Long reloption values are truncated by pg_dump and pg_get_indexdef()
Hello
pg_dump's appendReloptionsArray() and ruleutils' get_reloptions()
print an option value without quotes if it is a valid identifier that
doesn't need quoting. They don't check the length, so a string value
of 64 bytes or more is printed unquoted. When the output is replayed
the scanner truncates it to 63 bytes. The only sign of this is a
NOTICE during restore, the option silently ends up with a different
value.
None of the built-in reloptions can have such a value, but string
options defined by extensions can. Simple repro with the
dummy_index_am test module:
CREATE EXTENSION dummy_index_am;
CREATE TABLE t (i int4);
CREATE INDEX ti ON t USING dummy_index_am (i)
WITH (option_string_null =
'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa');
SELECT pg_get_indexdef('ti'::regclass);
-- ... WITH (option_string_null=aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa)
pg_dump prints the same, and restoring it results in a 63 character value.
Currently this only affects index AMs with string options, but there
are 2 proposals that make it more likely to happen: tablespace options
defined by extensions [1]/messages/by-id/CAN4CZFNxu5BgBge12C=f3JV9LdZvUiG+pQHOCutATewSywzFuw@mail.gmail.com (where I ran into it), and table AM
reloptions [2]/messages/by-id/ea1c4d33-0780-473c-96dc-1468cf733a04@dunslane.net.
The attached patch quotes values that are too long to be identifiers
in both places, and adds a test to dummy_index_am.
[1]: /messages/by-id/CAN4CZFNxu5BgBge12C=f3JV9LdZvUiG+pQHOCutATewSywzFuw@mail.gmail.com
[2]: /messages/by-id/ea1c4d33-0780-473c-96dc-1468cf733a04@dunslane.net
On Nov 19, 2026, at 04:50, Zsolt Parragi <zsolt.parragi@percona.com> wrote:
Hello
pg_dump's appendReloptionsArray() and ruleutils' get_reloptions()
print an option value without quotes if it is a valid identifier that
doesn't need quoting. They don't check the length, so a string value
of 64 bytes or more is printed unquoted. When the output is replayed
the scanner truncates it to 63 bytes. The only sign of this is a
NOTICE during restore, the option silently ends up with a different
value.None of the built-in reloptions can have such a value, but string
options defined by extensions can. Simple repro with the
dummy_index_am test module:CREATE EXTENSION dummy_index_am;
CREATE TABLE t (i int4);
CREATE INDEX ti ON t USING dummy_index_am (i)
WITH (option_string_null =
'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa');
SELECT pg_get_indexdef('ti'::regclass);
-- ... WITH (option_string_null=aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa)pg_dump prints the same, and restoring it results in a 63 character value.
Currently this only affects index AMs with string options, but there
are 2 proposals that make it more likely to happen: tablespace options
defined by extensions [1] (where I ran into it), and table AM
reloptions [2].The attached patch quotes values that are too long to be identifiers
in both places, and adds a test to dummy_index_am.[1]: /messages/by-id/CAN4CZFNxu5BgBge12C=f3JV9LdZvUiG+pQHOCutATewSywzFuw@mail.gmail.com
[2]: /messages/by-id/ea1c4d33-0780-473c-96dc-1468cf733a04@dunslane.net
<v1-0001-Quote-reloption-values-too-long-to-be-identifiers.patch>
I can reproduce the issue with your repro example, and the fix looks good to me.
My only comment is about the new comment:
```
and is short enough not to be truncated as one.
```
From this comment, I cannot quickly understand why the length check is needed. I would propose something like:
```
and is shorter than NAMEDATALEN bytes. Longer values must be quoted as string literals to avoid identifier truncation when the output is parsed.
```
As a non-native English speaker, I always prefer simpler and more explicit code comments. If other folks consider the wording in the patch good enough, I’m fine with withdrawing my comment.
Best regards,
--
Chao Li (Evan)
HighGo Software Co., Ltd.
https://www.highgo.com/