Re: [HACKERS] Broken hint bits (freeze)

On Wed, Jun 28, 2017 at 10:11:35PM -0400, Bruce Momjian wrote:
> On Fri, Jun 23, 2017 at 06:17:47PM +0300, Sergey Burladyan wrote:
> > PS:
> > I successfully upgraded last night from 9.2 to 9.4 and find other issue :-)
> > 
> > It is about hash index and promote:
> > 1. create master
> > 2. create standby from it
> > 3. create unlogged table and hash index like:
> >  create unlogged table test (id int primary key, v text);
> >  create index on test using hash (id);
> > 3. stop master
> > 4. promote standby
> > 
> > now, if you try to upgrade this new promoted master pg_upgrade will stop
> > on this hash index:
> > error while creating link for relation "public.test_id_idx" ("s/9.2/base/16384/16393" to "m/9.4/base/16422/16393"):
Nosuch file or directory
 
> > Failure, exiting
> > 
> > I touch this file (s/9.2/base/16384/16393) and rerun pg_upgrade from
> > scratch and it complete successfully.
> 
> Sergey, can you please test if the table "test" is not unlogged, does
> pg_upgrade still fail on the hash index file?

I was able to reproduce this failure on my server.  :-)

What I found is that the problem is larger than I thought.  Sergey is
correct that pg_upgrade fails because there is no hash file associated
with the unlogged table, but in fact a simple access of the unlogged
table with a hash index generates an error:
test=> SELECT * FROM t_u_hash;ERROR:  could not open file "base/16384/16392": No such file or directory

What is interesting is that this is the only combination that generates
an error.  A unlogged able with a btree index or a logged table with a
hash index are fine, e.g.:
           List of relations Schema |   Name    | Type  |  Owner--------+-----------+-------+---------- public |
t_btree  | table | postgres public | t_hash    | table | postgres public | t_u_btree | table | postgres
 
fail-->     public | t_u_hash  | table | postgres

This doesn't fail on PG 10 since we WAL-log hash indexes.

I think we have two questions:

1.  do we fix this in the server
2.  if not, do we fix pg_upgrade

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com

+ As you are, so once was I.  As I am, so you will be. +
+                      Ancient Roman grave inscription +