Zhang, I think there was some trouble submitting the post, The web server kept throwing an error about how something went wrong sending an email, so in my attempt to post it (back->click submit again) it posted a duplicate while throwing the same error, while never indicating it actually posted it the first (or second) time. I deleted the duplicate post.

Ah-ha! Thank you, that was my missing link. I needed to use 'sa-learn -u user@domain.com --dump' to get the Bayes stats for the particular user (me) I was testing to train.

So that is showing it is learning at the moment, but now I ran into bad situation training it.
I tried about 10 at a time, and that seemed to work okay. I then did an IMAP move of about 300 emails into two training folders I made (TrainGood, TrainSpam).  4 Hours later, it was still working on moving/training them. The CPU usage was between 30-60% the whole time, I/O charts were very slow, the actual processes of sa-learn (I noted up to 2 at a time running) seemed to be only taking 5-15% cpu (viewed with top).  The mail server however effectively halted its other processing and during this time the mail server was too slow to use at all. This seemed very strange since it appeared not fully loaded, but I couldn't figure out why this wasnt working.

I decided to stop the process and revert to a non-learning mode to see what effect that had - I commented out the modified plugin section of dovecot.conf, restarted dovecot and amavisd, and all went back to normal. The IMAP client seemed to complete its pending operations, and I was able to use the mail server again.

I tested by moving a few hundred emails, and they went almost instantly.

Then I re-enabled the anti spam portion of dovecot.conf, restarted dovecot and amavisd again, and tested a little at a time. By the time I was confident enough I tried moving 90 emails that it would train as spam. This operation took it about 60 seconds to do (~1.5/second). The system is running on 4 CPUs x Intel(R) Xeon(R) CPU E3-1240 v6 @ 3.70GHz, running ESXi with only one VM, this mail server on CentOS 7. Somehow I thought it would be much faster, as my much slower 5 year old mac mini seemed to train faster than that. Is this normal? I'm not sure why that would be so slow.

But then came the real kicker - I looked at the spam training folder that messages were copying to when I had to first stop dovecot after 4 hours. It has a bunch of spam it had trained, and then it had 256 copies of a spam in the middle. Then there were 44 more unique spams, and then another 256 of a different spam in the middle. Then some more unique ones.
I can't imagine what went wrong initially to cause that, or how to ensure it does not happen again...

Have you ever seen anything like this?  I don't know what to look at or monitor to see what is going on here.

Thank you for your help here. My goal is to get this working well/reliably enough to make this my permanent home server solution for at least the next few years with the iRedAdmin-Pro. This has just been a painful struggle for some reason.

I expected you to read my reply in the link you posted in first post:
https://forum.iredmail.org/post59958.html#p59958

If you call 'sa-learn' every time you move a message (and multiple calls if you move multiple messages), it's very slow, and webmail may hang and no response.

A better way is: configure the plugin to save a copy of moved message, and call 'sa-learn' every, e.g. 10 minutes to learn all moved message at the same time (only one 'sa-learn' process). this way sa-learn runs fast.

I'm afraid that others cannot help without any detail of your implementation.

- Is it problem of your sieve rule/script?
- Problem of other part?

Okay, I think possibly I made a mistake in this config .. Can anyone look at this and tell me whats wrong here?  I'm afraid I don't understand this part very well.

use_bayes          1
bayes_auto_learn   1
bayes_auto_expire  1

# Store bayesian data in MySQL.
# Please make sure you have correct server address, port and database name.
bayes_store_module Mail::SpamAssassin::BayesStore::MySQL
bayes_sql_dsn      DBI:mysql:sa_bayes:127.0.0.1:3306

# Store bayesian data in PostgreSQL.
# Please make sure you have correct server address, port and database name.
#bayes_store_module Mail::SpamAssassin::BayesStore::PgSQL
#bayes_sql_dsn      DBI:Pg:sa_bayes:127.0.0.1:5432

# SQL username and password.
bayes_sql_username sa_bayes_admin
bayes_sql_password RemovedMyAdminPasswordForPosting

# Override the username used for storing data in the database.
# This could be used to group users together to share bayesian filter data.
# You can also use this config option to trick sa-learn to learn data as a specific user.
#bayes_sql_override_username vmail

[Edit- hit submit too early]
I noticed that my last line for override username is still commented out, and shows user vmail. this is I believe how this iRedMail config file came. The example I was trying to follow from https://docs.iredmail.org/store.spamass … n.sql.html shows the line uncommented, with the username specified as amavis.