FDW pushdown of non-collated functions

Dear Hackers,

I figured out this email was sent at release time. The worst time to ask
for thoughts on a subject IMHO. Anyway, I hope this email will pop the
topic over the stack!
Thank you!

Le ven. 8 sept. 2023 à 16:41, Jean-Christophe Arnu <jcarnu@gmail.com> a
écrit :

Dear hackers,

I recently found a weird behaviour involving FDW (postgres_fdw) and
planning.

Here’s a simplified use-case:

Given a remote table (say on server2) with the following definition:

CREATE TABLE t1(
ts timestamp without time zone,
x bigint,
x2 text
);
--Then populate t1 table:INSERT INTO t1
SELECT
current_timestamp - 1000*random()*'1 day'::interval
,x
,''||x
FROM
generate_series(1,100000) as x;

This table is imported in a specific schema on server1 (we do not use
use_remote_estimate) also with t1 name in a specific schema:

On server1:

CREATE SERVER server2
FOREIGN DATA WRAPPER postgres_fdw
OPTIONS (
host '127.0.0.1',
port '9002',
dbname 'postgres',
use_remote_estimate 'false'
);
CREATE USER MAPPING FOR jc
SERVER server2
OPTIONS (user 'jc');
CREATE SCHEMA remote;

IMPORT FOREIGN SCHEMA public
FROM SERVER server2
INTO remote ;

On a classic PostgreSQL 15 version the following query using date_trunc()
is executed and results in the following plan:

jc=# explain (verbose,analyze) select date_trunc('day',ts), count(1) from remote.t1 group by date_trunc('day',ts) order by 1;
QUERY PLAN -----------------------------------------------------------------------------------------------------------------------------------
Sort (cost=216.14..216.64 rows=200 width=16) (actual time=116.699..116.727 rows=1001 loops=1)
Output: (date_trunc('day'::text, ts)), (count(1))
Sort Key: (date_trunc('day'::text, t1.ts))
Sort Method: quicksort Memory: 79kB
-> HashAggregate (cost=206.00..208.50 rows=200 width=16) (actual time=116.452..116.532 rows=1001 loops=1)
Output: (date_trunc('day'::text, ts)), count(1)
Group Key: date_trunc('day'::text, t1.ts)
Batches: 1 Memory Usage: 209kB
-> Foreign Scan on remote.t1 (cost=100.00..193.20 rows=2560 width=8) (actual time=0.384..106.225 rows=100000 loops=1)
Output: date_trunc('day'::text, ts)
Remote SQL: SELECT ts FROM public.t1
Planning Time: 0.077 ms
Execution Time: 117.028 ms

Whereas the same query with date_bin()

jc=# explain (verbose,analyze) select date_bin('1day',ts,'2023-01-01'), count(1) from remote.t1 group by 1 order by 1;
QUERY PLAN ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Foreign Scan (cost=113.44..164.17 rows=200 width=16) (actual time=11.297..16.312 rows=1001 loops=1)
Output: (date_bin('1 day'::interval, ts, '2023-01-01 00:00:00'::timestamp without time zone)), (count(1))
Relations: Aggregate on (remote.t1)
Remote SQL: SELECT date_bin('1 day'::interval, ts, '2023-01-01 00:00:00'::timestamp without time zone), count(1) FROM public.t1 GROUP BY 1 ORDER BY date_bin('1 day'::interval, ts, '2023-01-01 00:00:00'::timestamp without time zone) ASC NULLS LAST
Planning Time: 0.114 ms
Execution Time: 16.599 ms

With date_bin() the whole expression is pushed down to the remote server,
whereas with date_trunc() it’s not.

I dived into the code and live debugged. It turns out that decisions to
pushdown or not a whole query depends on many factors like volatility and
collation. In the date_trunc() case, the problem is all about collation (
date_trunc() on timestamp without time zone). And decision is made in the
foreign_expr_walker() in deparse.c (
https://git.postgresql.org/gitweb/?p=postgresql.git;a=blob;f=contrib/postgres_fdw/deparse.c;h=efaf387890e3f85c419748ec3af5d1e9696c9c4c;hb=86648dcdaec67b83cec20a9d25b45ec089a7c624#l468
)

First the function is tested as shippable (able to be pushed down) and
date_trunc() and date_bin() both are.

Then parameters sub-expressions are evaluated with collation and
“shippability”, and they all are with both functions.

Then we arrive at this code portion:

if (fe->inputcollid == InvalidOid)
/* OK, inputs are all noncollatable */ ;else if (inner_cxt.state != FDW_COLLATE_SAFE ||
fe->inputcollid != inner_cxt.collation)
return false;

For date_trunc() function :

-

fe variable contains the sub-expressions/arguments merged constraints
such as fe->inputcollid. This field is evaluated to 100 (default
collation) so codes jumps to else statement and evaluates the if
predicates. This 100 inputcollationid is due to text predicate 'day'.
-

inner_cxt.state contains FDW_COLLATE_STATE but inner_cxt.collation
contains 0 (InvalidOid) so the control flow returns false thus the
function cannot be pushed down.

For date_bin() function :

- fe variable contains the sub-expressions/arguments merged
constraints. Here, fe->inputcollid is evaluated to 0 (InvalidOid) thus
skips the else statement and continues the control flow in the
function.

For date_bin(), all arguments are “non-collatable” arguments (timestamp
without time zone and interval).

So the situation is that date_trunc() is a “non-collatable” function
failing to be pushed down whereas it may be a good idea to do so.

Maybe we could add another condition to the first if statement in order to
allow a “no-collation” function to be pushed down even if they have
“collatable” parameters. I’m not sure about the possible regressions of
behaviour of this change, but it seems to work fine with date_trunc() and
date_part() (which suffers the same problem).

Here’s the following change

/*
* If function's input collation is not derived from a foreign
* Var, it can't be sent to remote.
*/if (fe->inputcollid == InvalidOid ||
fe->funccollid == InvalidOid)
/* OK, inputs are all noncollatable */ ;else if (inner_cxt.state != FDW_COLLATE_SAFE ||
fe->inputcollid != inner_cxt.collation)
return false;

I don’t presume this patch is free from side effects or fits all use-cases.

A patch (tiny) is attached to this email. This patch works against
master/head at the time of writing.
Thank you for any thoughts.

--
Jean-Christophe Arnu

--
Jean-Christophe Arnu

Hi Jean-Christophe,

On Fri, Sep 8, 2023 at 11:30 PM Jean-Christophe Arnu <jcarnu@gmail.com> wrote:

Maybe we could add another condition to the first if statement in order to allow a “no-collation” function to be pushed down even if they have “collatable” parameters. I’m not sure about the possible regressions of behaviour of this change, but it

seems to work fine with date_trunc() and date_part() (which suffers
the same problem).

That may not work since the output of the function may be dependent
upon the collation on the inputs.

There were similar discussions earlier. E.g.
/messages/by-id/CACowWR1ARWyRepRxGfijMcsw+H84Dj8x2o9N3kvz=z1p+6b45Q@mail.gmail.com.

Reading Tom's first reply there you may work around this by declaring
the collation explicitly.

Briefly reading Tom's reply, the problem seems to be trusting whether
the default collation locally and on the foreign server respectively
is same or not. May be a simple fix is to declare a foreign server
level option declaring that the default collation on the foreign
server is same as the local server may be a way to move forward. But
given that the problem remains unsolved for 7 years at least, may be
such a simple fix is not enough.

Another solution would be to attach another attribute to a function
indicating whether the output of that function depends upon the input
collations or not. Doing that just for FDW may not be acceptable
though.

--
Best Wishes,
Ashutosh Bapat

Hi Ashutosh,

Le ven. 6 oct. 2023 à 14:16, Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
a écrit :

Hi Jean-Christophe,

On Fri, Sep 8, 2023 at 11:30 PM Jean-Christophe Arnu <jcarnu@gmail.com>
wrote:

Maybe we could add another condition to the first if statement in order

to allow a “no-collation” function to be pushed down even if they have
“collatable” parameters. I’m not sure about the possible regressions of
behaviour of this change, but it
seems to work fine with date_trunc() and date_part() (which suffers
the same problem).

That may not work since the output of the function may be dependent
upon the collation on the inputs.

There were similar discussions earlier. E.g.

/messages/by-id/CACowWR1ARWyRepRxGfijMcsw+H84Dj8x2o9N3kvz=z1p+6b45Q@mail.gmail.com
.

Reading Tom's first reply there you may work around this by declaring
the collation explicitly.

Thanks for your reply. I did not catch these messages in the archive.
Thanks for spotting them.

Briefly reading Tom's reply, the problem seems to be trusting whether
the default collation locally and on the foreign server respectively
is same or not. May be a simple fix is to declare a foreign server
level option declaring that the default collation on the foreign
server is same as the local server may be a way to move forward. But
given that the problem remains unsolved for 7 years at least, may be
such a simple fix is not enough.

I studied postgres_fdw source code a bit and the problem is not as easy to
solve : one could set an option telling the default remote collation is
aligned with local per "server" but nothing guaranties that the parameter
collation is known on the «remote» side.

Another solution would be to attach another attribute to a function
indicating whether the output of that function depends upon the input
collations or not. Doing that just for FDW may not be acceptable
though.

Yes, definitely. I thought

Anyway, you're right, after 7 years, this is a really difficult problem to
solve and there's no straightforward solution (to my eyes).
Thanks again for your kind explanations
Regards

--
Jean-Christophe Arnu