On Mon, Aug 28, 2017 at 12:23 PM, Antonin Houska <ah@cybertec.at> wrote:
> Now that dynamic shared memory hash table has been committed
> (8c0d7bafad36434cb08ac2c78e69ae72c194ca20) I wonder if it's still a big deal
> to remove restrictions like this in (e.g. heap_update()):
>
> /*
> * Forbid this during a parallel operation, lest it allocate a combocid.
> * Other workers might need that combocid for visibility checks, and we
> * have no provision for broadcasting it to them.
> */
> if (IsInParallelMode())
> ereport(ERROR,
> (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
> errmsg("cannot update tuples during a
> parallel operation")));
Well, it certainly doesn't solve that problem directly, but I think it
is useful infrastructure that lends itself to some kind of a solution
to that problem. There are other problems, too, like group locking
vs. the relation extension lock, though Masahiko-san posted a patch
for that IIRC, and also group locking vs. GIN page locks, and also
proper interlocking of updates. I think that the order in which we
should tackle things is probably something like:
1. Make inserts parallel-restricted rather than parallel-unsafe - i.e.
allow inserts but only in the leader. This allows plans like Insert
-> Gather -> whatever. I'm not sure there's a lot to do here other
than allow such plans to be generated.
2. Make updates and deletes parallel-restricted rather than
parallel-unsafe - i.e. allow updates and deletes but only in the
leader. This similarly would allow Update -> Gather -> whatever and
Delete -> Gather -> whatever. For this, you'd need a shared combo CID
hash so that workers can learn about new combo CIDs created by the
leader.
3. Make inserts parallel-safe rather than parallel-restricted. This
allows plans like Gather -> Insert -> whatever, which is way better
than Insert -> Gather -> whatever. If there's no RETURNING, the
Gather isn't actually gathering anything. This requires sorting out
some group locking issues, but not tuple locking.
4. Make updates and deletes parallel-safe rather than
parallel-restricted. Now tuple locking has to be sorted out.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company