any

Thanks for the report.

Reproducible easily enough on all of 10.3-10.10 (probably earlier ones too, I didn't check), as well as MySQL 5.7 and 8.0.

<?php

$port=3306;

$connroot = new mysqli('127.0.0.1:'.$port,'root','','');

$connroot->query("DROP USER IF EXISTS test_user@localhost");

$connroot->query("CREATE USER test_user@localhost WITH MAX_USER_CONNECTIONS 1");

$connroot->close();

foreach (range(0, 50_000) as $i) {

    $conn=new mysqli('127.0.0.1:'.$port,'test_user','','');

    if ($conn->connect_error) {

      die("Connection #$i failed: ".$conn->connect_error."\n\n");

    }

    $conn->close();

}

echo "All done\n";

exit;

?>

Disclaimer: I'm not a PHP writer, feel free to fix it as needed.

It usually fails in seconds with something like

PHP Warning:  mysqli::__construct(): (HY000/1226): User 'test_user' has exceeded the 'max_user_connections' resource (current value: 1) in /mnt8t/src/mariadb-toolbox/templates/test.php on line 11

Connection #29777 failed: User 'test_user' has exceeded the 'max_user_connections' resource (current value: 1)

In the error log, if enabled:

                143019 Connect  test_user@localhost as anonymous on 

                143019 Quit     

                143020 Connect  test_user@localhost as anonymous on 

                143020 Quit     

                143021 Connect  test_user@localhost as anonymous on 

                143022 Connect  test_user@localhost as anonymous on 

                143021 Quit     

Sometimes it misses the mark though, re-run then (increase the number of loops if necessary).

I don't think it's a bug as such. More like an implementation detail. Our own tests work around it. Supposedly, clients appreciate lower latency, and that's why a response packet as sent as soon as possible

With many fast reconnects the "active connection counter" in MariaDB can show value like 5 even there is a single active connection socket. This does not only causes problems with tests, but also in a real apps that connect and disconnect often.

The current behaviour requires the max. connection limit to be set much higher than needed or it causes random connect issues hard to debug.

> clients appreciate lower latency

How much the disconnect latency will increase when the "active connection counter" will be counted synchronously with client connect/reconnect? I doubt it will be more than 100us.