comment?
The easiest explanation for this would be if something in the
SQLAlchemy code path were munging the newline.

A completely different line of thought is that the function
*does* get created, but inside a transaction that never
gets committed.  Either way, I think this is mainly a SQLAlchemy
question not a Postgres question.

As far as the comparison behavior goes, psql's parser strips
comments that start with double dashes, for $obscure_reasons.
The server is perfectly capable of ignoring those by itself,
though.  (Awhile back I tried to remove that psql behavior,
but it caused too much churn in our regression tests.)

            regards, tom lane

> As far as the comparison behavior goes, psql's parser strips
> comments that start with double dashes, for $obscure_reasons.

That story aught to be worth a $beer or two
> The server is perfectly capable of ignoring those by itself,
> though.  (Awhile back I tried to remove that psql behavior,
> but it caused too much churn in our regression tests.)
>
>             regards, tom lane
>
>

On 2022-03-25 11:32:24 -0400, Philip Semanchuk wrote:
> I'm trying to understand a behavior where, with our Postgres client, a
> leading comment in a SQL script causes the CREATE FUNCTION statement
> following it to be not executed. I can't figure out if this is a bug
> somewhere or just a misunderstanding on my part. I would appreciate
> some help understanding.
>
> Here's the contents of foo.sql --
>
> -- this is a comment
> CREATE FUNCTION foo(bar text) RETURNS text AS $$
>     SELECT bar
> $$
> LANGUAGE sql IMMUTABLE PARALLEL SAFE
> ;
>
>
> When I feed that to 'psql -f foo.sql', the function is created as I
> expect. In the Postgres log, the leading comment *doesn't* appear. I
> see the same behavior if I just copy/paste the function into psql.
>
> Our test system uses Python 3.8, SQLAlchemy 1.3.6, and psycopg 2.8.5,
> and when our test harness reads foo.sql and passes it to SQLAlchemy's
> execute(), I can see in the Postgres log that the leading comment is
> *not* stripped, and the function isn't created.

I cannot reproduce this with plain psycopg:

% cat foo
#!/usr/bin/python3

import psycopg2

db = psycopg2.connect("")
csr = db.cursor()

csr.execute(
        """
        -- this is a comment
        CREATE FUNCTION foo(bar text) RETURNS text AS $$
            SELECT bar
        $$
        LANGUAGE sql IMMUTABLE PARALLEL SAFE
        """)
db.commit()
% ./foo
% psql
psql (13.6 (Ubuntu 13.6-1.pgdg20.04+1), server 11.15 (Ubuntu 11.15-1.pgdg20.04+1))
Type "help" for help.

hjp=> \df foo
                        List of functions
╔════════╤══════╤══════════════════╤═════════════════════╤══════╗
║ Schema │ Name │ Result data type │ Argument data types │ Type ║
╟────────┼──────┼──────────────────┼─────────────────────┼──────╢
║ public │ foo  │ text             │ bar text            │ func ║
╚════════╧══════╧══════════════════╧═════════════════════╧══════╝
(1 row)

hjp=> select foo('x*');
╔═════╗
║ foo ║
╟─────╢
║ x*  ║
╚═════╝
(1 row)

Time: 1.296 ms
hjp=> \q

So like others I suspect that SQLAlchemy is doing something weird here.

        hp

--
   _  | Peter J. Holzer    | Story must make more sense than reality.
|_|_) |                    |
| |   | hjp@hjp.at         |    -- Charles Stross, "Creative writing
__/   | http://www.hjp.at/ |       challenge!"

> On Mar 25, 2022, at 11:59 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>
> Philip Semanchuk <philip@americanefficient.com> writes:
>> I'm trying to understand a behavior where, with our Postgres client, a leading comment in a SQL script causes the
CREATEFUNCTION statement following it to be not executed. I can't figure out if this is a bug somewhere or just a
misunderstandingon my part. I would appreciate some help understanding. 
>
> Are you certain there's actually a newline after the comment?
> The easiest explanation for this would be if something in the
> SQLAlchemy code path were munging the newline.

I verified that there is a newline after the comment. But yes, thanks to your suggestion and others, I was able to
narrowthis down to something in SQLAlchemy behavior. In case anyone else comes across this and is wondering -- 

In addition to accepting a plain string, execute() accepts a number of different SQLAlchemy data types, including
TextClauseand DDLElement. We used to pass a DDLElement to execute(), but a few months ago we switched to passing a
TextClausebecause DDLElement interprets % signs anywhere in SQL scripts as Python string interpolation markers and that
wascausing us headaches in some scripts. Something about the way TextClause changes the raw SQL string causes the
behaviorI’m seeing, although we didn’t notice it at the time of the changeover. I don’t know what exactly it’s doing
yet,but when I switch back to passing a DDLElement to execute(), my SQL function is created as I expected.  

https://docs.sqlalchemy.org/en/13/core/connections.html#sqlalchemy.engine.Connection.execute

As David J pointed out, execute() is deprecated as of version 1.4. We’re still on 1.3 but we’ll have to move away from
thiscode eventually so maybe this is a good inspiration to move away from execute() now and reduce the number of
deprecationwarnings we have to deal with in the future. 


> As far as the comparison behavior goes, psql's parser strips
> comments that start with double dashes, for $obscure_reasons.
> The server is perfectly capable of ignoring those by itself,
> though.  (Awhile back I tried to remove that psql behavior,
> but it caused too much churn in our regression tests.)


Thanks, this is most helpful. I use psql to double check I think SQLAlchemy is doing something odd. It’s good to know
thatpsql's behavior in this case is a choice and not required behavior for clients. Peter J. Holzer’s psycopg2 example
couldhave showed me the same; I wish I had thought of that. 


I appreciate all the help!

Cheers
Philip

Rob Sargent <robjsargent@gmail.com> writes:
>> As far as the comparison behavior goes, psql's parser strips
>> comments that start with double dashes, for $obscure_reasons.

> That story aught to be worth a $beer or two

Hmm.  The original reasoning is lost in the mists of time;
I dug in our git history and traced the behavior as far back
as a45195a19 of 1999-11-04, but I'll bet a nickel that Peter
doesn't remember exactly why he did that.

But I can show you why I gave up on removing the behavior:
it's an important part of psql's strategy of discarding
leading whitespace before a query.  Our regression test
scripts are full of cases like

-- comments here

SELECT intentionally-wrong-query;

and what they're expecting to get from that is output like

ERROR:  column "intentionally" does not exist
LINE 1: SELECT intentionally-wrong-query;
               ^

When I changed psql's parser to not remove comments, that output
suddenly changed to say "LINE 3:", because now the query string
sent to the server included the "-- comments here" line as well
as the blank line after it.  While we could have changed all the
expected output, or changed how the server counts lines within
a query, we concluded that this would confuse too many people and
break too many applications; so we left it alone.

(As of v15, psql will send -- comments that begin *after* the
first non-whitespace token of a query [1].  But leading comments
and whitespace will still get stripped.)

            regards, tom lane

[1] https://git.postgresql.org/gitweb/?p=postgresql.git&a=commitdiff&h=83884682f4df96184549b91869a1cf79dafb4f94

Rob Sargent <robjsargent@gmail.com> writes:

As far as the comparison behavior goes, psql's parser strips
comments that start with double dashes, for $obscure_reasons.

That story aught to be worth a $beer or two

Hmm.  The original reasoning is lost in the mists of time;
I dug in our git history and traced the behavior as far back
as a45195a19 of 1999-11-04, but I'll bet a nickel that Peter
doesn't remember exactly why he did that.

But I can show you why I gave up on removing the behavior:
it's an important part of psql's strategy of discarding
leading whitespace before a query.  Our regression test
scripts are full of cases like

-- comments here

SELECT intentionally-wrong-query;

and what they're expecting to get from that is output like

ERROR:  column "intentionally" does not exist
LINE 1: SELECT intentionally-wrong-query;               ^

When I changed psql's parser to not remove comments, that output
suddenly changed to say "LINE 3:", because now the query string
sent to the server included the "-- comments here" line as well
as the blank line after it.  While we could have changed all the
expected output, or changed how the server counts lines within
a query, we concluded that this would confuse too many people and
break too many applications; so we left it alone.

(As of v15, psql will send -- comments that begin *after* the
first non-whitespace token of a query [1].  But leading comments
and whitespace will still get stripped.)
			regards, tom lane

[1] https://git.postgresql.org/gitweb/?p=postgresql.git&a=commitdiff&h=83884682f4df96184549b91869a1cf79dafb4f94

On Fri, 2022-03-25 at 13:35 -0600, Rob Sargent wrote:
> On 3/25/22 13:30, Tom Lane wrote:
> > Bob Sargent <robjsargent@gmail.com> writes:
> > > That story aught to be worth a $beer or two
> >
> > Hmm. [story]
>
>  Thank you for the indulgence! I clearly owe you (another) one.

I think Tom prefers wine: https://postgr.es/m/14929.1358317689@sss.pgh.pa.us

Yours,
Laurenz Albe

>Something about the way TextClause changes the raw SQL string causes the behavior I’m seeing, although we didn’t notice it at the time of the changeover.
>I don’t know what exactly it’s doing yet, but when I switch back to passing a DDLElement to execute(), my SQL function is created as I expected. 

query = em.createNativeQuery("/*requete_enregistree_num_" + requete.getId() + "*/ " + requete.getReqRequete().trim());

On Mar 25, 2022, at 11:59 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:

Philip Semanchuk <philip@americanefficient.com> writes:

I'm trying to understand a behavior where, with our Postgres client, a leading comment in a SQL script causes the CREATE FUNCTION statement following it to be not executed. I can't figure out if this is a bug somewhere or just a misunderstanding on my part. I would appreciate some help understanding.

Are you certain there's actually a newline after the comment?
The easiest explanation for this would be if something in the
SQLAlchemy code path were munging the newline.

I verified that there is a newline after the comment. But yes, thanks to your suggestion and others, I was able to narrow this down to something in SQLAlchemy behavior. In case anyone else comes across this and is wondering --

In addition to accepting a plain string, execute() accepts a number of different SQLAlchemy data types, including TextClause and DDLElement. We used to pass a DDLElement to execute(), but a few months ago we switched to passing a TextClause because DDLElement interprets % signs anywhere in SQL scripts as Python string interpolation markers and that was causing us headaches in some scripts. Something about the way TextClause changes the raw SQL string causes the behavior I’m seeing, although we didn’t notice it at the time of the changeover. I don’t know what exactly it’s doing yet, but when I switch back to passing a DDLElement to execute(), my SQL function is created as I expected. 

https://docs.sqlalchemy.org/en/13/core/connections.html#sqlalchemy.engine.Connection.execute

As David J pointed out, execute() is deprecated as of version 1.4. We’re still on 1.3 but we’ll have to move away from this code eventually so maybe this is a good inspiration to move away from execute() now and reduce the number of deprecation warnings we have to deal with in the future.

As far as the comparison behavior goes, psql's parser strips
comments that start with double dashes, for $obscure_reasons.
The server is perfectly capable of ignoring those by itself,
though.  (Awhile back I tried to remove that psql behavior,
but it caused too much churn in our regression tests.)

Thanks, this is most helpful. I use psql to double check I think SQLAlchemy is doing something odd. It’s good to know that psql's behavior in this case is a choice and not required behavior for clients. Peter J. Holzer’s psycopg2 example could have showed me the same; I wish I had thought of that.


I appreciate all the help!

Cheers
Philip

> On Mar 28, 2022, at 5:42 AM, Philippe Doussot <philippe.doussot@arche-mc2.fr> wrote:
>
> >Something about the way TextClause changes the raw SQL string causes the behavior I’m seeing, although we didn’t
noticeit at the time of the changeover. 
> >I don’t know what exactly it’s doing yet, but when I switch back to passing a DDLElement to execute(), my SQL
functionis created as I expected.  
>
>
> Alternate option if you want continue to use  TextClause:
>
> use /* comment */ for first prefix comment.
>
> Comment is logged and query executed (tested on Java ( not on SQLAlchemy )).
> We use it to track back the request id executed like that
>
> query = em.createNativeQuery("/*requete_enregistree_num_" + requete.getId() + "*/ " +
requete.getReqRequete().trim());

Thanks for the suggestion! In my testing, both single line and multiline comment blocks cause the same problem for me.
I*was* able to resolve this with a simple change. I was calling SQLAlchemy’s engine.execute(). When I call
connection.execute()instead, the problem resolves. This also solves a future deprecation problem for us.
engine.execute()is deprecated in SQLAlchemy 1.4, but connection.execute() is not. 

I didn’t expect this to fix the problem. There’s no difference in the Postgres log that I can see, so I think the SQL
thatSQLAlchemy sends to postgres is the same. If it’s a commit/transaction problem, it should affect all of our
functionsequally, not just the ones that start with comments.  

I clearly don’t understand this problem fully. Although I'm curious about it, I’m eager to move on to other things. I
planto proceed with this fix and not investigate any more.  

THanks everyone for all the help and suggestions

Cheers
Philip



>
> On 25/03/2022 19:05, Philip Semanchuk wrote:
>>
>>> On Mar 25, 2022, at 11:59 AM, Tom Lane <tgl@sss.pgh.pa.us>
>>>  wrote:
>>>
>>> Philip Semanchuk
>>> <philip@americanefficient.com>
>>>  writes:
>>>
>>>> I'm trying to understand a behavior where, with our Postgres client, a leading comment in a SQL script causes the
CREATEFUNCTION statement following it to be not executed. I can't figure out if this is a bug somewhere or just a
misunderstandingon my part. I would appreciate some help understanding. 
>>>>
>>> Are you certain there's actually a newline after the comment?
>>> The easiest explanation for this would be if something in the
>>> SQLAlchemy code path were munging the newline.
>>>
>> I verified that there is a newline after the comment. But yes, thanks to your suggestion and others, I was able to
narrowthis down to something in SQLAlchemy behavior. In case anyone else comes across this and is wondering -- 
>>
>> In addition to accepting a plain string, execute() accepts a number of different SQLAlchemy data types, including
TextClauseand DDLElement. We used to pass a DDLElement to execute(), but a few months ago we switched to passing a
TextClausebecause DDLElement interprets % signs anywhere in SQL scripts as Python string interpolation markers and that
wascausing us headaches in some scripts. Something about the way TextClause changes the raw SQL string causes the
behaviorI’m seeing, although we didn’t notice it at the time of the changeover. I don’t know what exactly it’s doing
yet,but when I switch back to passing a DDLElement to execute(), my SQL function is created as I expected.  
>>
>>
>> https://docs.sqlalchemy.org/en/13/core/connections.html#sqlalchemy.engine.Connection.execute
>>
>>
>> As David J pointed out, execute() is deprecated as of version 1.4. We’re still on 1.3 but we’ll have to move away
fromthis code eventually so maybe this is a good inspiration to move away from execute() now and reduce the number of
deprecationwarnings we have to deal with in the future. 
>>
>>
>>
>>> As far as the comparison behavior goes, psql's parser strips
>>> comments that start with double dashes, for $obscure_reasons.
>>> The server is perfectly capable of ignoring those by itself,
>>> though.  (Awhile back I tried to remove that psql behavior,
>>> but it caused too much churn in our regression tests.)
>>>
>>
>> Thanks, this is most helpful. I use psql to double check I think SQLAlchemy is doing something odd. It’s good to
knowthat psql's behavior in this case is a choice and not required behavior for clients. Peter J. Holzer’s psycopg2
examplecould have showed me the same; I wish I had thought of that. 
>>
>>
>> I appreciate all the help!
>>
>> Cheers
>> Philip
>>
>>
>>
>>
>>
>
>
> --
>
> 📌 Le nom de domaine de nos adresses mails évolue et devient @arche-mc2.fr.
>
>
> arche-mc2.fr
>
>
>
>
> Philippe DOUSSOT
>
> ARCHITECTE TECHNIQUE
>
> DIRECTION DES SOLUTIONS ARCHE MC2 DOMICILE
>
> philippe.doussot@arche‑mc2.fr
>