Обсуждение: Fix to enum hashing for dump and restore

Hello,

I have discovered a bug in one usage of enums. If a table with hash
partitions uses an enum as a partitioning key, it can no longer be
backed up and restored correctly. This is because enums are represented
simply as oids, and the hash function for enums hashes that oid to
determine partition distribution. Given the way oids are assigned, any
dump+restore of a database with such a table may fail with the error
"ERROR: new row for relation "TABLENAME" violates partition constraint".

This can be reproduced with the following steps:
************************************************
create database test;
\c test
create type colors as enum ('red', 'green', 'blue', 'yellow');
create table part (color colors) partition by hash(color);

create table prt_0 partition of part for values 
with (modulus 3, remainder 0);

create table prt_1 partition of part for values 
with (modulus 3, remainder 1);

create table prt_2 partition of part for values 
with (modulus 3, remainder 2);
insert into part values ('red');

/usr/local/pgsql/bin/pg_dump -d test -f /tmp/dump.sql
/usr/local/pgsql/bin/createdb test2
/usr/local/pgsql/bin/psql test2 -f /tmp/dump.sql
************************************************

I have written a patch to fix this bug (attached), by instead having the
hashenum functions look up the enumsortorder ID of the value being
hashed. These are deterministic across databases, and so allow for
stable dump and restore. This admittedly comes at the performance cost
of doing a catalog lookup, but there is precedent for this in
e.g. hashrange and hashtext.

I look forward to your feedback on this, thank you!

Sincerely,
Andrew J Repp (VMware)

Andrew <pgsqlhackers@andrewrepp.com> writes:
> I have discovered a bug in one usage of enums. If a table with hash
> partitions uses an enum as a partitioning key, it can no longer be
> backed up and restored correctly. This is because enums are represented
> simply as oids, and the hash function for enums hashes that oid to
> determine partition distribution. Given the way oids are assigned, any
> dump+restore of a database with such a table may fail with the error
> "ERROR: new row for relation "TABLENAME" violates partition constraint".

Ugh, that was not well thought out :-(.  I suppose this isn't a problem
for pg_upgrade, which should preserve the enum value OIDs, but an
ordinary dump/restore will indeed hit this.

> I have written a patch to fix this bug (attached), by instead having the
> hashenum functions look up the enumsortorder ID of the value being
> hashed. These are deterministic across databases, and so allow for
> stable dump and restore.

Unfortunately, I'm not sure those are as deterministic as all that.
They are floats, so there's a question of roundoff error, not to
mention cross-platform variations in what a float looks like.  (At the
absolute minimum, I think we'd have to change your patch to force
consistent byte ordering of the floats.)  Actually though, roundoff
error wouldn't be a problem for the normal exact-integer values of
enumsortorder.  Where it could come into play is with the fractional
values used after you insert a value into the existing sort order.
And then the whole idea fails, because a dump/restore won't duplicate
those fractional values.

Another problem with this approach is that we can't get from here to there
without a guaranteed dump/reload failure, since it's quite unlikely that
the partition assignment will be the same when based on enumsortorder
as it was when based on OIDs.  Worse, it also breaks the pg_upgrade case.

I wonder if it'd work to make pg_dump force --load-via-partition-root
mode when a hashed partition key includes an enum.

            regards, tom lane

Andrew <pgsqlhackers@andrewrepp.com> writes:

> I have discovered a bug in one usage of enums. If a table with hash

> partitions uses an enum as a partitioning key, it can no longer be

> backed up and restored correctly. This is because enums are represented

> simply as oids, and the hash function for enums hashes that oid to

> determine partition distribution. Given the way oids are assigned, any

> dump+restore of a database with such a table may fail with the error

> "ERROR: new row for relation "TABLENAME" violates partition constraint".

Ugh, that was not well thought out :-(.  I suppose this isn't a problem

for pg_upgrade, which should preserve the enum value OIDs, but an

ordinary dump/restore will indeed hit this.

> I have written a patch to fix this bug (attached), by instead having the

> hashenum functions look up the enumsortorder ID of the value being

> hashed. These are deterministic across databases, and so allow for

> stable dump and restore.

Unfortunately, I'm not sure those are as deterministic as all that.

They are floats, so there's a question of roundoff error, not to

mention cross-platform variations in what a float looks like.  (At the

absolute minimum, I think we'd have to change your patch to force

consistent byte ordering of the floats.)  Actually though, roundoff

error wouldn't be a problem for the normal exact-integer values of

enumsortorder.  Where it could come into play is with the fractional

values used after you insert a value into the existing sort order.

And then the whole idea fails, because a dump/restore won't duplicate

those fractional values.

Another problem with this approach is that we can't get from here to there

without a guaranteed dump/reload failure, since it's quite unlikely that

the partition assignment will be the same when based on enumsortorder

as it was when based on OIDs.  Worse, it also breaks the pg_upgrade case.

I wonder if it'd work to make pg_dump force --load-via-partition-root

mode when a hashed partition key includes an enum.

regards, tom lane