3 ms·

This is so misinformed by industry propaganda. Modern cloud services are very often unavailable from a region without any apparent reasons. Or the service appea

This is so misinformed by industry propaganda. Modern cloud services are very often unavailable from a region without any apparent reasons. Or the service appears to be available but some specific feature doesn't work. Or it's available but just really slow or dropping packets. When you have a "simple" (already quite complex) BGP(TCP(HTTP)) tunnel, chances are things just work and it's easy enough to diagnose issues. Between anycast, auto-scaling, and WAFs (among others) you have added so many layers of complexity that the chance of a "random" error somewhere in the stack has dramatically increased and diagnosis is close to impossible. It used to be simple that web services are either up or down. Now it's not so obvious anymore, and that's definitely not being reflected in the 5 9's SLAs falsely promised by cloud providers. I can say with confidence most selfhosted systems i've been close to have much better uptime than modern cloud services and are much cheaper to service and maintain.

exikyut 5y ago
I agree with specific dimensions of this sentiment. Firstly, I think that the GP comment ever so slightly misapplies the localized awareness available in smaller environments to the cloud. Yes, there are tons more engineers to look at stuff, but those engineers are a) already bogged down keeping up with internal infra, politics and bureaucratic machinery, and b) quite some distance further away from actual errors occurring on the ground because everything's aggregated to the hilt so the stats remain comprehensible at the higher scale. I also agree that the newer stuff that is less mature has an order of magnitude more intrinsic leaky abstractions than old designs, which I do think were more hygienic and espoused more effective separation of concerns than what is used today. Also, not to nitpick, but the "classical" canonical definition would probably look closer to BGP(TCP(TLSv1.3(HTTP/1.1))), with the modern equivalent being BGP(UDP(HTTP2)) and the future being BGP(UDP(QUIC)). I do agree that the rapid consolidation of HTTP and TLS, without wide-scale awareness and slow, methodological development of general introspection tooling, does make things net worse in general. I suspect infra will still be using HTTP/1.1 for a long time into the future until this materially changes.

exikyut 5y ago

I agree with specific dimensions of this sentiment. Firstly, I think that the GP comment ever so slightly misapplies the localized awareness available in smaller environments to the cloud. Yes, there are tons more engineers to look at stuff, but those engineers are a) already bogged down keeping up with internal infra, politics and bureaucratic machinery, and b) quite some distance further away from actual errors occurring on the ground because everything's aggregated to the hilt so the stats remain comprehensible at the higher scale. I also agree that the newer stuff that is less mature has an order of magnitude more intrinsic leaky abstractions than old designs, which I do think were more hygienic and espoused more effective separation of concerns than what is used today. Also, not to nitpick, but the "classical" canonical definition would probably look closer to BGP(TCP(TLSv1.3(HTTP/1.1))), with the modern equivalent being BGP(UDP(HTTP2)) and the future being BGP(UDP(QUIC)). I do agree that the rapid consolidation of HTTP and TLS, without wide-scale awareness and slow, methodological development of general introspection tooling, does make things net worse in general. I suspect infra will still be using HTTP/1.1 for a long time into the future until this materially changes.