Re: Regarding GIN Fast Update Technique

On 07/06/13 16:39, Tom Lane wrote:
> Amit Langote <amitlangote09(at)gmail(dot)com> writes:
>> Okay, aside from that case, what else would move those to the main
>> structure? They (the entries in the unsorted pending list) are in the
>> local memory (work_mem?) of the backend, right?
> No. If they were, it wouldn't be crash-safe.
Thats how it is, but if we could push in wishes, then I would
wish that is woulndn't be crash-safe, and be flushed by the backends
commit. The way it currently operates is that a "random backend"
pays the penalty of other backends pushes to the pending-list and "all
queries"
pays the penalty of searching the pendinglist in queries.

If the pending list were backend only it would not have to be searched
by "all queries" since commit needs to flush it and random backends
wouldn't be penalized. Allthough we'd still have the benefit of batching up
gin-inserts over mulitiple changes to the index done within the same
transaction.

Jesper

--
Jesper