Managing Computers with Automation

Using virtualization to provide "HA at wholesale"

Traditionally, the way people have implemented high availability is by using a high-availability management package like Linux-HA[1], then configure it in detail for each application, file system mount, IP address and so on.  This traditional method works quite well, but can be a bit labor intensive - particularly when using custom or uncommon applications.  You may have to understand the structure of your applications, write some resource agents[2], debug them, and test them in detail.  In addition, every time you change your mount structure, or other details you've told your HA system, you have to be sure and update your HA configuration to match - or it might not fail over correctly the next time.

When you have good resource agents, your HA system will also recover from application failures - by restarting applications that have failed.  This is a good thing.  On the other hand, this is enough work that virtually no one runs all their applications in an HA configuration.  It's just too much work for most applications.  I call this traditional boutique-like method "HA at retail".  It works well, but it is a little costly to set up and maintain all the details just so.

With virtualization, another approach is possible, and (big surprise), I call it "HA at wholesale".  In this paradigm, instead of needing to write scripts for each type of application, you just have one resource agent - one for managing a virtual machine.  You also don't need to know the structure of the applications - the OS still starts them in whatever way it has been starting them all along.  Wow, this sounds good - less work, fewer chances for errors!  As expected, there is still no such thing as a free lunch here - you do wind up with some disadvantages.

For example, you can no longer easily detect the failure of an application.  In addition, if an application fails, the only thing you can do about it is reboot the entire virtual machine.   Inevitably, this takes longer than just restarting the failed application.

So, HA at wholesale has these properties:

• Simple enough that you can implement it for every machine
• Works well for hardware failures
• When coupled with hardware predictive failure analysis[3] and smart HA software, outages can sometimes be completely avoided.
• Can't easily detect or recover from application failures
• The only thing you can do about any failure is reboot the virtual machine

HA at retail has these properties:

• It is complex enough that you need to limit how broadly you apply it in your environment
• Works well for hardware failures
• It can easily detect and recover from application failures
• Individual applications can easily be restarted - and don't require a reboot

HA software like Linux-HA[1] can manage either type of environment.  In an ideal world, one would like to be able to do both in the same software infrastructure.


[1] http://linux-ha.org/
[2] http://linux-ha.org/ResourceAgent
[3] http://www-05.ibm.com/hu/termekismertetok/xseries/dn/pfa.pdf

Watch that basket!

The computing industry has lots of trends, numerous buzzwords, and a number of hot topics.  Sometimes these are in conflict with each other, or at least start out that way...  But, in the end, there are often good ways to harmonize all these various things.

Let's wander into virtual machine territory again today.  If you have gone to the trouble to create a bunch of virtual machines, the chances are you hope to do a little server consolidation - because when that's properly done it can save you some money.

This sounds good, and indeed has lots of good things going for it.  It's buzzword compliant, it's green, it saves you green (money).  What's not to like?

To see what you might not like if this is all you do, let's take an example to make it obvious...

If you put all your virtual machines on one physical server, then if that server fails, you lose all your virtual machines.  If you put ten virtual machines on one server, then the impact of that server crashing is roughly ten times as great as if a single server crashed.    If you work at it, you might be able to consolidate the ten most critical virtual machines onto a single server - and bring your entire data center to a halt with just one crash - bringing a suddenly much more personal meaning to the term "shock and awe"

This is not typically what people are looking for in their data center - and could easily be one of those career-limiting mistakes that you'd like to avoid - unless you already have your next job lined up.

This falls under the "putting all your eggs into one basket" way of doing business.  This part of a famous quote - but not the whole quote.  Mark Twain said "Put all your eggs in the one basket and --- WATCH THAT BASKET"[1].  So, to follow Mark Twain's advice, we need to not just put our eggs into one basket, we also need to watch that basket.

As most of you already know, watching servers and services is most commonly done by high-availability software - something like Linux-HA[2].  A properly configured HA system will watch the basket for you, and keep the worst from happening to your basket, your servers or your career.

As you can see, doing virtualization for reasons of consolidation doesn't make much sense unless you also add management software (HA software or otherwise) to watch your basket of virtual machines for you.

In the end, it's easy to see that all these things are connected - virtualization, server consolidation, power savings (green computing), availability management, and you want to manage them all.

[1] http://herbison.com/herbison/broken_eggs_watch.html
[2] http://linux-ha.org/

Virtual machine snapshots considered (nearly) worthless...

With apologies to Edgar Dijkstra...

Usually when people talk about virtual machine snapshotting, they include with it snapshotting both the server and any filesystems its directly connected to.  Although this is more complex than just snapshotting the virtual machine, it isn't that hard.

This works in some very narrow technical sense for some few cases, but it involves loss of data in every case.  If you take a checkpoint every 30 minutes (or every 5 or whatever), then all the updates made during that period of time, are lost when you restore this snapshot and its storage to a consistent (but old) state.  This means that all the checks you deposited during that time, or all the bonuses your boss put you in for during that time, or the books you ordered, or whatever, are lost.  Lost to the point that they probably have to be restored manually - to the tune of great customer dissatisfaction.

In addition, if this application has connections as a client, or as a server to other servers or clients, then although the application and its immediately mounted storage are now consistent, but unless you do simultaneous snapshots between this virtual machines and all the world it is connecting with (some of which may be outside your enterprise), and then restore your entire world to this older state, then there are likely to be many client/server connections which will no longer work - because the client and server are in mutually inconsistent states.

The worst case of this is if you have a Service Oriented Architecture, where any given server is only a small part of the overall service - every service has connections to something else all the time, and to make matters worse, the clients and/or servers are often outside your own enterprise.

And, of course, don't forget that you lost transactions in the process too.  So, a reboot interval of 1 to 3 minutes sounds really good by comparision.  Because all you'll lose in that case is transactions that were not yet committed - which are many fewer than the number of transactions lost by backing up to the previous checkpoint.

As an example of a common special case where this obviously doesn't work, imagine that the server in question is a file server.  So, you restore the virtual machine and all its storage (the file server) to some older state.  Now all the connected applications which _thought_ they had committed some particular piece of work (a spreadsheet, a database transaction) - just had all that work undone.  And, depending on the file server protocol and the application, bad things will happen - certainly loss of data, and probably some of the applications will create corrupt data - since updates they thought they'd made are now gone -  unbeknownst to them.  This corrupt data can cause any number of problems - inability to make further updates, cascading application crashes - these are all possibilities.

Or what if it's a client of a file server?  The file server is a separate machine (possibly virtual, possibly real, possibly an appliance).  Then you can't put its storage state back to a known state - without restoring all its clients back to the same consistent state - and if you somehow did, then _all_ of them now suffer data loss.

Not a very pretty picture.

There are some few cases where you can isolate the application from the "real world" and snapshot the whole "mini-enterprise" in a synchronous way.  Those are mostly limited to large scale scientific applications.  Given how hard it is to make them more available in any other way, this is a good thing.  But, its a practice with narrow applicability.  After reading the paragraphs above, perhaps you can see why...

Quorum Server Illustrated - updated

In two earlier posts [1] [2], I gave brief descriptions of the quorum server which seem to have left as much confusion as they provided clarity.  This post is only about the Linux-HA quorum server, and includes illustrations for clarity.

The Linux-HA Quorum API

In the Linux-HA quorum API, you can configure a number of quorum modules which are used as follows.  If a quorum module returns HAVEQUORUM, then the cluster has quorum.  If it returns NOQUORUM then the cluster does not have quorum.  If a quorum module returns QUORUMTIE, then the next quorum module in the list is consulted.  If the final module returns QUORUMTIE, then it is treated as a NOQUORUM event.

The quorum daemon is normally used in conjunction with the nomal arithmetic voting quorum module, so that it is only consulted when the number of nodes in the cluster is exactly half the number of configured modules in the system.  So, it is worth noting that the quorum server will never be consulted if a cluster has an odd number of nodes.

Quorum Server Scenarios

Below, I'll go through the basic quorum server cases so you can see how all this works in more detail - with pictures, even!

Normal Situation - Everything up

In the picture above, everything is normal.  The quorum server is up, and both sites are also up.  Because the cluster has all its nodes up, the quorum server is irrelevant.

Single Site Failure

In the situation above, we show the "New Jersey" site as down.  In this case, the conventional voting quorum has a tie (1/2 - exactly half of the nodes).  In this case the quourm server is consulted.  Since only New York is talking to the quorum server, the quorum server grants quorum to the New York site.

Split Brain Avoided

In the case above, the link between the sites has been lost, but both sites and the quorum server are all up.  In this case, both New York and New Jersey contact the quorum server because each sees 1/2 nodes as being up - resulting in a tie condition.

In this case, the quorum server will choose one of the two sites to provide quorum to, and I assume in this case that New York was chosen.  Because New Jersey  wasn't granted quorum, it will shut its resources down.

What happens when the quorum server goes down?

That is the situation shown above.  Because New York and New Jersey are both up, they have 2/2 votes and both provide service as they should.  This illustrates the point that the quorum server is not a single point of failure.

Multiple Failures -> Loss of Service

In this final case, multiple failures have occurred - both New Jersey and the quorum server are down.  In this case, New York doesn't have quorum, so it shuts down services and none are provide by any node in the cluster.  Of course, this situation can be overridden in the cluster configuration by changing the quorum policy, but from an automated perspective, this is all that can be (should be) done.

Security Concerns

If you want to run your quorum server communications across networks which mig

Availability, MTBF, MTTR and other bedtime tales

If we let A represent availability, then the simplest formula for availability is:
    A = Uptime/(Uptime + Downtime)

Of course, it's more interesting when you start looking at the things that influence uptime and downtime.  The most common measures that can be used in this way are MTBF and MTTR.

    MTBF is  Mean Time Between Failures
    MTTR is Mean Time To Repair   

 A = MTBF / (MTBF+MTTR)

One interesting observation you can make when reading this formula is that if you could instantly repair everything (MTTR = 0), then it wouldn't matter what the MTBF is - Availability would be 100% (1) all the time.

That's exactly what HA clustering tries to do.  It tries to make the MTTR as close to zero as it can by automatically (autonomically) switching in redundant components for failed components as fast as it can.   Depending on the application architecture and how fast failure can be detected and repaired, a given failure might not be observable by at all by a client of the service.  If it's not observable by the client, then in some sense it didn't happen at all.  This idea of viewing things from the client's perspective is an important one in a practical sense, and I'll talk about that some more later on.

It's important to realize that any given data center, or cluster provides many services, and not all of them are related to each other.  Failure of one component in the system may not cause failure of the system.  Indeed, good HA design eliminates single points of failure by introducing redundancy.  If you're going to try and calculate MTBF in a real-life (meaning complex) environment with redundancy and interrelated services, it's going to be very complicated to do.

    MTBFx is  Mean Time Between Failures for entity x
    MTTRx is Mean Time To Repair for entity x
    Ax is the Availability of entity x   

Ax = MTBFx / (MTBFx+MTTRx)

In practice, these measures (MTBFx and MTTRx) are hard to come by for nontrivial real systems - in fact, they're so tied in to application reliability and architecture, hardware architecture, deployment strategy, operational skill and training, and a whole host of other factors, that you can actually compute them only very very rarely.  So, why did I spend your time talking about it?  That's simple - although you probably won't compute them, you can learn some important things from these formulas, and you can see how mistakes you make in viewing these formulas might lead you to some wrong conclusions.

Let's get right into one example of a wrong conclusion you might draw from incorrectly applying these formulas.

Let's say we have a service which runs on a single machine, which you put onto a cluster composed of two computers with a certain individual MTBF (Mi) and you can fail over to the other computer ("repair") a computer in a certain repair time (Ri).  With two computers, they'll fail twice as often as a single computer, so the system MTBF becomes Mi/2.  If you compute the availability of the cluster, it then becomes:

    A = Mi/2 / (Mi/2+Ri)

Using this (incorrect) analysis for a 1000 node cluster performing the same service, the system MTBF becomes Mi/1000.

    A = Mi/1000 / (Mi/1000+Ri)

If you take the number of nodes in the cluster to the limit (approaching infinity), the Availability approaches zero.

    A = 0/(0+Ri) = 0/Ri = 0

This makes it appear that adding cluster nodes decreases availability.  Is this really true?  Of course not!  The mistake here is thinking that the service needed all those  cluster nodes to make it go.  If your service was a complicated interlocking scientific computation that would stop if any cluster node failed, then this model might be correct.  But if the other nodes were providing redundancy or unrelated services, then they would have no effect on MTBF of the service in question.  Of course, as they break, you'd have to repair them, which would mean replacing systems more and more often, which would be both annoying and expensive, but it wouldn't cause the service availability to go down.

To properly apply these formulas, even intuitively, you need to make sure you understand what your service is, how you define a failure, how the service components relate to each other, and what happens when one of them fails.  Here are a few rules of thumb for thinking about availability

• Complexity is the enemy of reliability (MTTR).  This can take many forms
  
  • Complex software fails more often than simple software
  • Complex hardware fails more often than simple hardware
  • Software dependencies usually mean that if any component fails, the whole service fails
  • Configuration complexity lowers the chances of the configuration being correct
  • Complexity drastically increases the possibility of human error
    • What is complex software? - Software whose model of the universe doesn't match that of the staff who manage it.
• Redundancy is the friend of availability - it allows for quick autonomic recovery - significantly improving MTTR.  Replication is another word for redundancy.
• Good failure detection is vital - HA and other autonomic software can only recover from failures it detects.  Undetected failures have human-speed MTTR or worse, not autonomic-speed MTTR.  They can be worse than human-speed MTTR because the humans are surprised that it wasn't automatically recovered and they respond more slowly than normal.  In addition, the added complexity of correcting an autonomic service and trying to keep their fingers out of the gears may slow down their thought processes.
• Non-essential components don't count - failure of inactive or non-essential components doesn't affect service availability.  These inactive components can be hardware (spare machines), or software (like administrative interfaces), or hardware only being used to run non-essential software.  More generally, for the purpose of calculating the availability of service X, non-essential components include anything not running service X or services essential to X.

The real world is much more complex than any simple rules of thumb like these, but these are certainly worth taking into account.

More about quorum - updated

In a previous article[1], I talked about quorum, and alluded to some more details about quorum which I'll discuss here in a little more detail.  Let's examine a couple of common quorum tie-breaker methods, and see what's useful, and what's hype, and what's painful to use.

Can the standard voting quorum method fail?
By fail, I mean, can it grant quorum to two partitions simultaneously?  The answer, unfortunately, is yes, even though it seems like a mathematical impossibility.  This is because the world, unfortunately, is more complicated than simple mathematics, and quorum methods don't stand alone.  Quorum methods are tied to membership algorithms.  If a membership method fails, and for a period of time, a given node appears to be in the membership of two partitions, then while that's true, both partitions could legitimately think they have quorum.  Like many possible failures, this one is unlikely, but it is certainly possible.  Sigh...  Paranoia is so depressing.

SCSI Reserve Limitations

In the earlier article, I mentioned that SCSI reserve was often painful to use, without being explicit on why.  Let's explore that now.  SCSI reserve is an operation which happens at the physical disk volume level - that is, at the SCSI LUN  level.  With "dumb" disks this typically corresponds to an entire disk spindle - which nowadays is something like 180G-750G of disk data - which is clearly a significant waste of resources.

However, most people who use shared disks use them in a SAN are using it with a smart SAN disk controller which allows the creation of "logical" volumes which correspond to more much smaller sizes for a single SCSI LUN, using any RAID method you want.  For a disk used just as a quorum disk, you don't need to actually write anything on it, so you want it as small as you can make it.  But, most people probably make it RAID 1, which means that the minimum size is probably something like 2 gigabytes of SAN disk. If you have a large data center, it could easily have 50 clusters in it, and each one requires such a quorum device.  In this case, this makes for a lot of extra volumes, extra administration and possibilities for confusion and human error going on here.  In addition, smaller SAN disk units may have limitations on the total number of disk partitions they can manage.

So, perhaps you think - I'm smarter than that, I'll just use software volume managers to take care of this for me, and avoid all those extra logical disks in my SAN.  Unfortunately, that typically doesn't work.  This is because when you issue the SCSI reserve, it can't reserve a logical partition, only a physical partition, so many logical volume managers block reserve operations.  So, logical volume managers are not much help here.

To make it even more complicated, multi-pathing to disk devices often confuses (i.e., "breaks") disk reservation issues - particularly with SCSI II non-persistent reservations.

Of course, if you're replicating data instead of sharing it, disk reserve operations are of no help at all - since reserving a disk on one disk volume has no effect whatsoever on the other volume.

None of these considerations are changed by what kind of SCSI reserve you issue (persistent or the older non-persistent reserve).  However, there are even more problems that occur when you use the older non-persistent SCSI II reserve (the most commonly available kind), since it isn't persistent, and is broken by a bus reset or a device reset.  So, if you use SCSI II reserve, then you have to continually verify that you still have the reserve.

Count Key Data Disk Reserve

Mainframe disk subsystems support a non SCSI-based disk model, called the extended count key data (ECKD) disks.  These disks also support reserve methods similar to those provided by SCSI.

Quorum Daemon - Helpful or Hype?
Earlier, I said that the Linux-HA[2] quorum daemon[3] can help out here - not only for local shared disk situations, but for disaster-recovery type situations where you're replicating data between sites, and gave a hand-waving style argument on why that's so.  Let's see if we can go past the hand waving into more of the details so you can see what this is, how this works, and decide for yourself it would be helpful to you, or if this is just more hype written by someone who likes his project's work.

If you recall, I also described it as analogous to a software implementation of SCSI reserve.  But, in this case, there are no disks involved, so the hardware and SCSI protocol limitations mentioned above go away.  So, if the hardware has gone away, what kind of software has replaced it, and how does it work?

In the simplest view, the quorum daemon is simple - it takes TCP connections from clients, and when multiple clients both want to have quorum for the same cluster, it grants it to exactly one of them.  So, at this level of detail, it's quite simple, and the logic also straightforward.  But, there's a little more to it when you get into the details, so let's spend a few words describing the next level of detail, for those of you who are still skeptical.

TCP doesn't do a good job of telling you when your peer goes away, so both the client and the server processes send the other side heartbeats.  If the client stops hearing heartbeats, then he notifies his cluster that he no longer has quorum.  When the server stops hearing heartbeats, he takes quorum away from the client, and sends him noquorum messages.  Before switching quorum from one client to the other, the quorum daemon induces a configurable delay to allow the previous quorum owner time to notice that they've lost quorum and to shut down any resources it might be running.

What other interesting design features does it have?  Well, for one, all communication between client and server goes over SSL, with certificate authentication.  The server has a copy of the client's public certificate, and the clients have copies of the server's public certificate - so you don't need to get your certificates from a certificate authority - because we authenticate them by certificate, not through certificate authority (which could be a single point of failure).  Because of the way we use the SSL certificates, both authentication and authorization are bundled together.

Another nice design feature, is that a single quorum daemon doesn't have any fixed upper limit on the number of clusters it can support.  So, a single quorum daemon process should be able to support hundreds of clusters.  This makes it an advantage over just adding a third node to your cluster - because it can server hundreds of clusters, whereas typically a single computer can't join more than one cluster.  Indeed, since this doesn't run the cluster stack for any given cluster, it would be possible for the quorum daemon to work with multiple cluster implementations - as long as they put the hooks in to talk to it.

The combination of these two features allows you to put your quorum daemon on a completely different site, even at a colocation facility, and let your communication flow over the public Internet - securely.  This is really nice economical choice for your split-site DR-style clusters, for companies that don't have 3 or more major data centers.  Of course, if you have dozens of clusters that aren't split-site, you can make a single quorum daemon cluster with fencing to serve all the other clusters in your site.  All without bothering your storage group to create and manage all those little tiny quorum partitions for you.  In fact, without requiring shared storage at all.

Some kinds of software support resource-level quorum, that is, quorum on a resource-level rather than a whole cluster level.  Some of these are called resource-driven clusters.  The quorum daemon idea could also be used for those arrangements as well.

So, what's not to like about the quorum daemon?  Well, you do have to make a certificate for the server, and one for each client.  But, you don't have to pay for them, because they don't need to come from one of the well-known certificate producers.  It would be nice if the quorum daemon would sync its state to a slave copy that could be used in a cluster as a master/slave resource in case the machine running the quorum daemon failed.  You can still run it in a cluster with another machine to take over for it, but the takeover isn't as graceful as it would be if there were a slave copy ready to take over for the master complete with its quorum state.

I've recently learned that HP ServiceGuard[4] has a feature similar to this which they call the quorum server daemon (qs)[5].  The documentation indicates that HP's qs authenticates by IP address and does not use SSL for communication or authentication/authorization.  As a result, there are security concerns associated with it, and qs is probably unsuitable for split-site configurations.

And, even more recently, thanks to Nils Gorrol I've learned that Sun has a similar feature in their sqsd [6] since SunCluster version 3.2.  Looking at the documentation, it looks as though it has similar security issues that HP's qs has for split-site arrangements and potentially insecure networks.

Thanks to all those who keep me on the straight and narrow regarding the facts ;-)

Ping tiebreaker

Some HA systems provide  a ping tiebreaker.  To make this work, you pick a address outside the cluster to ping, and any partition that can ping that address has quorum.  The obvious advantage is that it's very simple to set up - doesn't require any additional servers or shared disk.  The disadvantage (and it's a big one) is that it's very possible for multiple partitions to think they have quorum.  In the case of split-site (disaster recovery) type clusters, it's going to happen fairly often.  If you can use this method for a single site in conjunction with fencing, then it will likely work out quite well.  It's a lot better than no tiebreaker, or one that always says "you have quorum".  Having said that, it's significantly inferior to any of the other methods.

If I've omitted things you want to know about, or this brings up ideas or questions in your mind, or you know of other implementations of the quorum daemon idea, by all means post a reply to this article.

References

[1] http://techthoughts.typepad.com/managing_computers/2007/10/split-brain-quo.html
[2] http://linux-ha.org/
[3] http://www.linux-ha.org/QuorumServerGuide
[4] http://h71036.www7.hp.com/enterprise/cache/6468-0-0-0-121.html
[5] http://www.docs.hp.com/en/B3936-90065/ch03s01.html
[6] http://docs.sun.com/app/docs/doc/819-5360/gbdud?a=view

Bad application design => Bad availability (more Rockies ticket debacle)

One final quick note on the Rockies ticket sales debacle - following up on my previous posting[1] on the subject.  This note discusses how to including the humans in your system design can improve both your perceived availability and your customer satisfaction while cutting your costs.

Pacolian did tweak their application to be a little more friendly to their infrastructure and perhaps fixed some networking issues, so that people did eventually get tickets to the Rockies/Red Sox games in Denver.  But, many, many people got told they were going to get tickets, but the system was still so slow and under such heavy load that it timed out many or even most users before they could pay for their tickets.  So, even the second try was highly unsatisfactory, and increased infrastructure costs.  Although you could argue that the system didn't crash, having it be so slow as to be unusable certainly creates the perception of unavailability in the minds of many users.

If you want to sell tickets to a hot event, there are many ways to do it.  For any system you come up with, you can expect the users who use it to try and game the system somehow to increase their odds.

What are the goals of selling tickets to events?

• sell all the tickets
• fairness
  • discourage professional ticket scalpers
  • limit number of tickets sold to any one party
  • make "gaming" the system harder
• minimize infrastructure costs
• minimize customer frustration

The way Pacolian did it, was to allow people to queue up in the web site and then sell the tickets first-come-first served beginning at a certain time.  Although this might sound reasonable because it's modeled after how tickets are sold at box offices in the real physical world, it's a bad idea in the Internet.  This encourages exactly the behavior that took their systems down, maximized customer frustration, and infrastructure costs, while still giving an edge to professional scalpers.  In their system, if you want tickets, obviously you open as many browser windows at a time as you can, hit the system as hard as you can at exactly the same time to increase your chances of getting in line first.  This creates a situation similar to what operating system people sometimes call thundering herd[2] [3] behavior.  The difference here is that the thundering herd loop included people and browsers.  It makes kind of a comical sight if you think about it...

Another method which I seem to recall hearing that other vendors use is a lottery system.  Such a system might work like this (details may need tweaking):

1. Announce a registration period of 8-18 hours where people can register for which games they want to buy tickets for, and which price ranges they're willing to buy.
2. Anyone registering during this time can register once, and give an email address where they can be reached, and which can be accessed from this computer.  Warn people not to do this on shared computers.
3. Each computer gets one cookie and one email address for registering.  If you try and re-register the same browser or use the email address multiple times, you're rejected.
4. At the end of the registration period, winners are randomly selected, and notified by email including a random cookie which is correlated to the cookie given to their browser earlier.  (you have to have both the cookie and the email to purchase a ticket).
5. They have a specified period of time to purchase the tickets, and any given credit card number can't be used more than once per game.  Any tickets not paid for by that time get sold to other applicants.

This means that there is no advantage to registering early or late.  As long as you register, you have the same chance as anyone else.  The peak load on such a lottery system system is probably one or two orders of magnitude below that of the Pacolian thundering herd design.

The result of this is:

• System availability goes up,
• Customer satisfaction goes up
• Infrastructure costs go down.
• The possibility of gaming the system still exists, but is probably no worse than the original Pacolian system.

All in all, a big win.  To be fair, I didn't think of this idea myself, but before the Rockies/Pacolian debacle, I'd never put much thought into ticket sales either.  Because I'm not an expert in this area, there are no doubt many improvements that an expert would make to my proposal to make it harder to game.  However, from an availability perspective, this application design is much more robust than the original Pacolian system because it takes into account the motivations of the humans that are part of the computer system.

References

[1] http://techthoughts.typepad.com/managing_computers/2007/10/the-cost-of-un-.html
[2] http://catb.org/jargon/html/T/thundering-herd-problem.html
[3] http://en.wikipedia.org/wiki/Thundering_herd

Split-brain, Quorum, and Fencing - updated

In some ways, an HA system is pretty simple - it starts services, it stops them, and it sees if they and the computers that run them are still running.  But, there are a few bits of important "rocket science" hiding in there among all these apparently simple tasks.  Much of the rocket science that's there centers around trying to solve a single thorny problem - split brain.  The methods that are used to solve this problem are quorum and fencing.  Unfortunately, if you manage an HA system you need to understand these issues.  So this post will concentrate on these three topics: split-brain, quorum, and fencing.

If you have three computers and some way for them to communicate with each other, you can make a cluster out of them and,each can monitor the others to see if their peer has crashed.  Unfortunately, there's a problem here - you can't distinguish a crash of a peer from broken communications with the peer.  All you really know is that you can't hear anything from them.  You're really stuck in a Dunn's law[1] situation - where you really don't know very much, but desperately need to.  Maybe you don't feel too desperate yet.  Perhaps you think that you don't need to be able to distinguish these two cases.  The truth is that sometimes you don't need to, but much of the time you very much need to be able to tell the difference.  Let's see if I can make this clearer with an illustration.

Let's say you have three computers, paul, silas, and mark, and paul and silas can't hear anything from mark and vice versa.  Let's further suppose that mark had a filesystem /importantstuff from a SAN volume mounted on it when we lost contact with it. and that mark is alive but out of contact.  What happens if we just go ahead and mount /importantstuff up on paul? The short answer is that bad things will happen[2]. /importantstuff will be irreparably corrupted as two different computers update the disk independently.  The next question you'll ask yourself is "Where are those backup tapes?". That's the kind of question that's been known to be career-ending.

Split-Brain

This problem of a subset of computers in a cluster beginning to operate autonomously from each other is called Split Brain[3]. In our example above, the cluster has split into two subclusters: {paul, silas} and {mark}, and each subset is unaware of the others.  This is the perhaps most difficult problem to deal with in high-availability clustering.  Although this situation does not occur frequently in practice, it does occur more often than one would guess.  As a result, it's vital that a clustering system have a way to safely deal with this situation.

Earlier I mentioned that there was information you really want to know, but don't know.  Exactly what information did I mean?   What I wanted to know was "is it safe to mount up /importantstuff somewhere else?".  In turn, you could figure that out if you knew the answer to one of these two questions:  "Is mark really dead?" which is one way of figuring out "Is mark going to write on the volume any more?"  But, of course, since we can't communicate with mark, this is pretty hard to figure out.  So, cluster developers came out with a kind of clever way of ensuring that this question can be answered.  We call that answer fencing.

Fencing

Fencing is the idea of putting a fence around a subcluster so that it can't access cluster resources, like  /importantstuff.  If you put a fence between it and its resources, then suddenly you know the answer to the question "Is mark going to write on the volume any more?" - and the answer is no - because that's what the fence is designed to prevent.  So, instead of passively wondering what the answer to the safeness question is, fencing takes action to ensure the "right" answer to the question.

This sort of abstract idea of fencing is fine enough, but how is this fencing stuff actually done? There are basically two general techniques:  resource fencing [4] and node fencing.[5]. 

• Resource fencing is the idea that if you know what resources a node might be using, then you can use some method of keeping it from accessing those resources. For example, if one has a disk which is accessed by a fiber channel switch, then one can talk to the fiber channel switch and tell it to deny the errant node access to the SAN.

• Node fencing is the idea that one can keep a node from accessing all resources - without knowing what kind of resources it might be accessing, or how one might deny access to them.  A common way of doing this is to power off or reset the errant node.  This is a very effective if somewhat inelegant method of keeping it from accessing anything at all.  This technique is also called STONITH[6] - which is a  graphic and colorful acronym standing for Shoot The Other Node In The Head.

With fencing, we can easily keep errant nodes from accessing resources, and we can now keep the world safe for democracy - or at least keep our little corner of it safe for clustering.  An important aspect of good fencing techniques is that they're performed without the cooperation of the node being fenced off, and that they give positive confirmation that the fencing was done.  Since errant nodes are suspect, it's by far better to rely on positive confirmation from a correctly operating fencing component than to rely on errant cluster nodes you can't communicate with to police themselves.

Although fencing is sufficient to ensure safe resource access, it is not typically considered to be sufficient for happy cluster operation because without some other mechanism, there are some behaviors it can get into which can be significantly annoying (even if your data really is safe).  To discuss this, let's return our sample cluster.

Earlier we talked about how paul or silas could use fencing to keep the errant node mark from accessing /importantstuff.  But, what about mark?  If mark is still alive, then it is going to regard paul and silas as errant, not itself.  So, it would also proceed to fence paul and silas - and progress in the cluster would stop.  If it is using STONITH, then one could get into a sort of infinite reboot loop, with nodes declaring each other as errant and rebooting each other, coming back up and doing it all over again.  Although this is kind of humorous the first time you see this in a test environment - in production with important services, the humor of the situation probably wouldn't be your first thought.  To solve this problem, we introduce another new mechanism - quorum.

Quorum

One way to solve the mutual fencing dilemma described above is to somehow select only one of these two subclusters to carry on and fence the subclusters it can't communicate with.  Of course, you have to solve it without communicating with the other subclusters - since that's the problem - you can't communicate with them.  The idea of quorum represents the process of selecting a unique (or distinguished for the mathematically inclined) subcluster.

The most classic solution to selecting a single subcluster is a majority vote.  If you choose a subcluster with more than half of the members in it, then (barring bugs) you know there can't be any other subclusters like this one. So, this is looks like a simple and elegant solution to the problem. For many cases, that's true.  But, what if your cluster only has two nodes in it?  Now,  if you have a single node fail, then you can't do anything - no one has quorum.  If this is the case, then two machines have no advantage over a single machine - it's not much of an HA cluster.  Since 2-node HA clusters are by far the most common size of HA cluster, it's kind of an important case to handle well.  So, how are we going to get out of this problem?

Quorum Variants and Improvements

What you need in this case, is some kind of a 3rd party arbitrator to help select who can fence off the other nodes and allow you to bring up resources - safely.  To solve this problem there is a variety of other methods available to act as this arbitrator - either software or hardware. Although there are several methods available to use as arbitrator, we'll only talk about one each of hardware and software methods: SCSI reserve and Quorum Daemon.

• SCSI reserve:  In hardware, we fall back on our friend SCSI reserve.  In this usage, both nodes try and reserve a disk partition available to both of them, and the SCSI reserve mechanism ensures that only one of the two of them can succeed.  Although I won't go into all the gory details here, SCSI reserve creates its own set of problems including it won't work reliably over geographic distances.  A disk which one uses in this way with SCSI reserve to determine quorum is sometimes called a quorum disk.  Some HA implementations (notably Microsoft's) require a quorum disk.

• Quorum Daemon:  In Linux-HA[7], we have implemented a quorum daemon - whose sole purpose in life is to arbitrate quorum disputes between cluster members.  One could argue that for the purposes of quorum this is basically SCSI reserve implemented in software - and such an analogy is a reasonable one.  However, since it is designed for only this purpose, it has a number of significant advantages over SCSI reserve - one of which is that it can conveniently and reliably operate over geographic distances, making it ideal for disaster recovery (DR) type situations.  I'll cover the quorum daemon and why it's a good thing in more detail in a later posting.  Both HP and Sun have similar implementations, although I have security concerns about them, particularly over long distances.  Other than the security concerns (which might or might not concern you), both HP's and Sun's implementations are also good ideas.

Arguably the best way to use these alternative techniques is not directly as a quorum method, but rather as a way of breaking ties when the number of nodes in a subcluster is exactly half the number of nodes in the cluster.  Otherwise, these mechanisms can become single points of failure - that is, if they fail the cluster cannot recover.

Alternatives to Fencing

There are times when it is impossible to use normal 3rd-party fencing techniques.  For example, in a split-site configuration (a cluster which is split across geographically distributed sites), when inter-site communication fails, then attempts to fence will also fail.  In these cases, there are a few self-fencing alternatives which one can use when the more normal third-party fencing methods aren't available.  These include:

• Node suicide.  If a node is running resources and it loses quorum, then it can power itself off or reboot itself (sort of a self-STONITH).  The remaining nodes wait "long enough" for the other node to notice and kill itself.  The problem is that a node which is sick might not succeed in self-suicide, or might not notice that it had a membership change, or had lost quorum.   It is equally bad if notification of these events is simply delayed "too long".  Since there is a belief that the node in question is, or at least might be, malfunctioning, this is not a trivial question.  In this case, use of hardware or software watchdog timers becomes critical.

• Self-shutdown.  This self-fencing method is a variant on suicide, except that resources are stopped gracefully.  It has many of the same problems, except it is somewhat less reliable because the time to shut down resources can be quite long.  Like the case above, use of hardware or software watchdog timers becomes critical.

Note that without fencing, the membership and quorum algorithms are extremely critical.  You've basically lost a layer of protection, and you've switched from relying on a component which gives positive confirmation to relying on a probably faulty component to fence itself, and then hoping without confirmation that you've waited long enough before continuing.

Summary

Split-brain is the idea that a cluster can have communication failures, which can cause it to split into subclusters.  Fencing is the way of ensuring that one can safely proceed in these cases, and quorum is the idea of determining which subcluster can fence the others and proceed to recover the cluster services.

An Important Final Note

It is fencing which best guarantees the safety of your resources.  Nothing else works quite as well.  If you have fencing in your cluster software, and you have irreparable resources (i.e. that would be irreparably damaged in a split-brain situation), then you must configure fencing.  If your HA software doesn't support (3rd party) fencing, then I suggest that you consider getting a different HA package.

See Also

General cluster concepts[8]

References

[1]   http://linux-ha.org/DunnsLaw
[2]   http://linux-ha.org/BadThingsWillHappen
[3]   http://linux-ha.org/SplitBrain
[4]   http://linux-ha.org/ResourceFencing
[5]   http://linux-ha.org/NodeFencing
[6]   http://linux-ha.org/STONITH
[7]   http://linux-ha.org/
[8]   http://linux-ha.org/ClusterConcepts