Thread: pg_dumpall problem when roles have default schemas
I've run into a problem while migrating an existing 8.2.7 data base to a new server running 8.3.3 (although I think the version numbers may not matter -- I think I've seen this problem in the past and just lived with it since so much of Postgresql is so great!).
The problem stems from the fact that for certain roles, I have defined default search paths, and because the output of pg_dumpall includes role definitions first, then data base definitions, then schema definitions.
Thus, when piping the output (from legacy host 192.168.2.2) to populate the newly initialized cluster, by way of running (on the new host 192.168.2.3)
pg_dumpall -h 192.168.2.2|psql
an error occurs in that first section when the script attempts to set a role-specific search path ... because the schema named in the search path hasn't been created yet.
Not sure if there is some different way I should be using these tools to accomplish this, or if there is a work-around, or if this feature needs improvement.
btober@ct.metrocast.net writes: > Thus, when piping the output (from legacy host 192.168.2.2) to > populate the newly initialized cluster, by way of running (on the new > host 192.168.2.3) > pg_dumpall -h 192.168.2.2|psql > an error occurs in that first section when the script attempts to > set a role-specific search path ... because the schema named in the > search path hasn't been created yet. Could we see a complete example? Recent versions of the backend are not supposed to throw hard errors in this situation, because of exactly that hazard. For instance: regression=# create role joe; CREATE ROLE regression=# alter role joe set search_path to foo, bar; NOTICE: schema "foo" does not exist NOTICE: schema "bar" does not exist ALTER ROLE regression=# which AFAICS is exactly what will happen while restoring a pg_dumpall dump. regards, tom lane
Tom Lane wrote:
> btober@ct.metrocast.net writes:
>> Thus, when piping the output (from legacy host 192.168.2.2) to
>> populate the newly initialized cluster, by way of running (on the new
>> host 192.168.2.3)
>> pg_dumpall -h 192.168.2.2|psql
>> an error occurs in that first section when the script attempts to
>> set a role-specific search path ... because the schema named in the
>> search path hasn't been created yet.
>
> Could we see a complete example?
>
> Recent versions of the backend are not supposed to throw hard errors in
> this situation, because of exactly that hazard. For instance:
>
> regression=# create role joe;
> CREATE ROLE
> regression=# alter role joe set search_path to foo, bar;
> NOTICE: schema "foo" does not exist
> NOTICE: schema "bar" does not exist
> ALTER ROLE
> regression=#
>
> which AFAICS is exactly what will happen while restoring a pg_dumpall
> dump.
Ah, I understand. I think.
What I saw was not literally and "error", but rather a "notice": The "alter role ... set search_path" statement doesn't actually fail.
The real problem is my preconceived notion. I was thinking in terms of, say, the analogous situation for a database insert to a table with a foreign key, i.e., that since the schema to be set doesn't exist, the the statement should actually fail.
The implication of your illustration above is that the "alter role ... set search_path" statement completes successfully even though it is setting an invalid search path (invalid because the schema referenced in the search path to be set does not exist at the time the statement is executed).
So this behavior, which seems a little counter intuitive, actually makes for a robust end result.
> btober@ct.metrocast.net writes:
>> Thus, when piping the output (from legacy host 192.168.2.2) to
>> populate the newly initialized cluster, by way of running (on the new
>> host 192.168.2.3)
>> pg_dumpall -h 192.168.2.2|psql
>> an error occurs in that first section when the script attempts to
>> set a role-specific search path ... because the schema named in the
>> search path hasn't been created yet.
>
> Could we see a complete example?
>
> Recent versions of the backend are not supposed to throw hard errors in
> this situation, because of exactly that hazard. For instance:
>
> regression=# create role joe;
> CREATE ROLE
> regression=# alter role joe set search_path to foo, bar;
> NOTICE: schema "foo" does not exist
> NOTICE: schema "bar" does not exist
> ALTER ROLE
> regression=#
>
> which AFAICS is exactly what will happen while restoring a pg_dumpall
> dump.
Ah, I understand. I think.
What I saw was not literally and "error", but rather a "notice": The "alter role ... set search_path" statement doesn't actually fail.
The real problem is my preconceived notion. I was thinking in terms of, say, the analogous situation for a database insert to a table with a foreign key, i.e., that since the schema to be set doesn't exist, the the statement should actually fail.
The implication of your illustration above is that the "alter role ... set search_path" statement completes successfully even though it is setting an invalid search path (invalid because the schema referenced in the search path to be set does not exist at the time the statement is executed).
So this behavior, which seems a little counter intuitive, actually makes for a robust end result.
Tom Lane wrote: > btober@ct.metrocast.net writes: >> Thus, when piping the output (from legacy host 192.168.2.2) to >> populate the newly initialized cluster, by way of running (on the new >> host 192.168.2.3) >> pg_dumpall -h 192.168.2.2|psql >> an error occurs in that first section when the script attempts to >> set a role-specific search path ... because the schema named in the >> search path hasn't been created yet. > > Could we see a complete example? > > Recent versions of the backend are not supposed to throw hard errors in > this situation, because of exactly that hazard. For instance: > > regression=# create role joe; > CREATE ROLE > regression=# alter role joe set search_path to foo, bar; > NOTICE: schema "foo" does not exist > NOTICE: schema "bar" does not exist > ALTER ROLE > regression=# > > which AFAICS is exactly what will happen while restoring a pg_dumpall > dump. Ah, I understand. I think. What I saw was not literally and "error", but rather a "notice": The "alter role ... set search_path" statement doesn't actually fail. The real problem is my preconceived notion. I was thinking in terms of, say, a database insert to a table with a foreign key, i.e., that since the schema to be set doesn't exist, the the statement should actually fail. The implication of your illustration above is that the "alter role ... set search_path" statement completes successfully even though it is setting an invalid search path (invalid because the schema referenced in the search path to be set does not exist at the time the statement is executed). So this behavior, which seems a little counter intuitive, actually makes for a robust end result.