We run an official mirror for various linux distros including CentOS, Fedora, EPEL etc.

We see this on our mirror nodes:

This comes from conntrack -L | wc -l

Analyzing the logs gives more hits close to the hour for:
/mirrors/linux/centos/7{}.9{}.2009{}/extras/x86{}_64{}/repodata/repomd.xml
/mirrors/linux/centos/7{}.9{}.2009{}/os/x86{}_64{}/repodata/repomd.xml
/mirrors/linux/centos/7{}.9{}.2009{}/updates/x86{}_64{}/repodata/repomd.xml

Seeing time distribution for repomd.xml for CentOS7 gives this pattern:

I believe this might come from yum-cron and
/etc/yum/yum-cron-hourly.conf (el7) which includes
random_sleep = 15 # in minutes

Can this be changed globally (an updated package) and put 60 minutes like in RHEL 8, 9
/etc/systemd/system/timers.target.wants/dnf-makecache.timer
which has RandomizedDelaySec=60m ?

It will help spreading the load.

Thanks