prb.io

» Ahhh... That's Better #

I cut over to perpubplat about a week ago, so it's worth a quick check to see if I met my goal of getting a few more nines.

On the left, we have an every-30-minutes response time chart for the Typo version of this site, and on the right, the perpubplat version. (The charts are from the free monitoring service from mon.itor.us.) The Typo configuration used Apache2 with mod_fcgid (see my earlier post on the subject) configured for seven processes with a maximum 60-minute lifetime; the perpubplat configuration uses Apache2 with mod_fastcgi configured for a single process that runs on 50 lightweight Haskell threads internally. (N.B.: The graphs use different scales on the y-axis.)

The graph hints that things are better, but some quick text processing on server logs make the difference more explicit. Here's an analysis of the response codes from the most recent log file for perpubplat:

$ head -n 1 multifarious-combined.log | awk '{print $4}'
[07/Jan/2008:00:32:50
$ ^head^tail
[11/Jan/2008:15:03:50
$ awk '{print $9}' multifarious-combined.log | sort | uniq -c | sort
      1 400
      2 206
     73 304
   1572 302
   2976 404
   5574 301
  14772 200

No 500s. (Most of the 404's are comment spammers trying to hit old URLs for comments.) Here's the same analysis for Typo from a week back in December of last year (2007):

$ zcat mult.ifario.us-access.log.3.gz | head -n 1 | awk '{print $4}'
[10/Dec/2007:19:49:06
$ ^head^tail
[19/Dec/2007:14:28:17
$ zcat mult.ifario.us-access.log.3.gz | awk '{print $9}' | sort | uniq -c | sort
      5 206
      5 400
     23 503
     30 302
    237 404
   1098 301
   2259 500
   9319 304
  20885 200

A change from one-ish nines (-log₁₀ (33861 - 2282)/33861 ~ 1.17) to 100% uptime is a positive change, and the CPU trace for the virtual server suggests that the perpubplat configuration uses a tiny fraction of the machine resources of the Typo configuration.