On Sun, May 10, 2015 at 7:50 AM, Albe Laurenz <laurenz.albe@wien.gv.at> wrote:
> Scott Marlowe wrote:
>> On Sat, May 9, 2015 at 11:20 PM, Albe Laurenz <laurenz.albe@wien.gv.at> wrote:
>>> Maxim Boguk wrote:
>>>> It's depend where a corruption happen, if pages become corrupted due to some
>>>> problems with physical storage (filesystem) in that case a replica data should be ok.
>
>>> I would not count on that.
>>> I have had a case where a table file got corrupted due to hardware problems.
>>> Pages that contained data were suddenly zeroed.
>>> PostgreSQL recognizes such a block as empty, so the user got no error, but
>>> data were suddenly missing. What does a user do in such a case? He/she grumbles
>>> and enters the data again. This insert will be replicated to the standby (which was
>>> fine up to then) and will cause data corruption there (duplicate primary keys).
>
>> You had zero corrupted pages turned on. PostgreSQL by default does NOT
>> DO THIS. That setting is for recovering a corrupted database not for
>> everyday use!
>
> No, I didn't.
>
> It was not PostgreSQL that zeroed the page, but the hardware or operating system.
> The problem was a flaky fibre channel cable that intermittently was connected and disconnected.
> That corrupted the file system, and I guess it must have been file system recovery
> that zeroed the pages. I'm not 100% certain, at any rate the symptoms were silently missing data.
Ahh OK. So broken hardware. I've seen some RAID controlelrs do that.
Sorry but your post didn't make it clear where the zeroing came from.
