Re: Proposal: Local indexes for partitioned table

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> On Fri, Oct 27, 2017 at 7:27 PM, Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> wrote:
>> I noticed that RelationBuildPartitionKey is generating a partition key
>> in a temp context, then creating a private context and copying the key
>> into that. That seems leftover from some previous iteration of some
>> other patch; I think it's pretty reasonable to create the new context
>> right from the start and allocate the key there directly instead. Then
>> there's no need for copy_partition_key at all.

> We could do that, but the motivation for the current system was to
> avoid leaking memory in a long-lived context. I think the same
> technique is used elsewhere for similar reasons. I admit I haven't
> checked whether there would actually be a leak here if we did it as
> you propose, but I wouldn't find it at all surprising.

Another key point is to avoid leaving a corrupted relcache entry behind
if you fail partway through. I think that the current coding is
RelationBuildPartitionKey is safe, although it's far too undercommented
for my taste. (The ordering of the last few steps is *critical*.)

It might work to build the new key in a context that's initially a
child of CurrentMemoryContext, then reparent it to be a child of
CacheMemoryContext when done. But you'd have to look at whether that
would end up wasting long-lived memory, if the build process generates
any cruft in the same context.

regards, tom lane