On 8/24/07, Tom Lane <tgl@sss.pgh.pa.us> wrote:"Steven Flatt" <steven.flatt@gmail.com> writes:
> Why do we even need to consider calling RelationGetNumberOfBlocks or looking
> at the pg_class.relpages entry? My understanding of the expected behaviour
> is that while a reindex is happening, all queries run against the parent
> table are planned as though the index isn't there (i.e. it's unusable).
Where in the world did you get that idea?
Maybe that's what I was *hoping* the behaviour would be. :)
From the docs:
"REINDEX locks out writes but not reads of the index's parent table."
"It also takes an exclusive lock on the specific index being processed..."
I believe those two statements imply that reads of the parent table don't take any lock whatsoever on the index being processed, i.e. they ignore it.
If we had a REINDEX CONCURRENTLY it might work that way. A normal
REINDEX cannot "mark" anything because it runs within a single
transaction; there is no way that it can emit any catalog changes
that will be visible before it's over.
... but I understand this difficulty.
So, can we simply trust what's in pg_class.relpages and ignore looking directly at the index? This is a fairly serious concern for us, that reindex is blocking all readers of the parent table.
Thanks,
Steve