On read latency budgets
When the read SLO is p99 < 50ms, the network round-trip alone often spends a third of that. I’ve been measuring our internal cache hits and the long tail is dominated by GC pauses, not the network.
When the read SLO is p99 < 50ms, the network round-trip alone often spends a third of that. I’ve been measuring our internal cache hits and the long tail is dominated by GC pauses, not the network.
An upstream that closes its side first leaves a half-open socket sitting
in FIN_WAIT_2 until the kernel’s timeout. Set
tcp_fin_timeout sanely or your conntrack table fills up.
The best technical decisions I’ve made this year were choosing the older, less interesting option. The novel one always promises more and bills more.