Re: Causal reads take II

On Sun, Oct 1, 2017 at 9:05 AM, Dmitry Dolgov <9erthalion6(at)gmail(dot)com> wrote:
>>> LOG: could not remove directory "pg_tblspc/47733/PG_10_201707211/47732":
>>> Directory not empty
>>> ...
>>
>> Hmm. The first error ("could not remove directory") could perhaps be
>> explained by temporary files from concurrent backends.
>> ...
>> Perhaps in your testing you accidentally copied a pgdata directory over
>> the
>> top of it while it was running? In any case I'm struggling to see how
>> anything in this patch would affect anything at the REDO level.
>
> Hmm...no, I don't think so. Basically what I was doing is just running
> `installcheck` against a primary instance (I assume there is nothing wrong
> with
> this approach, am I right?). This particular error was caused by
> `tablespace`
> test which was failed in this case:
>
> ```
> INSERT INTO testschema.foo VALUES(1);
> ERROR: could not open file "pg_tblspc/16388/PG_11_201709191/16386/16390":
> No such file or directory
> ```
>
> I tried few more times, and I've got it two times from four attempts on a
> fresh
> installation (when all instances were on the same machine). But anyway I'll
> try
> to investigate, maybe it has something to do with my environment.
>
> ...
>
> 2017-09-30 15:47:55.154 CEST [6030] LOG: invalid magic number 0000 in log
> segment 000000010000000000000020, offset 10092544

Hi Dmitry,

Thanks for testing. Yeah, it looks like the patch may be corrupting
the WAL stream in some case that I didn't hit in my own testing
procedure. I will try to reproduce these failures.

--
Thomas Munro
http://www.enterprisedb.com