View
912
Download
0
Category
Tags:
Preview:
DESCRIPTION
The talk focuses on a new unified approach to deploying and managing modern versions of FreeBSD across a wide variety of technical and administrative circumstances: different countries, data centers, hardware, access policies, boot methods, networking, support contracts, machine roles, etc. While avoiding any popular Linux-centric CM systems, such as Puppet, Chef, and CFEngine, we achieve very low complexity by leveraging rc(8), loader(8), glabel(8) and other existing instruments, such as pkgng, to their potential as necessary. The cornerstone is keeping configuration and deployment versioned and unified — same across all cases, with no duplication of common parts and very simple specification of per-role/per-case peculiarities. The approach spans everything from installation and booting to managing third-party and custom site-specific software. The method is being actively developed and applied in production environment of a popular online music service.
Citation preview
FreeBSD UnifiedConfigurationAndrew Pantyukhin
infofarmer@FreeBSD.org
once upon a timea private cloud
petabytes of datadozens of gigabits of transfers
teraflops of processing
4 countries10 cities
13 data centers
11 service providers15 support contracts
5 SLA types
~100 machines~20 hardware configurations
~1000 hard drives
30 local networks5 network types
7 out-of-band console types
1 operating system(potentially more)
5 boot types
1 systems engineer1 network engineer
1 field engineer
initial tacticsowned -> cluster
leased -> setup & forget
briefly consideredpuppet, chef, cfengine
scripted per-node management
prioritiesextremely low ops load and
complexityextremely high performance and
flexibility
solutionunified configuration management
unified deployment
unified?exactly same root fs everywhereexactly same configs everywhere
/.git/usr/local/project/.git
/usr/home/*/.git
fully distributedflexible semi-auto master-master
syncno symlinking, copying (almost)
concentratedcomplexity
smarter specializationrole-aware configs
rolespasswd, group
aware.map
role-aware bootwho am I? what are my MACs?
MAC -> aware.map -> host -> roles
rc.conf - role-awareshell script
intricate evaluation
ntpd_enable="YES"role.www() { nginx_enable="YES"
}role.host1() { hack_enable="YES"
}
for i in $myrolesrole.$i
nginx.conf role-compatible
{ server_name www1; }{ server_name www2; }
syslog.conf role-unaware
syslog.conf - most nodessyslog.conf.collect - log collector
rc.conf-based work-around
role.logcol() {syslog_flags="-c
syslog.conf.collect" }
fstab role-unaware#empty
loader.conf, scripts
boot drive/dev/ufs/root1 - 10G/dev/ufs/root2 - 10G
boot drive/dev/gpt/swapserial - 4G/dev/ufs/serial - leftover
loader.confvfs.mountroot
falls back to NFS root
deploymentaware.map, configs adjustment
dhcp, etc
deploymentfind & partition a suitable driveuntar recent image into root1
full upgradeuntar new image into root2
pivot root1<->root2 (kernel!!)
full upgradersync? pkgng?
freebsd-update?
pkg upgradepkgng
continuous upgradegit pull
edit on any boxcommit, push
powerful conflict resolution
pretty scalable
git is awfulrsync is lacking
need more smart configs
pretty simplefool-proof
single-view cloud-wide config
Q&A
Recommended