4 ms·
It's a bizarre article. On one hand I understand their frustration with OpenStack. On the other hand, if they expected to pretty much build their own AWS withou
by vidarh 11y ago
It's a bizarre article. On one hand I understand their frustration with OpenStack. On the other hand, if they expected to pretty much build their own AWS without any warning bells going off all over the place I find it very hard to be sympathetic - they're not remotely large enough for that. If they'd set their goals lower they could have had a very capable system fairly easily.
That they're having problems getting load balancing working properly is a real sign that their ops team just didn't know what they're doing.
- rrmm 11y agoif the choice is "let's deploy software on AWS," vs "let's build an AWS work-a-like and then deploy software on it," it'll always be easier to use AWS. Managing physical servers is a pain. I've been in a similar situation in a small research group, but AWS doesn't fly for privacy and regulatory reasons. Managing hardware and infrastructure is a huge amount of overhead that doesn't directly relate to the core mission.
- vidarh 11y ago> if the choice is "let's deploy software on AWS," vs "let's build an AWS work-a-like and then deploy software on it," it'll always be easier to use AWS Yes, but that's a false equivalence. You don't need an AWS work-a-like unless you plan on competing with AWS. So much of AWS is dedicated to doing things you don't need to do in most single tenant systems. But if managing hardware and infrastructure is a huge amount of overhead, what in the world are you doing? If it's more than sliding a server into a rack, hooking up network and power, and leaving it to itself until/unless there are hardware failures or SMART alerts, then you've missed out on automation steps somewhere. For server environments I set up, I spend perhaps on average 30m per server in a datacentre per year, including travel. On top of that we estimate on average perhaps 10m of remote hands per server per year.
- _Codemonkeyism 11y ago"they're not remotely large enough for that." Second that. Maybe there are only ten companies in the world large enough to have positive ROI to build an internal AWS, like Google, eBay, Facebook, ... who don't want to depend on Amazon.
- leoc 11y agoSounds like there's a space for a Visa- or Underwriters' Laboratories-like cloud company owned by and run for a consortium of medium-sized publishers.
- zurn 11y agoOn the other hand this is from an Amazon marketing event and they are probably combining hindsight with colored glasses. Ultimately you can always diagnose failures as "they didn't know what they were doing". But usually there are more nuanced explanations for wrong directions. For example it might be that OpenStack seemed to be heading in a better direction in 2012. Maybe it was reasoned that it's worth it to hedge against AWS lock-in and you can always easily fall back to Amazon, or even mix and match AWS and OpenStack. Etc. If OpenStack worked well I don't see any inherent reason why you'd have to be bigger than Guardian to use it for autoscaling and load balancing.
- vidarh 11y agoHedging against AWS, I would understand. Building out your own virtualised setup I would understand. Even making use of portions of OpenStack. My issue with their description is that they seems to have tried to essentially emulate what is a very complex platform with the help of OpenStack, which is a very complex platform in a situation where they almost certainly didn't actually need more than a fraction of the functionality. > If OpenStack worked well I don't see any inherent reason why you'd have to be bigger than Guardian to use it for autoscaling and load balancing. On the other hand, if that was all they were using it for, OpenStack is massive overkill and there were plenty of simpler options.
- slyall 11y agoWe I think the problem was that "setting their goals lower" would have not given them what they wanted (self-service, autoscale, etc). If they had just built VMware say then they would have not been able to keep up with companies that did go with AWS. As for load balancing I'd believe they couldn't get it working within Openstack. We're trying to deploy OpenStack at work and the the number of things that don't quite work are a pain.
- vidarh 11y agoSelf service and auto-scale is almost trivial. In a typical setup I'll be sliding a new server into a rack, plugging in power and network, powering it up and have it auto-boot to a suitable Linux distro. Then I'll leave the data centre. Trigger a provisioning script, and install libvirtd+kvm, or a suitable container technology (e.g. in the timeframe they were working on this, openvz or lxc were both perfectly fine alternatives for a single-tenant system). Now you need a scheduling system: Split each server into N units of type M, add it to an inventory, and add a small API to reserve or release one of them and trigger a kvm/openvz etc. install. Both components here (bootstrap script + resource allocation) can be small. The most recent bootstrap script I wrote was 60 lines. For resource allocation, you can go insanely complex (in which case you should probably consider AWS or installing OpenStack or similar after all), or very, very simple. I've run off the shelf solutions for it, and I've written my own in a couple of hundred lines - it very much depends on your needs for a specific platform. What almost nobody needs, is the complexity of an AWS-level platform when setting up a single-tenant cloud where you know and control the requirements. Public cloud providers need the complexity because they need to be able to support a vast number of different customer requirements. When you don't have that issue, a ton of complexity falls away. > then they would have not been able to keep up with companies that did go with AWS. They're a media company, not a cloud hosting provider. They need to be able to support the functionality they need internally, nothing more. > As for load balancing I'd believe they couldn't get it working within Openstack. I'm not surprised. Which raises the question of why the picked Openstack in the first place. Clearly it was not evaluated very well.