Velocity 2013 - Tuesday

gistfile1.txt

Keynote
-------

All systems will become Safety critical systems
Rasmussen system model
ctlab.org
prerender
httpwatch <- mobile
Network / Render / Compute
Gather / Insight / Action
goo.gl/UAZQvl <- Chrome Dev tools - Memory Stuff


Latency
-------

Anything that keeps state - Affects latency
Should be non-sticky
Learn openstack
Learn Hadoop


Alerts
------

Signal to noise


Going beyond onLoad
-------------------


SpeedIndex = "The average time at which visible parts of the page are displayed" (milliseconds)
aft final event
Page (visual) completeness

Microsoft PPT (Page Phase Time) - Spilts page load into phases

"Done" = User ready


Google Cloud Platform
---------------------

[Marketing monster]

App engine looks nice and straightforward.

Easy to spin up instances with API call
Billed by minute

Compute engine:
Very like AWS

App engine: (No hands scaling)
A set of APIs to run on. You pick your language.

Has a "Git push to deploy" style like Heroku

[2K USD free starter pack]


Production Incident System
--------------------------

Capacity metrics - CPU / Memory etc
Discrete metrics - No. of users / No. of errors
Continuous metrics - Response times, etc (But, there might be a lot of noise)

Connection pool:

Normal: Connection pool use (at 90%, 100% ?)
Better: Connection aquistion time (Pre-warning, don't need to wake someone)

Normal: CPU usage
Better: CPU and load average combination
Even better: App response times


Don't use std. deviation if the data doesn't follow that pattern (bi-modal, etc)


Braintree payments infra (Paul Gross) <- Great guy
------------------------

MySQL -> Posgres

Python/Tornado used in Broxy (Apache to Redis middleware)

Automate everything.
Add safeguards
Check everything
Nagios and PagerDuty (Phone calls)

Learn Capistrano and Puppet

Environment name in prompt on console.
Graphite <- Vis