On Thu, May 14, 2015 at 10:29 PM, Andrew Dunstan <
andrew@dunslane.net> wrote:
>
>
> On 05/14/2015 10:52 AM, Robert Haas wrote:
>>
>> On Thu, May 14, 2015 at 12:12 AM, Amit Kapila <
amit.kapila16@gmail.com> wrote:
>>>
>>> On Thu, May 14, 2015 at 2:10 AM, Andrew Dunstan <
andrew@dunslane.net> wrote:
>>>>
>>>> How about if we simply abort if we find a non-symlink where we want the
>>>> symlink to be, and only remove something that is actually a symlink (or a
>>>> junction point, which is more or less the same thing)?
>>>
>>> We can do that way and for that I think we need to use rmdir
>>> instead of rmtree in the code being discussed (recovery path),
>>> OTOH we should try to minimize the errors raised during
>>> recovery.
>>
>> I'm not sure I understand this issue in detail, but why would using
>> rmtree() on something you expect to be a symlink ever be a good idea?
>> It seems like if things are the way you expect them to be, it has no
>> benefit, but if they are different from what you expect, you might
>> blow away a ton of important data.
>>
>> Maybe I am just confused.
>>
>
>
> The suggestion is to get rid of using rmtree. Instead, if we find a non-symlink in pg_tblspc we'll make the user clean it up before we can continue. So your instinct is in tune with my suggestion.
>
Find the patch which gets rid of rmtree usage. I have made it as
a separate function because the same code is used from
create_tablespace_directories() as well. I thought of extending the
same API for using it from destroy_tablespace_directories() as well,
but due to special handling (especially for ENOENT) in that function,
I left it as of now.
