Hi,
I'm having some trouble getting parallel restores to work. When running pg_restore like this:
pg_restore --no-password --host=$MAIN_CLUSTER-pg-cluster --port=5432 --dbname=$name --jobs=4 --format=d --no-owner
--clean--if-exists --disable-triggers --role=${name}_owner "/temp_data/data-$name-$UNIQUE_SUFFIX"
the restore errors on foreign key constraints while removing tables, is there anyway to keep this parallel but have it
respectthe tree of dependencies.
Errors like this:
pg_restore: error: could not execute query: ERROR: cannot drop table public.organisations because other objects depend
onit
DETAIL: constraint fk_groups_organisations_organisation_id on table public.groups depends on table
public.organisations
constraint fk_organisation_role_organisations_organisation_id on table public.organisation_role depends on table
public.organisations
HINT: Use DROP ... CASCADE to drop the dependent objects too.
Command was: DROP TABLE IF EXISTS public.organisations;
When executed with --jobs=1 this command works.
So I tried the following: (to restore data and make indices with more than one thread but do the initial table creating
andcleaning without)
pg_restore --no-password --host=$MAIN_CLUSTER-pg-cluster --port=5432 --dbname=$name --jobs=1 --section=pre-data
--format=d--no-owner --clean --if-exists --disable-triggers --role=${name}_owner
"/temp_data/data-$name-$UNIQUE_SUFFIX"
pg_restore --no-password --host=$MAIN_CLUSTER-pg-cluster --port=5432 --dbname=$name --jobs=8 --section=data --format=d
--no-owner--clean --if-exists --disable-triggers --role=${name}_owner "/temp_data/data-$name-$UNIQUE_SUFFIX"
pg_restore --no-password --host=$MAIN_CLUSTER-pg-cluster --port=5432 --dbname=$name --jobs=4 --section=post-data
--format=d--no-owner --clean --if-exists --disable-triggers --role=${name}_owner
"/temp_data/data-$name-$UNIQUE_SUFFIX"
But now pre-data does not remove the foreign key constraints at all so all tables can't be removed by --clean, breaking
thewhole thing.
Is there some way to make it do this internally? And this doesn't always reproduce either, seems like a bit of a race
conditionbecause one thread is just a bit too fast dropping tables.
We really want the indices to be created after the data gets put in because it is much much faster, than doing a
schema-onlyand data-only restore.
Kind regards,
Erwin de Haan