Simon Riggs <simon@2ndquadrant.com> wrote:
> The LOCK is only required because we defer the inserts into unique
> indexes, yes?
No, as far as present pg_bulkload. It creates a new relfilenode like REINDEX,
therefore, access exclusive lock is needed. When there is violations of
unique constraints, all of the loading is rollbacked at the end of loading.
BTW, why REINDEX requires access exclusive lock? Read-only queries
are forbidden during the operation now, but I feel they are ok
because REINDEX only reads existing tuples. Can we do REINDEX
holding only shared lock on the index?
> I very much like the idea of index merging, or put another way: batch
> index inserts. How big do the batch of index inserts have to be for us
> to gain benefit from this technique?
Hmm, we might need to know *why* COPY with indexes is slow. If the major
cause is searching position to insert, batch inserts will work well.
However, if the cause is index splitting and following random i/o,
batch insertion cannot solve the problem; "rebuild" is still required.
Regards,
---
ITAGAKI Takahiro
NTT Open Source Software Center
