CREATE SUBSCRIPTION fails with long passwords

On Wed, Apr 24, 2019 at 10:07:45AM -0700, Mike Lissner wrote:

This made me a bit crazy tonight. Steps to reproduce:
1. Create a new role and give it a really long password. Ours was 108
chars.
2. Set up all the necessary replication permissions/roles.
3. From another computer, try to connect as that user using psql — it
works!
4. Now try a CREATE SUBSCRIPTION command like:
create subscription my_sub connection 'host=192.168.0.111 port=5432
user=the_new_user password=long_password dbname=my_db sslmode=require'
publication new_server;
You'll get the following error:
ERROR:  could not connect to the publisher: FATAL:  password
authentication failed for user "the_new_user"
Now, go back and shorten that password, and try again. You'll get:
create subscription my_sub connection 'host=192.168.0.111 port=5432
user=the_new_user password=long_password dbname=my_db sslmode=require'
publication new_server;
NOTICE:  created replication slot "my_sub" on publisher
CREATE SUBSCRIPTION

And it'll be off to the races. I watched the logs on both servers during
these experiments. Nothing much in there, aside from the logs above.
Is this known functionality? Seems like a nasty bug and it took me a while
to figure it out.

I've tried to reproduce this on a master, and it works for me (with 200
char passwords). Which version have you used? Can you share an actual
example of commands creating the role/subscription that triggers the
failure, including the password value?

regards

--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

Possibly the password might have non-ascii or something causing it to
unmatched but when stripped off, that non-ascii wasn't included. I'm
curious if you try to have passwords Mike with 108 chars for which chars
are of [0-9] and [a-zA-Z] strings.

On Thu, Apr 25, 2019 at 5:19 AM Tomas Vondra <tomas.vondra@2ndquadrant.com>
wrote:

Paul Mansueto Namuag <paulnamuag@gmail.com> writes:

Possibly the password might have non-ascii or something causing it to
unmatched but when stripped off, that non-ascii wasn't included.

Oh --- another variant of that theory, if there's any non-ASCII in the
password, is that you're getting bit by an encoding conversion issue.
That is, whatever the client is sending is in a different encoding
from the way the password was stored.

regards, tom lane

---one month passes---

Just to close this loop, because I like things closed. I wasn't able to
reproduce this in the end. I'll keep trying, but seems like I must have
mis-diagnosed the issue, I think.

Thanks for all the suggestions and help here. This kind of help and support
is so wonderful and valuable.

Mike

On Thu, Apr 25, 2019 at 6:59 AM Tom Lane <tgl@sss.pgh.pa.us> wrote: