The stated capacity of 10^6 submissions per hour seems utterly ludicrous - the sums shown in this article imply that the suppliers are incompetent. It will be interesting to see how the fallout from this is played.
Considering it was designed and hosted by IBM who are trying to recreate themselves as a primarily cloud PaaS company (see IBM BlueMix), I'd say it was a matter of IBM's engineers poorly configuring the scalability of the deployment, which I think also reflects poorly on their ability to service the cloud hosting market.
ABS shouldn't have outsourced the entire project to IBM, they could have just developed the application in-house and deployed it to an autoscaling CE instance group on Google Cloud, they probably would have saved $7-8 million paid to IBM while also avoiding this catastrophic failure.
I don't think the cloud and/or auto-scaling would help here. Most likely this is due to poor architecture choices.
Also, from what I can see, gov't departments normally outsource for projects and only ever retain staff for maintenance activities (i.e. keep the lights running). It's probably due to the way budgeting is done.
It would almost certainly not be legal for them to host it on any servers owned by a private company. It's a national census, and they specifically are planning to store personally identifiable information for 4 years (imagine that I'm screaming the last sentence). Basically there's no way on earth it would be permitted for a government department to host the census collection on someone else's servers.
