LiveBlog: Death of a Web Server

The first Velocity 2009 session for me today. Semi-useful, but seemed very .NET focused. I didn’t take a lot of notes around it for that reason.

For me, most of his prezzo reduced to:

• Make sure your instrumentation has a light touch
• Know what you’re caching
• Be sure it’s getting hit (don’t cache singleton queries)
• Be sure your TTL is well-set (short, perhaps sliding)

Otherwise, most of the interesting bits came from Q&A:

Everyone has a favorite load generator; many love the appliances. They “make a webserver cry”. Mercury LoadRunner seemed like a crowdpleaser.

Log playback can be challenging as you frequently don’t have all the post data, response data, etc. Hard to know how “real” it is. An alternative is to inject a transparent proxy on the front end and capture everything both ways for short periods.

One fellow in the audience is using “Siege” and EC2 instances. [I think this is the Siege he meant. I wonder why no one mentions using Tsung, formerly Tsunami, and EC2 … here’s the one decent ref I could find]

Realisitic tests: be sure what you’re using to load it resembles your production load or it’s not very useful.

All rely on scripted sessions; there’s a gap in converting log data into that script.