Re: Re: new patch of MERGE (merge_204) & a question about duplicated ctid

On Mon, Jan 3, 2011 at 11:08 AM, Heikki Linnakangas
<heikki(dot)linnakangas(at)enterprisedb(dot)com> wrote:
> On 03.01.2011 18:02, Robert Haas wrote:
>>
>> On Mon, Jan 3, 2011 at 10:58 AM, Heikki Linnakangas
>> <heikki(dot)linnakangas(at)enterprisedb(dot)com>  wrote:
>>>
>>> On 03.01.2011 17:56, Stephen Frost wrote:
>>>>
>>>> * Robert Haas (robertmhaas(at)gmail(dot)com) wrote:
>>>>>
>>>>> Like Heikki, I'd rather have the feature without a workaround for the
>>>>> concurrency issues than no feature.
>>>>
>>>> I'm still trying to figure out the problem with having the table-level
>>>> lock, unless we really think people will be doing concurrent MERGE's
>>>> where they won't overlap..?  I'm also a bit nervous about if the result
>>>> of concurrent MERGE's would actually be correct if we're not taking a
>>>> bigger lock than row-level (I assume we're taking row-level locks as it
>>>> goes through..).
>>>>
>>>> In general, I also thought/expected to have some kind of UPSERT type
>>>> capability with our initial MERGE support, even if it requires a big
>>>> lock and won't operate concurrently, etc.
>>>
>>> You can of course LOCK TABLE as a work-around, if that's what you want.
>>
>> That work-around completely fails to solve the concurrency problem.
>> Just because you have a lock on the table doesn't mean that there
>> aren't already tuples in the table which are invisible to your
>> snapshot (for example because the inserting transactions haven't
>> committed yet).
>
> It works in read committed mode, because you acquire a new snapshot after
> the LOCK TABLE, and anyone else who modified the table must commit before
> the lock is granted.

Oh, I forgot we hold the ROW EXCLUSIVE lock until commit. That might
be OK, then.

> In serializable mode you get a serialization error.

I don't think this part is true. You can certainly do this:

CREATE TABLE test (a int);
BEGIN TRANSACTION ISOLATION LEVEL SERIALIZABLE;
SELECT * FROM test;
<in another session, insert (1) into test>
LOCK TABLE test IN SHARE MODE; -- or just LOCK TABLE test, if you prefer
SELECT * FROM test; -- still ain't there
INSERT INTO test VALUES (1);

I don't see what would make MERGE immune to this.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company