#FORMAT IRC

<marcelo> The old model of high availability is "fault tolerance" usually hardware-based.
<marcelo> Expensive, proprietary.
<marcelo> This old model goal is to have the hardware system running 
<andres> plas
<riel> so basically, a single computer is an unreliable piece of shit  (relatively speaking) ...
<riel> ... and High Availability is the collection of methods to make the job the computer does more reliable
<riel> you can do that by better hardware structures
<riel> or by better software structures
<riel> usually a combination of both
<marcelo> the Linux model of high availability is software based.
<marcelo> Now let me explain some basic concepts of HA
<marcelo> First, its very important that we dont rely on unique hardware components in a High Availability system
<marcelo> for example, you can have two network cards connected to a network
<marcelo> In case one of the cards fail, the system tries to use the other card.
<marcelo> A hardware component that cannot fail because the whole system depends on it is called a "Single Point of Failure"
<marcelo> SPOF, to make it short. :)
<marcelo> Another important concept which must be known before we continue is "failover" 
<marcelo> Failover is the process which one machine takes over the job of another node
<riel> "machine" in this context can be anything, btw ...
<riel> if a disk fails, another disk will take over
<riel> if a machine from a cluster fails, the other machines take over the task
<riel> but to have failover, you need to have good software support
<riel> because most of the time you will be using standard computer components
<marcelo> well, this is all the "theory" needed to explain the next parts. 
<riel> so let me make a quick condensation of this introduction
<riel> 1. normal computers are not reliable enough for some people (like: internet shop), so we need a trick .. umm method ... to make the system more reliable
<riel> 2. high availability is the collection of these methods
<riel> 3. you can do high availability by using special hardware (very expensive) or by using a combination of normal hardware and software
<riel> 4. if one point in the system breaks and it makes the whole system break, that point is a single point of failure .. SPOF
<riel> 5. for high availability, you should have no SPOFs ... if one part of the system breaks, another part of the system should take over
<riel> (this is called "failover")
<riel> now I think we should explain a bit about how high availability works .. the technical side
<riel> umm wait ... sorry marcelo ;)
<marcelo> ok
<marcelo> Lets talk about the basic components of HA 
<marcelo> Or at least some of them,
<marcelo> A simple disk running a filesystem is clearly an SPOF
<marcelo> If the disk fails, every part of the system which depends on the data contained on it will stop.l
<marcelo> To avoid a disk from being a SPOF of a system, RAID can be used.
<marcelo> RAID-1, which is a feature of the Linux kernel...
<marcelo> Allows "mirroring" of all data on the RAID device to a given number of disks...
<marcelo> So, when data is written to the RAID device, its replicated between all disks which are part of the RAID1 array.
<marcelo> This way, if one disk fails, the other (or others) disks on the RAID1 array will be able to continue working
<riel> because the system has a copy of the data on each disk
<riel> and can just use the other copies of the data
<riel> this is another example of "failover" ... when one component fails, another component is used to fulfill this function
<riel> and the system administrator can replace (or reformat/reboot/...) the wrong component
<riel> this looks really simple when you don't look at it too much
<riel> much
<riel> but there is one big problem ... when do you need to do failover?
<riel> in some situations, you would have _2_ machines working at the same time and corrupting all data ... when you are not careful
<riel> think for example of 2 machines which are fileservers for the same data
<riel> at any time, one of the machines is working and the other is on standby
<riel> when the main machine fails, the standby machine takes over
<riel> ... BUT ...
<riel> what if the standby machine only _thinks_ the main machine is dead and both machines do something with the data?
<riel> which copy of the data is right, which copy of the data is wrong?
<riel> or worse ... what if _both_ copies of the data are wrong?
<riel> for this, there is a special kind of program, called a "heartbeating" program, which checks which parts of the system are alive
<riel> for Linux, one of these programs is called "heartbeat" ... marcelo and lclaudio have helped writing this program
<riel> marcelo: could you tell us some of the things "heartbeat" does?
<marcelo> sure
<marcelo> "heartbeat" is a piece of software which monitors the availability of nodes
<marcelo> it "pings" the node which it wants to monitor, and, in case this node doesnt answer the "pings", it considers it to be dead.
<marcelo> when a node is considered to be dead when can failover the services which it was running
<marcelo> the services which we takeover are previously configured in both systems.
<marcelo> Currently heartbeat works only with 2 nodes.
<marcelo> Its been used in production environments in a lot of situations...
<riel> there is one small problem, however
<riel> what if the cleaning lady takes away the network cable between the cluster nodes by accident?
<riel> and both nodes *think* they are the only one alive?
<riel> ... and both nodes start messing with the data...
<riel> unfortunately there is no way you can prevent this 100%
<riel> but you can increase the reliability by simply having multiple means of communication
<riel> say, 2 network cables and a serial cable
<riel> and this is reliable enough that the failure of 1 component still allows good communication between the nodes
<riel> so they can reliably tell if the other node is alive or not
<riel> this was the introduction to HA
<riel> now we will give some examples of HA software on Linux
<riel> and show you how they are used ...
<riel> ... <we will wait shortly until the people doing the translation to Espa�ol have caught up> ... ;)
<marcelo> Ok
<marcelo> Now lets talk about the available software for Linux
<riel> .. ok, the translators have caught up .. we can continue again ;)
<marcelo> Note that I'll be talking about the opensource software for Linux
<marcelo> As I said above, the "heartbeat" program provides monitoring and basic failover of services 
<marcelo> for two nodes only
<marcelo> As a practical example...
<marcelo> The web server at Conectiva (www.conectiva.com.br) has a standby node running heartbeat
<marcelo> In case our primary web server fails, the standby node will detect and start the apache daemon
<marcelo> making the service available again 
<marcelo> any service can be used, in theory, with heartbeat.
<riel> so if one machine breaks, everybody can still go to our website ;)
<marcelo> It only depends on the init scripts to start the service
<marcelo> So any service which has a init script can be used with heartbeat
<marcelo> arjan asked if takes over the IP address
<marcelo> There is a virtual IP address used by the service
<marcelo> which is the "virtual serverIP" 
<marcelo> which is the "virtual server" IP address. 
<marcelo> So, in our webserver case...
<marcelo> the real IP address of the first node is not used by the apache daemon
<marcelo> but the virtual IP address which will be used by the standby node in case failover happens
<marcelo> Heartbeat, however, is limited to two nodes.
<marcelo> This is a big problem for a lot of big systems.
<marcelo> SGI has ported its FailSafe HA system to Linux recently (http://oss.sgi.com/projects/failsafe)
<marcelo> FailSafe is a complete cluster manager which supports up to 16 nodes.
<marcelo> Right now its not ready for production environments
<marcelo> But thats being worked on by the Linux HA project people :)
<marcelo> SGI's FailSafe is GPL.
<riel> another type of clustering is LVS ... the Linux Virtual Server project
<riel> LVS uses a very different approach to clustering
<riel> you have 1 (maybe 2) machines that request http (www) requests
<riel> but those machines don't do anything, except send the requests to a whole bunch of machines that do the real work
<riel> so called "working nodes"
<riel> if one (or even more) of the working nodes fail, the others will do the work
<riel> and all the routers (the machines sitting at the front) do is:
<riel> 1. keep track of which working nodes are available
<riel> 2. give the http requests to the working nodes
<riel> the kernel needs a special TCP/IP patch and a set of usermode utilities for this to work
<riel> RedHat's "piranha" tool is a configuration tool for LVS, that people can use to setup LVS clusters in a more easy way
<riel> in Conectiva, we are also working on a very nice HA project
<riel> the project marcelo and Olive are working on is called "drbd"
<riel> the distributed redundant block device
<riel> this is almost the same as RAID1, only over the network
<riel> to go back to RAID1 (mirroring) ... RAID1 is using 2 (or more) disks to store your data
<riel> with one copy of the data on every disk
<riel> drdb extends this idea to use disks on different machines on the network
<riel> so if one disk (on one machine) fails, the other machines still have the data
<riel> and if one complete machine fails, the data is on another machine ... and the system as a whole continues to run
<riel> if you use this together with ext3 or reiserfs, the machine that is still running can very quickly take over the filesystem that it has copied to its own disk
<riel> and your programs can continue to run
<riel> (with ext2, you would have to do an fsck first, which can take a long time)
<riel> this can be used for fileservers, databases, webservers, ...
<riel> everything where you need the very latest data to work
<riel> ...
<riel> this is the end of our part of the lecture, if you have any questions, you can ask them and we will try to give you a good answer ;)

<> See also http://www.linux-ha.org/
----
CategoryDocs