Обсуждение: BUG #5174: [minor] directories symlinked into base/ are not recursively removed

The following bug has been logged online:

Bug reference:      5174
Logged by:          Richard Neill
Email address:      rn214@cam.ac.uk
PostgreSQL version: 8.4.1
Operating system:   Linux
Description:        [minor] directories symlinked into base/ are not
recursively removed
Details:

This is rather a minor nit, but it might be a useful report. If not, sorry
for wasting your time.

Summary:
If subdirectories of base/ are actually symlinks, then postgresql deletes
just the symlink, not the directory.


To reproduce:

I had a system which was running out of disk space, and required a CLUSTER
to recover some. However, there wasn't actually enough space to run the
cluster operation.  Therefore, I moved some of the larger subdirectories in
base/ to a different partition and symlinked them back.

For example:

service postgresql stop
cd /var/lib/postgresql/8.4/main/base/
mv 12345  /elsewhere
ln -s  /elsewhere/12345 .
#repeat for a few directories.
service postgresql start

This allowed me to get the system running again, and to recover space on the
main partition.

On shutting down postgres again to clean up, I found that all the symlinks
were gone (good), but that the directories on the /elsewhere partition were
still all present, and full of (now-useless) data.

The (possible) bug:
I'd expect Postgresql to first follow the link, then recursively clean-out
the linked directory; then delete the link, leaving (at most) an empty
directory /elsewhere/12345

Actually, postgresql deleted the symlink, then left everything in /elsewhere
unchanged.

Comparison with rm:

mkdir a
touch a/b
ln -s a c
rm -rf c
#c is deleted; a and b remain untouched,

i.e. it does exactly the same thing that Postgresql did.


Conclusion:
I don't know if this is intentional; if not, then a minor RFE would be to
fix it.

Thanks for your help - Richard

"Richard Neill" <rn214@cam.ac.uk> writes:
> If subdirectories of base/ are actually symlinks, then postgresql deletes
> just the symlink, not the directory.

Doing that is completely unsupported, and we cannot be expected to cope
with random manual modifications to the structure of the database.  Why
didn't you use a tablespace like you're supposed to?

            regards, tom lane

Tom Lane wrote:
> "Richard Neill" <rn214@cam.ac.uk> writes:
>> If subdirectories of base/ are actually symlinks, then postgresql deletes
>> just the symlink, not the directory.
>
> Doing that is completely unsupported, and we cannot be expected to cope
> with random manual modifications to the structure of the database.

That's a fair point, although as I understand it, it's reasonable to
move, say, pg_xlog in that manner.

Why
> didn't you use a tablespace like you're supposed to?

Because it wasn't supposed to be anything other than a temporary
workaround, and because I had only 5 minutes of downtime in which to
find more diskspace!

Anyway, I guess that's another "bug" report of mine that turns out to be
useless :-)

Best wishes,

Richard

>
>             regards, tom lane
>