Thread: is this a bug or feature?

hi,

  i was building a testdb for development.  i did a pg_dump -t

to get the tables that i needed.  i then loaded those objects

with psql testdb << dump.file.  when i tried to do an update

against one of the tables i got this error message:


<color><param>ffff,0000,0000</param>psql:update.sql:5: pqReadData() --
backend closed the channel unexpectedly.

        This probably means the backend terminated abnormally

        before or while processing the request.

psql:update.sql:5: connection to server was lost


</color>i got a similar error for insert, delete, or truncate but i could
read

the table without error.  i finally figured out that what had happened

was that the foreign key constraints to and from that table, table x,

that existed in production ok, had loaded into my test area but the

associated tables that would have made those constraints valid did not

exist;  i didn't load them as i didn't need them and had forgotten

that there were RI constraints dependent on them.


of course since i could still select from the table i guess i could go

microsoft and say that it's actually a "read only" feature of postgres!
:)


my question is, why does postgres load up those constraints, when
they'll

be invalid and error out later with a vague error message, instead of

indicating up front that those constraints are there and refuse to load

them?


it took me a while to figure this one out.  you cannot create a foreign

key to or from a non-existing table but the already generated triggers

from previous valid constraint creations will reload, in situations
such

as my test database scenario up above, without initial complaint.


xxxooo

mikeo

> the table without error.  i finally figured out that what had happened
> was that the foreign key constraints to and from that table, table x,
> that existed in production ok, had loaded into my test area but the
> associated tables that would have made those constraints valid did not
> exist;  i didn't load them as i didn't need them and had forgotten>
> that there were RI constraints dependent on them.

Yep, it got noticed on -hackers a few days ago when someone reported
similar symptoms.

The message is because the backend that was handling the request crashes
in that case (so, definately a bug).  The foreign key triggers need to
change a bit anyway and match partial needs to be implemented, so I
plan to put something in to give a message that is meaningful inside
the trigger if you get into this state.  As for not getting into that
state, see below.

> my question is, why does postgres load up those constraints, when
> they'll
> be invalid and error out later with a vague error message, instead of
> indicating up front that those constraints are there and refuse to load
> them?

Mainly because the routine that re-adds the constraint is a much more
general routine (create constraint trigger) which doesn't know anything
about foreign keys.  It's effectively the same thing as calling create
trigger for a c function on a table that exists, but giving it an arg
to one that doesn't and having the trigger try to access the table
assuming it exists (without checking that you successfully did).

Theoretically, I'd guess that the new pg_dump is going to use
alter table add constraint to add the constraints.  That will prevent
this particular way of getting to the problem, although you could still
get into that state if you used a create constraint trigger manually.