I'm pretty sure you're not going to go with my solution, but having just produced some live stats, I approached it a bit differently. I put the raw data into a database, and then, in another process, scanned the database for NULL fields, and calculated them. Then I just used some SQL to query counts, grouped by interesting fields as appropriate. Right before the query, I would DELETE FROM LOGS WHERE STAMP < CURRENT TIMESTAMP - 7 DAYS, to keep the logs circular.

This meant that I figured out what was important on each line only once, and then pulled all the numbers from the db as many times as needed (~168 times - once an hour for seven days before that line got deleted).

Performance-wise, mojotoad's approach (which seemed to parse the entire set of data each time) took about 3 seconds to produce, whereas mine is steadily hovering between 0.3 and 0.7 seconds - most of the time, around 0.4 seconds. Basically, an order of magnitude greater speed (most of the hard work is done outside the generation). Mind you, I may also be using an order of magnitude more memory, too, hard to tell ;-) though I do have more stats than he did.