Em 22 de jun de 2020, à(s) 13:54, Tom Lane <tgl@sss.pgh.pa.us> escreveu:
Flaris Feller <flaris.feller@gmail.com> writes:
Em seg., 22 de jun. de 2020 às 12:33, Peter J. Holzer <hjp-pgsql@hjp.at>
escreveu:
On 2020-06-22 11:13:33 -0300, Flaris Feller wrote:
When using Postgres 9.6.15 on "CentOS Linux release 7.5.1804 (Core)" on
Intel x86_64 I noticed "invalid memory alloc request size" error at PostgreSQL
logs.
That's 2**64 - 3. So probably the size of some object is (erroneously)
computed as -3 bytes.
Yeah. I don't recall the bit-level details offhand, but this is an extremely common symptom of a corrupted length word in a variable-length field.
The usual approach is to try to isolate which row or rows contains corrupt data and then delete it. I'm not aware of any tools for doing that automatically, but you can usually home in on a bad row by a process of binary search, eg testing how many rows you can fetch without seeing the error. Also "select * from mytab where ctid = '(m,n)'" is useful for probing individual rows, once you get close to the problem area.
