> On Oct 6, 2021, at 10:39 AM, Peter Geoghegan <pg@bowt.ie> wrote:
>
>> The differences in the upcoming version are
>>
>> 1) --all no longer means "all relations" but rather "all checkable relations"
>
> Clearly pg_amcheck never checked all relations, because it never
> checked indexes that are not B-Tree indexes. I'm pretty sure that I
> can poke big holes in almost any positivist statement like that with
> little effort.
There is a distinction here that you are (intentionally?) failing to acknowledge. On the one hand, there are relation
typesthat cannot be checked because no checking functions for them exist. (Hash, gin, gist, etc.) On the other hand,
thereare relations which could be check but for the current state of the system, or could be checked in some particular
waybut for the current state of the system. One of those has to do with code that doesn't exist, and the other has to
dowith the state of the system. I'm only talking about the second.
>
>> 2) checking options should be automatically downgraded under circumstances where they cannot be applied
>> 3) unlogged relations during replication are by definition not corrupt
>>
>> I think #1 and #3 are unsurprising enough that they need no documentation update. #2 is slightly surprising (at
leastto me) so I updated the docs for it.
>
> To me #2 sounds like a tautology. It could almost be phrased as
> "pg_amcheck does not check the things that it cannot check".
I totally disagree. It is uncomfortable to me that `pg_amcheck --parent-check` will now silently not perform the
parentcheck that was explicitly requested. That reported an error before, and now it just downgrades the check. This
ishardly tautological. I'm only willing to post a patch with that change because I can see a practical argument that
somebodymight run that as a cron job and they don't want the cron job failing when the database happens to go into
recovery. But again, not at all tautological.
—
Mark Dilger
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
