Alvaro Herrera writes:
> Why not just use an unsigned 64 bit variable? Also, perhaps
> palloc_huge() avoids the whole problem in the first place ...
Ja, that crossed my mind too, but the current limit is already far
beyond anything that is usually configured for per-backend memory use,
so I dismissed it.
> though it might only move the issue around, if you cannot ship the
> longer-than-1GB resulting escaped value.
For example, when client and server encodings do not match:
,----[ mbutils.c ]
| result = palloc(len * MAX_CONVERSION_GROWTH + 1);
`----
This results in the fun fact that the maximum size for bytea values that
are guaranteed to be pg_dumpable regardless of encoding/escaping
settings is lower than 64MB.
One thing that would also mitigate the problem is supporting a more
efficient output format. For example, there's already means for
base64-encoding in the backend:
self=# select c, length(encode(mkbytea(28),c)) from (values ('hex'),('base64')) as v(c); c | length
--------+-----------hex | 536870912base64 | 362623337
(2 rows)
Maybe it is reasonable to make it available as another option for use in
bytea_output?
regards,
Andreas