On Wed, Jun 24, 2015 at 3:43 PM, Michael Paquier wrote:
>> I just realized another problem: We recently learned the hard way that some
>> people have files in the data directory that are not writeable by the
>> 'postgres' user
>> (http://www.postgresql.org/message-id/20150523172627.GA24277@msg.df7cb.de).
>> pg_rewind will try to overwrite all files it doesn't recognize as relation
>> files, so it's going to fail on those. A straightforward fix would be to
>> first open the destination file in read-only mode, and compare its contents,
>> and only open the file in write mode if it has changed. It would still fail
>> when the files really differ, but I think that's acceptable.
>
> If I am missing nothing, two code paths need to be patched here:
> copy_file_range and receiveFileChunks. copy_file_range is
> straight-forward. Now wouldn't it be better to write the contents into
> a temporary file, compare their content, and then switch if necessary
> for receiveFileChunks?
After sleeping on it, I have been looking at this issue again and came
up with the patch attached. Instead of checking if the content of the
target and the source file are the same, meaning that we would still
need to fetch chunk content from the server in stream mode, I think
that it is more robust to check if the target file can be opened and
check for EACCES on failure, bypassing it if process does not have
permissions on it. the patch contains a test case as well, and is
independent on the rest sent upthread.
Thoughts?
--
Michael