Re: [DOCS] Let's document a bytea bug

On Fri, Jul 31, 2020 at 10:13:48AM +0500, Andrey M. Borodin wrote:
> Hi Anna!
>
> > 23 мая 2018 г., в 20:33, Anna Akenteva <a(dot)akenteva(at)postgrespro(dot)ru> написал(а):
> >
> >
> > Some time ago I've encountered a problem with the bytea type: we can't SELECT
> > bytea strings whose textual representation is too big to fit into StringInfoData.
> > And as a side effect, pg_dump refuses to dump tables with big bytea strings.
> >
> > It's a bug, it's pretty confusing, but it seems like there's no pretty way
> > to fix it so far. Here's a link to a recent discussion on the issue:
> > https://www.postgresql.org/message-id/flat/c8bdf802d41ec37003ec3b726db79428(at)postgrespro(dot)ru#c8bdf802d41ec37003ec3b726db79428@postgrespro.ru
> >
> > Since it won't be fixed anytime soon, I thought it could be worth documenting.
> > Attaching a patch for the documentation: I added some text to the "Binary Data Types"
> > part where I tried to describe the issue and to explain how to deal with it.
> >
> > My patch in plain text (for convenience):
> >
> > It is not recommended to use bytea strings whose textual representation
> > exceeds 1GB, as it may not be possible to SELECT them due to output size
> > limitations. Consequently, a table containing such big strings cannot be
> > properly processed by pg_dump, as pg_dump will try to SELECT these values from the
> > table and fail. The exact size limit advised for bytea strings depends on their
> > content, the external format and encoding that you are using, the context in
> > which they will be selected. The general rule is that when you use SELECT,
> > the returned tuple should not exceed 1GB. Although even if SELECT does not
> > work, you can still retrieve big bytea strings using COPY in binary format.
>
> Thanks for this message. It took me a while to find out what was the problem.
> +1 for documenting this, maybe even with exact error like
> [ 2020-07-30 01:20:32.248 MSK pg_dump - 10.3.3.30,XX000 ]:ERROR: invalid memory alloc request size 1472599557
> It's really really scary. My first feeling was that it's TOAST corruption.

I still have Anna Akenteva's 2018 email in my mailbox because I wanted
to research this further. Now that you have re-reported the problem, I
am on it! ;-)

I looked for a clearer reproducible test case, and I have found this
one:

$ awk 'BEGIN { printf "\\\\x"; for (i=0; i < 30000000; i++) \
printf "7f7f7f7f7f7f7f7f7f7f"; print}' > /tmp/x

$ psql test

CREATE TABLE test(x BYTEA);

COPY test FROM '/tmp/x';

COPY test FROM '/tmp/x';

CREATE TABLE big_data AS SELECT (string_agg(x,'')) AS x FROM test;

SELECT length(x) FROM big_data;
length
-----------
600000000

SELECT octet_length(x) FROM big_data;
octet_length
--------------
600000000

SELECT x FROM big_data;
ERROR: invalid memory alloc request size 1200000003

\q

$ pg_dump -d test > /rtmp/a
pg_dump: Dumping the contents of table "big_data" failed: PQgetResult() failed.
pg_dump: Error message from server: ERROR: invalid memory alloc request size 1200000003
pg_dump: The command was: COPY public.big_data (x) TO stdout;

So, the character and octet length is 600 million, but on output, that
will be expanded, and both SELECT and pg_dump fail. I also can't see
how to improve the error message since it happens so low in the stack.

In reading the previous posts, I understand that expanding the length
limit would be very hard, and it is difficult to see how to know if the
output will error at input time. I am in favor of the 2018 proposed
patch, attached, perhaps with some small adjustments.

--
Bruce Momjian <bruce(at)momjian(dot)us> https://momjian.us
EnterpriseDB https://enterprisedb.com

The usefulness of a cup is in its emptiness, Bruce Lee