Hi;
I ran into a funny situation today regarding PostgreSQL replication and wal corruption and wanted to go over what I think happened and what I wonder about as a possible solution.
Basic information is custom-build PostgreSQL 9.6.3 on Gentoo, on a ~5TB database with variable load. Master database has two slaves and generates 10-20MB of WAL traffic a second. The data_checksum option is off.
The problem occurred when I attempted to restart the service on the slave using pg_ctl (I believe the service had been started with sys V init scripts). On trying to restart, it gave me a nice "Invalid memory allocation request" error and promptly stopped.
The main logs showed a lot of messages like before the restart:
2017-08-02 11:47:33 UTC LOG: PID 19033 in cancel request did not match any process
2017-08-02 11:47:33 UTC LOG: PID 19032 in cancel request did not match any process
2017-08-02 11:47:33 UTC LOG: PID 19024 in cancel request did not match any process
2017-08-02 11:47:33 UTC LOG: PID 19034 in cancel request did not match any process
On restart, the following was logged to stderr:
LOG: entering standby mode
LOG: redo starts at 1E39C/8B77B458
LOG: consistent recovery state reached at 1E39C/E1117FF8
FATAL: invalid memory alloc request size 3456458752
LOG: startup process (PID 18167) exited with exit code 1
LOG: terminating any other active server processes
LOG: database system is shut down
After some troubleshooting I found that the wal segment had become corrupt, I copied the correct one from the master and everything came up to present.
So It seems like somewhere something crashed big time on the back-end and when we tried to restart, the wal ended in an invalid way.
