When I eval'd it for my book last fall there were big delays in reporting waiting for segments and then also issues with custom reports. I think they have changed the default behavior to get around some of the former, but with MySQL it's always going to be tough for larger queries.
(if there's any performance issue on the measurement side it has more to do with the JavaScript payload because they include a lot in their standard JS bundle).
For comparison, I've also worked with a large variety of dashboards from companies using Snowflake/Tableau that are paying teams millions of dollars a year to build their interfaces, and it's normal to see sluggish load times that are taking 3-10s+.
The big difference is I know what charts and data I want to display, and I build my database from the ground up so it's optimized to show that data. These companies have teams of people that know absolutely nothing about databases and data structures, they just drag a bunch of data together to create some temporary table, then pass that data over to another piece of software. That gets imported into something else where someone that can barely write SQL writes a poorly optimized statement that somehow shows the expected result. They have more flexibility, and everything loads, although performance isn't ideal. Mine either works in 0.1s or it'll never load.
MySQL was too slow for running analytics on a single website? How much data are we talking about?
I've never actually run into performance issues, neither when using it in production professionally, nor for my self-hosted sites (with Matomo always running on-prem). I'd say the performance of PHP and MySQL/MariaDB is most likely decent as long as you don't go too far into specialized workloads, for example log aggregation/tracing; though even some APM solutions like Apache Skywalking also support using traditional RDBMSes for this purpose as well: https://skywalking.apache.org/docs/main/v9.0.0/en/setup/back...
That said, I can't help but to wonder at what actual scale (number of logged events/second, given certain hardware) you'd run into issues. Luckily, because adding basic analytics is usually quite easy, testing this for your own workloads shouldn't be out of the question - then you can let the data speak for itself.