BUG #17842: Adding a qual to a working query gets bogus syntax error

PG Bug reporting form <noreply@postgresql.org> writes:

During refactoring our application I got a bogus syntax error when adding a
simple where-clause to a working query.

My guess is that this:

('0'||split_part(split_part(nline,'(',2),')',1))::smallint as nlen

sometimes produces a string that fails to cast to smallint, which you
accidentally don't notice because it never gets evaluated for troublesome
values of nline --- until you add the "nlen > 0" condition. There isn't
anything in your query that prevents that from being evaluated fairly
early.

This isn't a bug, or at least you have provided no data that would
motivate (or indeed allow) anyone else to poke into it more closely.

regards, tom lane

Hello Tom,

thanks for the quick response.

You are probably right with your assumption of a not convertible substring, but some questions remain:

1. When the question runs succesfully, all lines are subject to the join of qj (newcol_imp and oldcol_imp heve exact the same number of lines, matching logically), but why does a limitation of the result set trigger the error?

(There really may be some substrings from declarations of numeric(4,1) which are not excluded properly from conversion to smallint.

The '0'||<rest> tries to handle an empty string)

2. Why is it reporting an syntax error (supposed at parse time) and not a runtime error (supposed at execution time)

3. Why does an uncorrelated qual like and onum < 100 trigger an error too?

( I know that a limitation can change the choosen plan, but which one cannot be seen because of the error.)

The behavior of having a succesfull full query, but getting the error with some quals seems to me very unintuitive.

PS: The exactness and correctness of this specific query is certainly not overall important for me, but I wanted to point out this very unexpected behavior for some other users.

When I find some time, I'll try to provide some anonymized data.

Best regards

Hans Buschmann

________________________________
Von: Tom Lane <tgl@sss.pgh.pa.us>
Gesendet: Dienstag, 14. März 2023 17:20
An: Hans Buschmann
Cc: pgsql-bugs@lists.postgresql.org
Betreff: Re: BUG #17842: Adding a qual to a working query gets bogus syntax error

PG Bug reporting form <noreply@postgresql.org> writes:

During refactoring our application I got a bogus syntax error when adding a
simple where-clause to a working query.

My guess is that this:

('0'||split_part(split_part(nline,'(',2),')',1))::smallint as nlen

sometimes produces a string that fails to cast to smallint, which you
accidentally don't notice because it never gets evaluated for troublesome
values of nline --- until you add the "nlen > 0" condition. There isn't
anything in your query that prevents that from being evaluated fairly
early.

This isn't a bug, or at least you have provided no data that would
motivate (or indeed allow) anyone else to poke into it more closely.

regards, tom lane

On Tue, Mar 14, 2023, 11:12 Hans Buschmann <buschmann@nidsa.net> wrote:

2. Why is it reporting an syntax error (supposed at parse time) and not a
runtime error (supposed at execution time)

Your parsed query doesn't have a syntax error. Data it processes does,
which is found when parsing that data during query execution.

David J.

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

On Tue, Mar 14, 2023, 11:12 Hans Buschmann <buschmann@nidsa.net> wrote:

2. Why is it reporting an syntax error (supposed at parse time) and not a
runtime error (supposed at execution time)

Your parsed query doesn't have a syntax error. Data it processes does,
which is found when parsing that data during query execution.

Right. It's kind of unfortunate perhaps that we use SQLSTATEs from
the "syntax error" class for malformed data input, but that seems
unlikely to change. In any case, it's pretty obvious that this error
is not complaining about anything in the text of the query.

regards, tom lane

Hello Tom,

I totally forgot to apply explain instead of explain analyze,

Here EXPLAIN for the erroneous question:

cpsdb=# explain -- select * from ( -- select count(*) from ( -- select length(sel) from (
cpsdb-# select
cpsdb-# onum
cpsdb-# ,'column'||(onum::varchar) as vname
cpsdb-# ,vlen
cpsdb-# ,nlen
cpsdb-# ,olen
cpsdb-# ,NULLIF(vlen-olen,0) as delta_len
cpsdb-# from (
cpsdb(# select *
cpsdb(# ,('0'||split_part(split_part(nline,'(',2),')',1))::smallint as nlen
cpsdb(# ,('0'||split_part(split_part(oline,'(',2),')',1))::smallint as olen
cpsdb(# from newcol_imp
cpsdb(# join oldcol_imp on onum=nnum
cpsdb(# join (
cpsdb(# select
cpsdb(# vnum
cpsdb(# ,split_part(vline,' ',1) as vname
cpsdb(# ,('0'||split_part(split_part(vline,'(',2),')',1))::smallint as vlen
cpsdb(# from varchar_imp
cpsdb(# ) qv on nline like '%'||vname||'%'
cpsdb(# where nline not like '%KEY%'
cpsdb(# ) qj
cpsdb-# --limit 30
cpsdb-# where vlen!=olen
cpsdb-# and nlen > 0
cpsdb-# ;
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Merge Join (cost=0.56..10438.53 rows=2209 width=44)
Merge Cond: (newcol_imp.nnum = oldcol_imp.onum)
Join Filter: ((('0'::text || split_part(split_part(varchar_imp.vline, '('::text, 2), ')'::text, 1)))::smallint <> (('0'::text || split_part(split_part(oldcol_imp.oline, '('::text, 2), ')'::text, 1)))::smallint)
-> Nested Loop (cost=0.28..10111.27 rows=2220 width=77)
Join Filter: (newcol_imp.nline ~~ (('%'::text || split_part(varchar_imp.vline, ' '::text, 1)) || '%'::text))
-> Index Scan using newcol_imp_pkey on newcol_imp (cost=0.28..107.37 rows=745 width=38)
Filter: ((nline !~~ '%KEY%'::text) AND ((('0'::text || split_part(split_part(nline, '('::text, 2), ')'::text, 1)))::smallint > 0))
-> Materialize (cost=0.00..14.94 rows=596 width=39)
-> Seq Scan on varchar_imp (cost=0.00..11.96 rows=596 width=39)
-> Index Scan using oldcol_imp_pkey on oldcol_imp (cost=0.28..66.86 rows=2372 width=44)
(10 Zeilen)

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

EXPLAIN for the same with a restrictive clause for onum:

explain -- select * from ( -- select count(*) from ( -- select length(sel) from (
select
onum
,'column'||(onum::varchar) as vname
,vlen
,nlen
,olen
,NULLIF(vlen-olen,0) as delta_len
from (
select *
,('0'||split_part(split_part(nline,'(',2),')',1))::smallint as nlen
,('0'||split_part(split_part(oline,'(',2),')',1))::smallint as olen
from newcol_imp
join oldcol_imp on onum=nnum
join (
select
vnum
,split_part(vline,' ',1) as vname
,('0'||split_part(split_part(vline,'(',2),')',1))::smallint as vlen
from varchar_imp
) qv on nline like '%'||vname||'%'
where nline not like '%KEY%'
) qj
--limit 30
where vlen!=olen
and onum < 100
--and nlen > 0
;

-------------
EXPLAIN ANALYZE for the same with an irrelevant clause for onum
cpsdb=#
cpsdb=#
cpsdb=#
cpsdb=# explain analyze -- explain analyze verbose -- explain -- select * from ( -- select count(*) from ( -- select length(sel) from (
cpsdb-# select
cpsdb-# onum
cpsdb-# ,'column'||(onum::varchar) as vname
cpsdb-# ,vlen
cpsdb-# ,nlen
cpsdb-# ,olen
cpsdb-# ,NULLIF(vlen-olen,0) as delta_len
cpsdb-# from (
cpsdb(# select *
cpsdb(# ,('0'||split_part(split_part(nline,'(',2),')',1))::smallint as nlen
cpsdb(# ,('0'||split_part(split_part(oline,'(',2),')',1))::smallint as olen
cpsdb(# from newcol_imp
cpsdb(# join oldcol_imp on onum=nnum
cpsdb(# join (
cpsdb(# select
cpsdb(# vnum
cpsdb(# ,split_part(vline,' ',1) as vname
cpsdb(# ,('0'||split_part(split_part(vline,'(',2),')',1))::smallint as vlen
cpsdb(# from varchar_imp
cpsdb(# ) qv on nline like '%'||vname||'%'
cpsdb(# where nline not like '%KEY%'
cpsdb(# ) qj
cpsdb-# --limit 30
cpsdb-# where vlen!=olen
cpsdb-# and onum < 10000
cpsdb-# --and nlen > 0
cpsdb-# ;
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Merge Join (cost=0.56..30929.15 rows=6630 width=44) (actual time=3.881..257.727 rows=185 loops=1)
Merge Cond: (oldcol_imp.onum = newcol_imp.nnum)
Join Filter: ((('0'::text || split_part(split_part(varchar_imp.vline, '('::text, 2), ')'::text, 1)))::smallint <> (('0'::text || split_part(split_part(oldcol_imp.oline, '('::text, 2), ')'::text, 1)))::smallint)
Rows Removed by Join Filter: 425
-> Index Scan using oldcol_imp_pkey on oldcol_imp (cost=0.28..72.79 rows=2372 width=44) (actual time=0.008..0.230 rows=2371 loops=1)
Index Cond: (onum < 10000)
-> Materialize (cost=0.28..30086.66 rows=6663 width=77) (actual time=1.577..256.773 rows=610 loops=1)
-> Nested Loop (cost=0.28..30070.00 rows=6663 width=77) (actual time=1.576..256.675 rows=610 loops=1)
Join Filter: (newcol_imp.nline ~~ (('%'::text || split_part(varchar_imp.vline, ' '::text, 1)) || '%'::text))
Rows Removed by Join Filter: 1322510
-> Index Scan using newcol_imp_pkey on newcol_imp (cost=0.28..71.79 rows=2236 width=38) (actual time=0.011..0.422 rows=2220 loops=1)
Filter: (nline !~~ '%KEY%'::text)
Rows Removed by Filter: 152
-> Materialize (cost=0.00..14.94 rows=596 width=39) (actual time=0.000..0.015 rows=596 loops=2220)
-> Seq Scan on varchar_imp (cost=0.00..11.96 rows=596 width=39) (actual time=0.009..0.046 rows=596 loops=1)
Planning Time: 0.259 ms
Execution Time: 257.766 ms
(17 Zeilen)

And indeed it shows, that qual pushdown (Filter line 8 in the first error case) causes the calculation of nlen (with splitpart) to be executed for every tuple in newcol_imp.

But as written with the inner joins I intended to calculate nlen only after the joins. This is always OK because varchar_imp has only 596 valid records where the numeric or decimal cases are eliminated.

I think this qual pushdown is dangerous, because it is applied before the join and so includes invalid cases.

Hope this is a certain clarification

Hans Buschmann

________________________________
Von: Tom Lane <tgl@sss.pgh.pa.us>
Gesendet: Dienstag, 14. März 2023 19:25
An: David G. Johnston
Cc: Hans Buschmann; PostgreSQL Bug List
Betreff: Re: BUG #17842: Adding a qual to a working query gets bogus syntax error

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

On Tue, Mar 14, 2023, 11:12 Hans Buschmann <buschmann@nidsa.net> wrote:

2. Why is it reporting an syntax error (supposed at parse time) and not a
runtime error (supposed at execution time)

Your parsed query doesn't have a syntax error. Data it processes does,
which is found when parsing that data during query execution.

Right. It's kind of unfortunate perhaps that we use SQLSTATEs from
the "syntax error" class for malformed data input, but that seems
unlikely to change. In any case, it's pretty obvious that this error
is not complaining about anything in the text of the query.

regards, tom lane

On Tue, Mar 14, 2023, 12:08 Hans Buschmann <buschmann@nidsa.net> wrote:

I think this qual pushdown is dangerous, because it is applied before the
join and so includes invalid cases.

Well, it's better than any alternative that is more expensive even when the
data in question isn't prone to parsing problems. We optimize for the
common case of consistent normalized data. You have tools if you need to
operate with bad data.

David J.