Starting a new thread to discuss the proposal I put forward in [1] to stop
creating explicit NOT NULL constraint on range partition keys that are
simple columns. I said the following:
On 2017/05/12 11:20, Robert Haas wrote:
> On Thu, May 11, 2017 at 10:15 PM, Amit Langote
> <Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> wrote:
>> On 2017/05/12 10:42, Robert Haas wrote:
>>> Actually, I think that not supporting nulls for range partitioning may
>>> have been a fairly bad decision.
>>
>> I think the relevant discussion concluded [1] that way, because we
>> couldn't decide which interface to provide for specifying where NULLs are
>> placed or because we decided to think about it later.
>
> Yeah, but I have a feeling that marking the columns NOT NULL is going
> to make it really hard to support that in the future when we get the
> syntax hammered out. If it had only affected the partition
> constraints that'd be different.
So, adding keycol IS NOT NULL (like we currently do for expressions) in
the implicit partition constraint would be more future-proof than
generating an actual catalogued NOT NULL constraint on the keycol? I now
tend to think it would be better. Directly inserting into a range
partition with a NULL value for a column currently generates a "null value
in column \"%s\" violates not-null constraint" instead of perhaps more
relevant "new row for relation \"%s\" violates partition constraint".
That said, we *do* document the fact that a NOT NULL constraint is added
on range key columns, but we might as well document instead that we don't
currently support routing tuples with NULL values in the partition key
through a range-partitioned table and so NULL values cause error.
Can we still decide to do that instead?
Thanks,
Amit
[1]
https://www.postgresql.org/message-id/f52fb5c3-b6ad-b6ff-4bb6-145fdb805fa6%40lab.ntt.co.jp
