Bug#988191: ionos-1 proxy causes lots of temporary build failures by being slow to respond
On Fri, May 07, 2021 at 01:50:04PM +0200, Helmut Grohne wrote: > I'm unsure what you mean here precisely. Do you mean adding a separate > squid on ionos9 used by only ionos9 yes, that. > If the proxy is only used locally, adding another cpu is unnecessary as > the times when the proxy is being used we are not cpu bound. Adding a > bit of ram for that case sounds sensible though 2gb should already > suffice in that case. ok, cool. > Possibly adding a bit of disk space for the squid spool (20G?) would > be necessary to avoid excessive refetching. ok. -- cheers, Holger ⢀⣴⠾⠻⢶⣦⠀ ⣾⠁⢠⠒⠀⣿⡁ holger@(debian|reproducible-builds|layer-acht).org ⢿⡄⠘⠷⠚⠋⠀ OpenPGP: B8BF54137B09D35CF026FE9D 091AB856069AAA1C ⠈⠳⣄ In Europe there are people prosecuted by courts because they saved other people from drowning in the Mediterranean Sea. That is almost as absurd as if there were people being prosecuted because they save humans from drowning in the sea. signature.asc Description: PGP signature
Bug#988191: ionos-1 proxy causes lots of temporary build failures by being slow to respond
Hi Holger, On Fri, May 07, 2021 at 11:32:21AM +, Holger Levsen wrote: > as a simple fix: we could install squid on your node (ionos9?) and a another > cpu for it? (and 4gb ram or such) I'm unsure what you mean here precisely. Do you mean adding a separate squid on ionos9 used by only ionos9 or do you mean moving the squid from ionos1 to ionos9 and moving all of its users over to ionos9? Both seem sensible to me, but the resource adjustments are slightly different in each case. If the proxy is only used locally, adding another cpu is unnecessary as the times when the proxy is being used we are not cpu bound. Adding a bit of ram for that case sounds sensible though 2gb should already suffice in that case. If other the other nodes from the same datacenter are supposed to use the moved proxy, your adjustments seem sensible to me. Note that ionos9 intentionally swaps a lot. The swapping is a lot less bursty and a magnitude lower than that observed on ionos1, but we may have to apply resource limits to keep the proxy responsive. Possibly adding a bit of disk space for the squid spool (20G?) would be necessary to avoid excessive refetching. Helmut
Bug#988191: ionos-1 proxy causes lots of temporary build failures by being slow to respond
On Fri, May 07, 2021 at 12:52:52PM +0200, Helmut Grohne wrote: > The proxy running on ionos-1 is frequently overloaded and slow to > respond. This frequently results in apt or mmdebstrap failing to > download packages and makes a lot of jobs fail. For rebootstrap it seems > to roughly kill 1/3 of all jobs. As such the situation is a significant > waste of resources. as a simple fix: we could install squid on your node (ionos9?) and a another cpu for it? (and 4gb ram or such) -- cheers, Holger ⢀⣴⠾⠻⢶⣦⠀ ⣾⠁⢠⠒⠀⣿⡁ holger@(debian|reproducible-builds|layer-acht).org ⢿⡄⠘⠷⠚⠋⠀ OpenPGP: B8BF54137B09D35CF026FE9D 091AB856069AAA1C ⠈⠳⣄ Stop saying that we are all in the same boat. We’re all in the same storm. But we’re not all in the same boat. signature.asc Description: PGP signature
Bug#988191: ionos-1 proxy causes lots of temporary build failures by being slow to respond
Package: jenkins.debian.org Severity: important The proxy running on ionos-1 is frequently overloaded and slow to respond. This frequently results in apt or mmdebstrap failing to download packages and makes a lot of jobs fail. For rebootstrap it seems to roughly kill 1/3 of all jobs. As such the situation is a significant waste of resources. Looking into munin (thanks for providing that!), there seems to be a good correlation between significant iowait https://jenkins.debian.net/munin/debian.net/ionos1-amd64.debian.net/cpu.html and the temporary failures. The system has 40G swap in use at the moment and significant swap traffic. It seems unsurprising that a squid running on such a system would not perform well and time out requests occasionally. A long term solution would be separating the proxy component from the component that causes heavy swapping (whatever that is). The simplest solution would be dedicating a separate small vm at the same datacenter to the proxying. I suppose that 1 vcpu, 1G ram and like 30G of disk space would be the minimum to support a proxy-only vm. Do you agree with the proposed solution? Helmut