Percona Server seems great to me overall. If I had to think of any small area for improvement for Percona Server, it would be around the proxy story. When adding a proxy to manage traffic better, we first tried SQL Proxy, but it never really worked for us, which was unfortunate because SQL Proxy offers better optimizations for read performance. We did have success with HAProxy, but it works on a network level, while SQL Proxy was designed to be SQL-aware and distribute parts of the query across the database cluster. The performance gain we were hoping for with SQL Proxy did not materialize, so we remained on the network level with HAProxy. Databases can be finicky regarding the performance they need. We had some disk I/O latency issues, which led us to use different types of nodes in our Kubernetes cluster where we provisioned Percona Server explicitly. There is a lot of IOPS happening under the hood, and while I suspect the application was not very efficient, it was also challenging to identify that we were experiencing performance issues due to disk performance. It would be helpful to clarify the metrics related to disk I/O latency so we could understand where our bottlenecks are.
