On 24 October 2010 17:20, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Greg Stark <gsstark@mit.edu> writes:
>> There's nothing magic about the integral types here. If you use a
>> string then you could always split by making the string longer.
>
> The entire objective here is to make enum comparisons fast. Somehow,
> going to a non-primitive data type to represent the sort values does
> not seem like a win from that perspective.
>
> (Likewise for Dean's suggestion for an array of another kind.)
>
The point with an OID array is that you wouldn't need to store the
enumsortorder values at all. The sort order would just be the index of
the OID in the array. So the comparison code would read the OID array,
traverse it building an array of enum_sort structs {oid, idx}, sort
that by OID and cache it. Then do binary searches to efficiently find
the index (i.e., sort order) for any given OID. That's pretty much
what the comparison code is doing now, except without explicitly
stored sort positions.
The reason I thought this might help is that it would be a single
atomic operation to read the whole enum sort order, so you'd never get
a mix of old and new sort positions, if another transaction was
altering the enum.
> What I'm currently thinking is that we should actually use float4 not
> float8. That eliminates any concerns about making the representation
> larger than it was before. We'll definitely have to have the logic
> to do renumbering, but it'll still be an exceedingly rare thing in
> practice. With float4 the implementation would fail at somewhere
> around 2^24 elements in an enum (since even with renumbering, there
> wouldn't be enough bits to give each element a distinguishable value).
> I don't see that as a real objection, and anyway if you were trying
> to have an enum with many elements, you'd want the in-memory
> representation to be compact.
>
That seems like a lot of additional complexity to do the renumbering,
and you'd have to make the comparison code safe against a concurrent
renumbering.
Regards,
Dean
> regards, tom lane
>