Content-Type: text/plain;
	charset=us-ascii
Mime-Version: 1.0 (Mac OS X Mail 16.0 \(3864.400.21\))
Subject: Re: COPY JSON: use trailing commas in FORCE_ARRAY output
From: Chao Li <li.evan.chao@gmail.com>
In-Reply-To: <B774FE54-19BB-4250-AF2E-967460816C1F@yesql.se>
Date: Wed, 6 May 2026 17:18:38 +0800
Cc: PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Content-Transfer-Encoding: quoted-printable
Message-Id: <F08C3C8E-9769-4A96-B392-DA93E11AC3A5@gmail.com>
References: <DFAC4097-2559-4DED-B7D5-EB53B02E9DA3@gmail.com>
 <B774FE54-19BB-4250-AF2E-967460816C1F@yesql.se>
To: Daniel Gustafsson <daniel@yesql.se>
Archived-At: 
 <https://www.postgresql.org/message-id/F08C3C8E-9769-4A96-B392-DA93E11AC3A5%40gmail.com>
Precedence: bulk


> On May 6, 2026, at 16:23, Daniel Gustafsson <daniel@yesql.se> wrote:
>=20
>> On 6 May 2026, at 08:40, Chao Li <li.evan.chao@gmail.com> wrote:
>=20
>> I was surprised by the comma placement. It is valid JSON, but it =
looks quite uncommon.
>=20
> It might look uncommon, but for very wide lines it's IMHO preferrable =
to not
> have to scroll all the way to the end of the line to know that the =
line is part
> of an array.
>=20
>> For comparison, the existing json_agg() places commas at the end of =
the line:
>=20
> That's true, but json_agg() and COPY TO in ndjson format have =
different use
> cases.
>=20
>> ..it should not have any performance impact.
>=20
> It does add branches though, and in one branch use a non-inlined =
function where
> previously it would unconditionally use an inline function.  ISTM it =
would
> still be valuable to do performance testing given that COPY is =
commonly used in
> performance sensitive settings.
>=20

Make sense. I just did a test to compare the performance between master =
and the patch:

For the data setup, since the patch only changes where the comma is =
emitted, I intentionally used a table with only one column, to minimize =
the cost of formatting each row:
```
DROP TABLE IF EXISTS copy_json_force_array_perf;

CREATE UNLOGGED TABLE copy_json_force_array_perf(id int);

INSERT INTO copy_json_force_array_perf
SELECT g
FROM generate_series(1, 10000000) AS g;

VACUUM ANALYZE copy_json_force_array_perf;

\timing on
```

On master:
```
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1208.694 ms (00:01.209)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1200.203 ms (00:01.200)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1238.639 ms (00:01.239)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1211.344 ms (00:01.211)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1252.197 ms (00:01.252)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1223.510 ms (00:01.224)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1212.378 ms (00:01.212)
```
Average: ~1221 ms

With the patch:
```
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1218.580 ms (00:01.219)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1212.913 ms (00:01.213)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1204.350 ms (00:01.204)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1205.276 ms (00:01.205)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1202.088 ms (00:01.202)
evantest=3D# COPY copy_json_force_array_perf TO '/dev/null' WITH (FORMAT =
json, FORCE_ARRAY);
COPY 10000000
Time: 1222.390 ms (00:01.222)
```
Average: ~1211 ms

The difference doesn't look quite meaningful from this test. I built =
with debug and asserts disabled, and compiled with -O2. The output was =
written to /dev/null to avoid client/network overhead. The tests ran on =
my MacBook M4.

Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/