Re: WIP: Deferrable unique constraints

2009/7/29 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>:
> Another thought on the index AM API issues: after poking through the
> code I realized that there is *nobody* paying any attention to the
> existing bool result of aminsert() (ie, did we insert anything or not).
> So I think that instead of adding a bool* parameter, we should repurpose
> the function result, along the lines of this spec:
>
>  <para>
>   The method's boolean result value is significant only when
>   <literal>checkUnique</> is <literal>UNIQUE_CHECK_PARTIAL</>.
>   In this case a TRUE result means the new entry is known unique, whereas
>   FALSE means it might be non-unique (and a deferred uniqueness check must
>   be scheduled).  For other cases a constant FALSE result is recommended.
>  </para>
>

And you'll be moving the ereport() back into the btree code? Makes
sense, provided that nothing is ever going to care whether the index
actually inserted an entry. I can see arguments for making the
recommended return value for "other cases" either TRUE or FALSE, but I
guess it doesn't matter since nothing is going to check it.

>  <para>
>   For non-unique indexes, it is not required that <function>aminsert</>
>   do anything; it might for instance refuse to index NULLs.
>  </para>
>

Doesn't this comment apply equally to unique indexes?

- Dean