FWIW pghr (the initial experiment) has largely served its purpose for me. As I mentioned a few posts ago, the idea was that I would layer on the complexity in our internal application and find the cliff that we were falling off of.
I found it today.
We were using pg_notify as part of an internal change notification mechanism. When I introduced that (even with no listeners), Ecto performance fell off by a factor of 3. pgbench performance showed no such impact.
I’ll post updated results on the pghr repo soon (hopefully tomorrow). In the meantime, armed with that information, we can achieve the desired performance by using other technology for our change notification.
