Re: Index corruption with CREATE INDEX CONCURRENTLY

On 2017-02-06 08:08:01 +0530, Amit Kapila wrote:
> I don't see in your patch that you are setting rd_bitmapsvalid to 0.

IIRC a plain relcache rebuild should do that (note there's also no place
that directly resets rd_indexattrs).

> Also, I think this suffers from the same problem as the patch proposed
> by Alvaro which is that different attrs (hot_attrs, key_attrs and
> id_attrs) will get computed based on different index lists which can
> cause a problem as mentioned in my mail up thread.

I'm not convinced that that's a legitimate concern. If that's an issue
with CIC, then we have a lot bigger problems, because relcache entries
and such will be inconsistent anyway if you have non-exclusive locks
while changing catalog data. In the situations where that happens it
better be harmless (otherwis CiC is broken), and the next access will
recompute them.

> I am not sure but I think your solution can be
> extended to make it somewhat similar to what we do with rd_indexvalid
> (basically, we should set rd_bitmapsvalid to 2 (bitmap is temporarily
> forced) at rel cache invalidation and then clean it up transaction
> end). I can try something on those lines.

Not sure I understand what you mean with "clean it up at transaction end"?

Greetings,

Andres Freund