It helps to divide this question into several parts.
1) When should one start using the HTML5 doctype?
The HTML5 doctype causes browsers to render pages in Full Standards mode. Unless you're currently using a strict doctype, pages may display differently. So you'll need to weigh up the amount of rework required if you're converting a existing site, and you'll need to understand the difference in rendering behaviour. Secondly, the doctype will cause the w3c validator to select HTML5 as the default language to validate against. Given that HTML5 is still in draft, this means aiming at a moving target. Is validity important to you? Some features that are valid in HTML4 are not currently valid in HTML5. Should you change your markup now to remove those features, given that HTML5 may change to make those features valid again? Provided you're comfortable with this, you can start using the HTML5 doctype now.
On the other hand, on it's own, what are you gaining? If you want browsers to render in Full Standards mode, you can use a strict doctype from HTML 4.01 or XHTML 1.x. Do you need HTML5 validation given that it will pass things that may not at present be handled consistently across browsers? Be aware that any HTML5 feature that currently works in browsers, works regardless of whether you use the HTML5 doctype or not.
2) When should one start using features that are currently interoperably implemented in the major browsers, but standardised for the first time in HTML5?
You can use these now, but try to make sure that your pages are properly accessible. Canvas, for example, does not provide good support for non-sighted users.
3) When should one start using features that are current implemented in some browsers, and will be standardised by HTML5?
This depends on your target users and delivery schedules. Will your users be using a browser that supports these features by the time your site goes live? What happens if they don't? Does the page fall back gracefully? Do you care if some users can't use your site?
4) When should one start using features of HTML5 that are not implemented at all yet?
Why would you want to do that? It might be reasonable for training purposes or investigating fall back behaviour, but HTML5 is still in draft form, and it's quite possible that some features will never be implemented, or will be implemented in a totally different way to that currently described, so you could waste a lot of effort.
Summarising, what matters is not so much what's in the HTML5 draft, as what is practically usable today. The draft describes a range of features, some of which have been in browsers for years, some which are new but have support or good fall back behaviour, and some which are not currently and may never be usable.
For me, I'll be holding off until I need an HTML5 feature, and I know that that feature is implemented in a consistent and stable way across the major browsers. That will probably happen some time during HTML5's Last Call phase, or possibly during the Candidate Recommendation phase.