the Unix OS

• 1970s — comprehensive platform complete with apps
• 1990s — nice platform to run third-party apps
• 2010s — thin sub-middleware layer

operating system

• meeting point
• implementation repo
• academic vehicle

good ol' times

• data
• server OS (Unix)
• user

wasted effort ...

• server OS (FreeBSD, etc)
• server middleware (hadoop, etc)
• server app (Facebook PHP, etc)

... wasted effort

• client OS (PC-BSD, etc)
• client middleware (browsers, etc)
• client app (Facebook JS, etc)

Chrome OS

• consolidating client side
• only in concept

large systems

• Google, Facebook, Twitter, ...
• a small company cluster
• any two machines, e.g. client-server

TCP/IP is evil

• never meant for intrasystem comms
• utterly tenacious
• preempted real distributed systems engineering

the cloud

• does not tolerate inconveniences in the OS
• happy to move any piece of logic up
• happy to forget about the OS altogether (bare metal Java!)

should BSD care?

• wasted effort
• highly fragmented distributed engineering effort
• extremely poor quality of net result

a way forward

• look for obvious things to move down into the OS
• provide useful conveniences
• set the trend and define guidelines

the usual impediment

• the larger the system —
• — the more the experience and ideas
• — the less time to speak and contribute

role aware rc

• one /etc for all machines
• infectious boot: pxe, hdd, usb
• / in git

rc.conf.d

case `hostname -s` in
z[0-9]*|h22|nemea|h4|lw[1-2])
	nginx_enable="YES"
	nginx_profiles="default zdata"
	;;
esac
    

machine generated data

• don't try to print human readable
• don't try to parse human readable
• expect multiple hosts
• cron(8), periodic(8)
• apache-style logs, logs in SQL

task contracts

• access to data
• resource allocation
• real-life metrics

task baby-sitting

• launchd, fcs
• crontab: pgrep || start
• monitd, etc

AAA

• LDAP/GSS
• master.passwd(5), group(5) in git(1)

SW fault tolerance

• watchdogd(8)
• timelimit(1)
• custom wrappers
• rctl(8), capsicum

HW fault tolerance

• ipmitool
• disk error counters
• memory testing

massive execution

• zz -grole1,role2 -Xrole3,role4 uptime

• distributed uptime, df, w, top, ...

storage

• RAID, HAST hard to justify on a scale
• ZFS limited to one node
• no useful distributed fs
• UFS2 + nginx + custom dispatcher

the good bits

• ssh hpn
• rctl
• pkgng
• SU+J
• bsnmpd