Обсуждение: pg_upgrade + Extensions

Hackers,

My co-workers tell me that pg_upgrade told them to drop the colnames and hostname extensions before upgrading from 9.3
to9.4. Fortunately, Postgres had  not recorded any dependencies on functions from these extensions (not sure why not,
sincewe do user them, but for the moment grateful), so it wasn’t a big deal to drop them and then add them back after
finishingthe upgrade. But frankly I don’t understand why this was necessary. It’s true that they’re C extensions with
sharedlibraries, but there are separate .so files for the 9.3 and 9.4 installs. 

Would there be a way to convince pg_upgrade that extensions don’t need to be dropped before upgrading?

Thanks,

David

"David E. Wheeler" <david@justatheory.com> writes:
> My co-workers tell me that pg_upgrade told them to drop the colnames and
> hostname extensions before upgrading from 9.3 to 9.4.

Really?  I see nothing in the source code that would print any such
advice.

There *is* a check on whether .so libraries used by the source
installation exist in the destination one.  But the preferred way to
deal with that type of complaint is to install the needed libraries
(in the destination's lib/ folder).  You shouldn't have to drop anything
as long as you have a copy of the extension that works for the new PG
version.
        regards, tom lane

On Jul 10, 2015, at 11:32 AM, Smitha Pamujula <smitha.pamujula@iovation.com> wrote:

> I just tested and yes that worked. Once we have the new library for the hostname, pg_upgrade is not complaining about
thehostname extension.  

Great, thank you Smitha -- and Tom for the pointer.

> Your installation references loadable libraries that are missing from the
> new installation.  You can add these libraries to the new installation,
> or remove the functions using them from the old installation.  A list of
> problem libraries is in the file:
>     loadable_libraries.txt
>
> Failure, exiting
> [postgres@pdxdvrptsrd04 ~]$ cat loadable_libraries.txt
> Could not load library "json_build"
> ERROR:  could not access file "json_build": No such file or directory

So you drop the json_build extension before upgrading, but pg_upgrade still complains that it’s missing? That seems
odd.

Best,

David

On 07/13/2015 01:12 PM, Smitha Pamujula wrote:
> Yes. I have checked that the extension didn't exist in any of the 
> databases. I used \dx to see if there was json_build was listed and i 
> didnt see any. Is that sufficient to check its existence. I am about 
> to do another testing in a few minutes on a different machine. I will 
> capture before/after shots
>
>

Please don't top-post on the PostgreSQL lists - see 
<http://idallen.com/topposting.html>

In theory it should be enough if it was installed in the standard way. 
But a more thorough procedure would be to run this command:
   select count(*)   from pg_proc   where prosrc ~ 'json_build';

Here's a one-liner that will check every database for you:
   for db in `psql  -t -c 'select datname from pg_database where   datallowconn'` ; do C=`psql -t -c "select count(*)
frompg_proc   where prosrc ~ 'json_build'" $db`; echo $db $C; done
 

cheers

andrew

On Tue, Jul 14, 2015 at 09:48:59AM -0700, Smitha Pamujula wrote:
> This error will go away only if I install the new json_build94.
> ........................................................................

> I was under the impression that we dont need to get the json_build
> libraries for 94. But the upgrade wont go forward without that. Are we
> missing doing something here or is it necessarty to get json_build94
> before upgrade.

I am coming very late to this discussion, but I do have some
information.  Basically, pg_upgrade is trying to predict if the
dump/restore will fail by checking references to shared objects in the
old cluster.  It sees a reference to json_build in the old cluster and
assumes this is needed in the new cluster, when in fact it isn't.

We could have hard-coded this knowledge into 9.4 pg_upgrade if we had
known it.  The simple solution is to install json_build94 as you did,
run pg_upgrade, then just uninstall json_build94 as nothing depends on
it.  Not sure if this should be in the pg_upgrade docs or not.

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + Everyone has their own god. +

On Mon, Aug 31, 2015 at 04:03:20PM -0700, Smitha Pamujula wrote:
> Thank you Bruce. So far installing it before have been working well so we will
> continue with that plan. 
> 
> I think it would help if its noted somewhere in the document as it would have
> helped us save some time understanding why it was failing and why it was
> looking for json_build. 

The problem is that this is a rare case where you had an extension that
was later included in Postgres.

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + Everyone has their own god. +

On Aug 31, 2015, at 4:20 PM, Bruce Momjian <bruce@momjian.us> wrote:

>> I think it would help if its noted somewhere in the document as it would have
>> helped us save some time understanding why it was failing and why it was
>> looking for json_build.
>
> The problem is that this is a rare case where you had an extension that
> was later included in Postgres.

Maybe not so rare. Thanks to Andrew, we’ve had to do this for both 9.2-9.3 (json_object) and 9.3-9.4 (json_build).

Best,

David

On 08/31/2015 07:21 PM, David E. Wheeler wrote:
> On Aug 31, 2015, at 4:20 PM, Bruce Momjian <bruce@momjian.us> wrote:
>
>>> I think it would help if its noted somewhere in the document as it would have
>>> helped us save some time understanding why it was failing and why it was
>>> looking for json_build.
>> The problem is that this is a rare case where you had an extension that
>> was later included in Postgres.
> Maybe not so rare. Thanks to Andrew, we’ve had to do this for both 9.2-9.3 (json_object) and 9.3-9.4 (json_build).
>


Yeah, a lot of people don't like to wait for new stuff. :-)

cheers

andrew

On Mon, Aug 31, 2015 at 07:28:00PM -0400, Andrew Dunstan wrote:
> 
> 
> On 08/31/2015 07:21 PM, David E. Wheeler wrote:
> >On Aug 31, 2015, at 4:20 PM, Bruce Momjian <bruce@momjian.us> wrote:
> >
> >>>I think it would help if its noted somewhere in the document as it would have
> >>>helped us save some time understanding why it was failing and why it was
> >>>looking for json_build.
> >>The problem is that this is a rare case where you had an extension that
> >>was later included in Postgres.
> >Maybe not so rare. Thanks to Andrew, we’ve had to do this for both 9.2-9.3 (json_object) and 9.3-9.4 (json_build).
> >
> 
> 
> Yeah, a lot of people don't like to wait for new stuff. :-)

It might make the most sense to mention this method in the release notes
of the extension.  However, I assume they are not using the extension in
the new server so their is no release to look at.

Still, I don't know how many people are doing this, but the right fix is
to get the names of the modules that are superceeded and tell pg_upgrade
to skip them.

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com
 + Everyone has their own god. +

On 08/31/2015 07:32 PM, Bruce Momjian wrote:
> On Mon, Aug 31, 2015 at 07:28:00PM -0400, Andrew Dunstan wrote:
>>
>> On 08/31/2015 07:21 PM, David E. Wheeler wrote:
>>> On Aug 31, 2015, at 4:20 PM, Bruce Momjian <bruce@momjian.us> wrote:
>>>
>>>>> I think it would help if its noted somewhere in the document as it would have
>>>>> helped us save some time understanding why it was failing and why it was
>>>>> looking for json_build.
>>>> The problem is that this is a rare case where you had an extension that
>>>> was later included in Postgres.
>>> Maybe not so rare. Thanks to Andrew, we’ve had to do this for both 9.2-9.3 (json_object) and 9.3-9.4 (json_build).
>>>
>>
>> Yeah, a lot of people don't like to wait for new stuff. :-)
> It might make the most sense to mention this method in the release notes
> of the extension.  However, I assume they are not using the extension in
> the new server so their is no release to look at.
>
> Still, I don't know how many people are doing this, but the right fix is
> to get the names of the modules that are superceeded and tell pg_upgrade
> to skip them.
>


I don't think this knowledge should be hardcoded in pg_upgrade. I could
see some point in a switch that would tell pg_upgrade a list of
extensions to ignore.

cheers

andrew

Andrew Dunstan <andrew@dunslane.net> writes:
> On 08/31/2015 07:32 PM, Bruce Momjian wrote:
>> Still, I don't know how many people are doing this, but the right fix is
>> to get the names of the modules that are superceeded and tell pg_upgrade
>> to skip them.

> I don't think this knowledge should be hardcoded in pg_upgrade. I could 
> see some point in a switch that would tell pg_upgrade a list of 
> extensions to ignore.

That would not be terribly helpful for cases where the pg_upgrade call is
embedded in some wrapper script or other.

In any case, there is plenty of precedent for hard-coding knowledge about
specific version updates into pg_upgrade.  The question here is whether
it's feasible to handle extensions that way.  I think we could reasonably
expect to know about cases where a formerly separate extension got
integrated into core, but are there other cases where pg_upgrade would
need to ignore an extension in the old database?
        regards, tom lane

On Aug 31, 2015, at 4:58 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:

> In any case, there is plenty of precedent for hard-coding knowledge about
> specific version updates into pg_upgrade.  The question here is whether
> it's feasible to handle extensions that way.  I think we could reasonably
> expect to know about cases where a formerly separate extension got
> integrated into core,

+1

> but are there other cases where pg_upgrade would
> need to ignore an extension in the old database?

Not that I can think of, unless it’s already present because it was in template1 or something.

David