IMHO, code should be commented nevertheless, saves time reading. Ie. doc blocks to explain what for a method is for. Later on this saves time when you can just generate code documentation, and can refer to. Larger the project is, the more important this is. Tho this might be a bit small project for such level of documentation, nevertheless a short comment here and there is a good idea.

If you check some major project commentation rules or philosophies, like Linux Kernel, they suggest that one comment per 10 lines is a good level, level which you should target.

As for the queries, if you can avoid on the fly created tables, that's better. Indeed, i did not even check how my queries did, as i was expecting them to work as expected.

I spent sometime working on the 3rd one, and i achieved in under an hour to make the work amount less, less subqueries and on the fly created tables. No temporarys, no filesorts etc. but on the testing machine i'm using (Athlon XP 1900+ so REALLY slow), i couldn't verify results as times went into margin of error easily, and might just be parsing the query & deciding indices etc. takes that damn long on that slow CPU, with my tiny dataset at the time. It was faster without mHash rate, slower with mHash rate added (comparing oranges to oranges, so i compared to yours without mhash rate, and with it). Now i have larger dataset so when i got time i'll check it out again.
Tho, i might begin from scratch as approach is wrong.

Now i got 50k rows in the table, but that's still too small, i want to get to atleast 150k+ to be able to properly go through it, despite handling on this slow machine.

Why are you using echo_html function in the view templates? Have you looked into smarty btw? It's very easy to implement, and makes views damn much simpler to view (just view as html and most highlighters are perfect for it). And those saying it forces bad code behavior: It's all about how you use it, it doesn't force anything, it simply compiles the views you ask for, and actually provides an EXCELLENT "isolation layer" for different layers of code, and eases debugging quite a bit saving tons of time. Overhead is extremely minimal too, and the first time you want to even think about it is when you are hitting high requests per sec.

Also, i noticed some core functionality issues with this: phoenix started crashing now and then, about once a day per instance. Using BTCGuild, Bitcoin.cz and Deepbit, got 2 workers put for it, another works with 2 discreet GPUs, another with just 1. Sometimes it will just not pass new work, and sometimes it pauses work for a while. Is it only me? I'm using LinuxCoin v0.2a and it's accompanying phoenix.

EDIT: Still getting quite fast load on dashboard with this dataset. Infact most of the time in the ms range. At slowest nearly a second. What is the point where people are starting to experience serious performance degradation, all the time?