High Availability (HA) Explained - second edition

Maciej Lasyk, High Availability Explained
Maciej Lasyk
11. Sesja Linuksowa
Wrocław, 2014-04-05
1/14
High Availability Explained

“Anything that can go wrong, will go wrong”
Murphy's law
2/14

An electrical explosion and fire Saturday at a Houston data
center operated by The Planet has taken the entire facility offline.
The company claimed power to the facility was interrupted when a
transformer exploded. Official reports that three walls were blown
down causing a fire.
Murphy's law
2/14

Three walls of the electrical equipment room on the first floor
blew several feet from their original position, and the underground
cabling that powers the first floor of H1 was destroyed.
An electrical explosion and fire Saturday at a Houston data
center operated by The Planet has taken the entire facility offline.
The company claimed power to the facility was interrupted when a
transformer exploded. Official reports that three walls were blown
down causing a fire.
Murphy's law
2/14

High Availability is in the eye of the beholder
3/14

CEO: we don't loose sales
3/14

Sales: we can extend our offer basing on HA level
3/14

Accounts managers: we don't upset our customers (that often)
3/14

Developers: we can be proud – our services are working ;)
3/14

System engineers: we can sleep well (and fsck, we love to!)
3/14

System engineers: we can sleep well (and fsck, we love to!)
Technical support: no calls? Back to WoW then.. ;)
3/14

So how many 9's?
4/14

So how many 9's?
Monthly: 1 hour of outage means 100% - 0.13888 ~= 99.86112 of availability
4/14

So how many 9's?
Yearly: 1 hour of outage means 100% - 0.01142 ~= 99.98858 of availability
4/14

So how many 9's?
Yearly: 1 hour of outage means 100% - 0.01142 ~= 99.98858 of availability
Availability Downtime (year) Downtime (month)
90% (“one nine”) 36.5 days 72 hours
95% 18.25 days 36 hours
97% 10.96 days 21.6 hours
98% 7.30 days 14.4 hours
99% (“two nines”) 3.65 days 7.2 hours
99.5% 1.83 days 3.6 hours
99.8% 17.52 hours 86.23 minutes
99.9% (“three nines”) 4.38 hours 21.56 minutes
99.99 (“four nines”) 52.56 minutes 4.32 minutes
99.999 (“five nines”) 5.26 minutes 25.9 seconds
4/14

So how many 9's?
https://jazz.net/wiki/bin/view/Deployment/HighAvailability
4/14

HA terminology
RPO: Recovery Point Objective; how much data can we loose?
5/14

HA terminology
RTO: Recovery Time Objective; how long does it take to recover?
5/14

HA terminology
RTO: Recovery Time Objective; how long does it take to recover?
MTBF: Mean-Times-Between-Failures; time between failures
(density fnc -> reliability fnc)
https://en.wikipedia.org/wiki/Mean_time_between_failures
5/14

HA terminology
SLA: Service Level Agreement;
formal definitions (customer <-> provider)
5/14

HA terminology
SLA: Service Level Agreement;
formal definitions (customer <-> provider)
OLA: Operational Level Agreement; definitions within organization;
help us keeping provided SLAs
5/14

SLAs..
So what is written in SLAs?
90% 36.5 days 72 hours
95% 18.25 days 36 hours
97% 10.96 days 21.6 hours
98% 7.30 days 14.4 hours
99% 3.65 days 7.2 hours
99.5% 1.83 days 3.6 hours
99.8% 17.52 hours 86.23 minutes
99.9% 4.38 hours 21.56 minutes
99.99 52.56 minutes 4.32 minutes
99.999 5.26 minutes 25.9 seconds
5/14

SLAs..
So what is written in SLAs?
90% 36.5 days 72 hours
95% 18.25 days 36 hours
97% 10.96 days 21.6 hours
98% 7.30 days 14.4 hours
99% 3.65 days 7.2 hours
99.5% (EC2, EBS) 1.83 days 3.6 hours
99.8% 17.52 hours 86.23 minutes
99.9% (SoftLayer, IBM) 4.38 hours 21.56 minutes
99.99 (OVH ded. cloud) 52.56 minutes 4.32 minutes
99.999 5.26 minutes 25.9 seconds
http://aws.amazon.com/ec2/sla/
http://www.softlayer.com/about/service-level-agreement
https://www.ovh.com/us/dedicated-cloud/security-and-sla.xml
5/14

SLAs..
Availability mentioned in SLAs are only goals of service provider
Usually when it's not met than company pays off the fees
5/14

SLAs..
5/14
Hetzner?
“We guarantee an annual average of 99% network availability”
“For indirect damages and loss of profits, we are liable only in
cases of intentional or gross negligence. In this case we are
liable only for the contract-typical predictable damage, a
maximum of 100% of the annually fee.”
http://www.hetzner.de/en/hosting/legal/agb

SLAs..
5/14
Leaseweb?
(yup - megaupload)
- no %s
- best effort
- $$$ for faster response
times
http://www.leaseweb.com/en/support/all-about/sla

How deep is this hole?
app layer (core, db, cache)
data storage
operating system
hardware
networking
location
So we would like to achieve 99,9999% which is about 30s of downtime per year
6/14

data storage
operating system
hardware
networking
location
Even Proof of Concept is very hard to provide: 2.5s of downtime per layer monthly!
6/14

data storage
operating system
hardware
networking
location
AOL has 99,999%!
http://highscalability.com/blog/2014/2/17/how-the-aolcom-architecture-evolved-to-99999-availability-8.html
6/14

Load-balancing and failover
LB:
http://www.netdigix.com/linux-loadbalancing.php
7/14

Load-balancing and failover
Failover:
http://www.simplefailover.com/
7/14

LB – 4th
layer or 7th
?
4th
layer:
- high performance
- just do the LB work!
- reliable
- scalable
7th
layer:
- low cost
- good for quickfixes / patches
- not that scalable
- low performance
- complex codebase
- custom code for protocols
- cookies? what about memcache..
8/14

Disaster Recovery
9/14

Disaster Recovery
http://disasterrecovery.starwindsoftware.com/planning-disaster-recovery-for-virtualized-environments
9/14

Disaster Recovery
Hot site: active synchronization, could be serving services. Cost can be high
Warm site: periodical synchronization, DR tests needed. Low costs
Cold site: Nothing here – just echo and some place to spin services; nightmare
9/14

Disaster Recovery
Hot site: active synchronization, could be serving services. Cost can be high
So maybe use “hot” as production and global LB? win-win scenario
9/14

Planning for failure
10/14

Everything starts here - DNS:
- keep TTLs low (300s). Can't make under 60min? That's bad!
- check SLA of DNS servers (dnsmadeeasy.com history)
- what do you know about DNSes?
- zero downtime here is a must!
- this can be achieved with complicated network abracadabra
- remember what 99.9999% means?
- round robin is a load – balancer but without failover!
- GSLB – killed by OS/browser/srvs cache'ing
(GlobalServerLoadBalancing)
- GlobalIP (SoftLayer etc) – workaround for GSLB via routing
10/14

E-mail servers:
- it's simple as MX records (delivering)
- it's almost simple as complicated system of SMTP servers (sending)
- it's not that simple when IMAP locking over DFS (reading)
5 gmail-smtp-in.l.google.com.
10 alt1.gmail-smtp-in.l.google.com.
When MXing – watch the spam!
10/14

WEB servers:
- it's simple as some frontend loadbalancer
- did you really stick user session to particular server? Memcache!
- LB balancing algorithm
- how many LBs?
- what if LB goes down?
10/14

DB servers:
- it's.. not that simple
- replication (master – master? App should be aware..)
- replication ring? Complicated, works, but in case of failure...
- let's talk about MySQL:
- NoSPOF solution: MySQL cluster
- MySQL Galera cluster – synch, active-active multi-master
- master – master – simply works
- MySQL fabric – HA + sharding; use with large farms
- Failover? Matsunobu Yoshinori mysql-master-ha
- MySQL utilities (http://www.clusterdb.com/mysql/mysql-utilities-webinar-qa-replay-now-available/)
10/14

DB servers:
Matsunobu Yoshinori mysql-master-ha
https://code.google.com/p/mysql-master-ha/
“I have heard a couple of cases where AWS users use MHA
instead of RDS because RDS takes much longer downtime
on slave promotion (more than 5 minutes usually, because
standby database is not running).
I'm surprised AWS users care about a few minutes of downtime..”
10/14

Caching servers:
- this is cache for God's sake – why would we use HA here?
- just use proper architecture like... redundancy.
10/14

Caching servers:
Load – balancers:
- remember about failovering IP addresses!
10/14

Caching servers:
Load – balancers:
- remember about failovering IP addresses!
Storage – DFSes:
- GlusterFS – we'll see it in action in a minute
- NFS? Could be – over some SAN / NAS (high cost solution)
- CephFS – just like GlusterFS – it's great and does the work
- DRBD – lower level, does the work on block – device layer – slow...
10/14

GlusterFS:
- low cost (could be..)
- distributed volumes
- replicated volumes
- striped volumes
- and...
- distributed – striped volumes
- distributed – replicated volumes
- distributed – striped – replicated volumes
- sound good? :)
10/14

GlusterFS: replicated volumes vs Geo-replication
- replicated:
- mirrors data
- provides HA
- synch – replication
- Geo-replication:
- mirrors data across geo – distributed clusters
- ensures backing up data for DR
- asynch – replica (periodic checks)
10/14

HA for virtualization solutions?
- it's really complicated, like...
11/14

- but it could be done simpler...
11/14

- it could be done simpler with containers!
- containers are very light
- deploy time from bare is tiny
- management is very easy
- resources throttling via cgroups
11/14

Tools
The most important tool would be the conclusion from the picture below:
12/14

Maciej Lasyk, High Availability Explained 12/14
Tools
The most important tool would be the conclusion from the picture below:

Tools
- DNS: roundrobin, GSLB, low ttls, globalIP
12/14

Tools
- Load-Balancers (l7, stateless services)): HaProxy, Pound, Nginx
12/14

Tools
- Failover (statefull services):
- IP: KeepAlived + sysctl
12/14

Tools
- Managing: pacemaker (manager) + corosync (message'ing)
12/14

Tools
- Managing: pacemaker (manager) + corosync (message'ing)
- (almost) All-In-One: Linux Virtual Server
12/14

Turn on HA thinking!
Main goal of HA? Improve user experience!
- keep the app fully functional
- keep the app resistant and tolerant to faults
- provide method for a successful audit
- sleep well (anyone awake?) ;)
13/14

Maciej Lasyk
11. Sesja Linuksowa
2014-04-05, Wrocław
http://maciek.lasyk.info/sysop
maciek@lasyk.info
@docent-net
High Availability Explained
Thank you :)
14/14

High Availability (HA) Explained - second edition

More Related Content

Viewers also liked (7)

Similar to High Availability (HA) Explained - second edition (20)

More from Maciej Lasyk (20)

Recently uploaded (20)

High Availability (HA) Explained - second edition