You don't need a new metric for this, you can use any existing ones mysql_slave_status_*.
For example

absent(mysql_slave_status_connect_retry{instance="foobar"})

returns 1 if slave is not configured or nothing if it is.

As I mentioned, absent() is not ideal here because we have hundreds of mysql hosts that we need to alert on in the event of replication misconfiguration, and defining an alert for instance="foo001" through instance="foo999" is not practical. With this metric all we need to do is define a single alert: mysql_slave_status_is_configured == 0.

When replication is not configured at all (after a reset slave command, for example) none of the mysql_slave_status_* metrics exist prior to this PR.

All you need is a regex to match all your mysql hosts {instance=~"foo\d{3}"} or some any other label you may want to tag your mysql instances to be captured.

Sorry, no, this doesn't work. in order for absent() to return one on a regex match, ALL hosts that match the regex must be missing the metric. I need to know if any single host is missing the metric.

I just spend hours trying to create an alert if one of my replica doesn't replicate. The stale metrics render things very difficult. I can confirm what @mmiller1 is explaining to @roman-vynar, I tried group_left, unless, offset, absent_over_time, nothing works.
Isn't stale metrics a bad practice for Prometheus anyway?

I'm using the alerts from awesome-prometheus and was surprised that no alerts are triggered for replica that has been reset (I have a switchover script that fails sometimes).

How about this?

expr: count(node_uname_info{instance=~"mysql. "}) by (instance) unless count(mysql_slave_status_connect_retry) by (instance)

This means instances named mysql* will have node_uname_info metric but not mysql_slave_status_connect_retry and then to alert.

Genius @roman-vynar thanks for your help!
My luck is that I created a metric called db_role with a textfile collector that read every minute a custom facts that is written by Ansible. This way, even after a switchover, db_role is always right.

So the correct expression for the alert that works if the metric == 0 AND is staled:

expr: count(db_role == 0) by (instance) unless count(mysql_slave_status_slave_io_running == 1) by (instance) == 1

db_role : 0 => replica
db_role : 1 => primary

It works perfectly. But you have to wait 5 minutes for the metric to be staled. But that's far better than no alert at all!