Re: expanding inheritance in partition bound order

Hi Amit,

On 2017/09/03 16:07, Amit Khandekar wrote:
> On 31 August 2017 at 13:06, Amit Langote <Langote_Amit_f8(at)lab(dot)ntt(dot)co(dot)jp> wrote:
>>> Mind you, that idea has some problems anyway in the face of default
>>> partitions, null partitions, and list partitions which accept
>>> non-contiguous values (e.g. one partition for 1, 3, 5; another for 2,
>>> 4, 6). We might need to mark the PartitionDesc to indicate whether
>>> PartitionDesc-order is in fact bound-order in a particular instance,
>>> or something like that.
>>
>> ISTM, the primary motivation for the EIBO patch at this point is to get
>> the partitions ordered in a predictable manner so that the partition-wise
>> join patch and update partition key patches could implement certain logic
>> using O (n) algorithm rather than an O (n^2) one. Neither of them depend
>> on the actual order in the sense of, say, sticking a PathKey to the
>> resulting Append.
>
> Now that the inheritance hierarchy is expanded in depth-first order,
> RelationGetPartitionDispatchInfo() needs to be changed to arrange the
> PartitionDispatch array and the leaf partitions in depth-first order
> (as we know this is a requirement for the update-partition-key patch
> for efficiently determining which of the leaf partitions are already
> present in the update result rels).

I was thinking the same.

> Amit, I am not sure if you are
> already doing this as part of the patches in this mail thread. Please
> let me know.

Actually, I had thought of changing the expansion order in
RelationGetPartitionDispatchInfo to depth-first after Robert committed his
patch the other day, but haven't got around to doing that yet. Will do
that in the updated patch (the refactoring patch) I will post sometime
later today or tomorrow on a differently titled thread, because the EIBO
work seems to be done.

> Also, let me know if you think there will be any loss of
> efficiency in tuple routing code if we arrange the Partition Dispatch
> indexes in depth-first order.

I don't think there will be any loss in the efficiency of the tuple
routing code itself. It's just that the position of the ResultRelInfos
(of leaf partitions) and PartitionDispatch objects (of partitioned tables)
will be different in their respective arrays, that is, pd->indexes will
now have different values than formerly.

And now because the planner will put leaf partitions subplans / WCOs /
RETURNING projections in that order in the ModifyTable node, we must make
sure that we adapt the same order in the executor, as you already noted.

Thanks,
Amit