ExtraVM's Singapore node feels oversold since August — anyone else or is it me?

Two boxes in Singapore, both since March, both fine until about the second week of August. Since then steal time on both averages 9-11% where it used to sit under 2, and disk latency spikes in a pattern that looks like a neighbour running something heavy on a schedule.

# 5 min samples over 24h, one of the two
vmstat 300 288 | awk 'NR>2 {s+=$17; n++} END {print "avg st:", s/n}'
# avg st: 9.7

Ticket came back saying the node is within normal parameters, which may well be true and doesn't help me. Anyone else in that region seeing it, or should I stop blaming the host? FWIW the US boxes on the same account are unchanged.

Not Singapore but Netherlands node was same in July. Went away after three week. Somebody left, probably.

Depends what you're running, but at three and a half dollars a month you're on shared cores and the contract is explicitly best-effort. Nine per cent steal isn't a fault, it's the product.

Not saying don't complain. Saying the ticket will always come back within normal parameters, because it is.

Did you benchmark it in March, or is under 2% a memory?

Benchmarked. I keep a week of vmstat on everything, which is the only reason I'm confident enough to post rather than just complain.

March median steal 1.4%, August 9.7%, September so far 10.2%.

The APAC nodes at the cheap end have been oversold since about 2019 and it isn't specific to this host. Singapore floor space costs what it costs, and a three-dollar box in Singapore is a three-dollar box in Singapore.

If the workload cares about steal time, the cheap tier in that region is the wrong product and no amount of ticketing changes it. If it doesn't care, you've built a monitoring habit that's making you unhappy about a number that isn't costing you anything.

I mean that second one seriously. Half the people who start measuring find things they then can't stop looking at.

Two weeks on and it hasn't improved. Moved the latency-sensitive half to a dedicated-core plan elsewhere and left the rest.

Not calling it a fault. Calling it a product I misread, which is closer to the truth.