What is to Show for Five Years of Cloud Computing?

Just a few short years ago launching a virtual machine in the cloud was a simple and basic. With a couple of API calls and maybe a button click or two you were up and running in a just a few minutes. The choices were limited but it was nice. You could even get a little storage to go along with the instance. Just before that we were leasing, renting and co-locating dedicated physical hardware in data centers and it took weeks to order, provision, deploy, and set up the gear. Fast forward to today and we are now full-on in a cloud computing revolution redefining how technology is deployed. There are so many choices and so many of them good that it can be complete overwhelming to those trying to make sense of it all. On top of it all, every day I meet people who have never deployed anything in “the cloud.” It’s just as easy as ever to launch a machine but there is so much more available to the on-demand computing as a service would be client today. I was trying to think of what really is different or better today than it was five years ago. What’s really new to show for 5+ years of cloud computing innovation and effort?

First and foremost, people figured out that cloud computing is good for something really important (meaning people not Google). They figured out that the cloud in its various forms is phenomenal for capturing and processing what has come to be known as big data. This is a really important point. It’s never been easy, and still isn’t, to aggregate and process voluminous, high speed, or wildly unstructured data. In fact, prior to cloud computing coming of age it was down right impossible fiscally and technically. Now, it’s all there at the click of a few buttons as pretty as you like. You can now spin up a super computer for just a few dollars an hour to crunch even your most gnarly data sets.

A second fairly dramatic improvement is in the category of orchestration of resources. There are far more resources available to orchestrate for an infinite number of purposes but doing so has never been easier (not, I did not say easy). Due to the proliferation in understanding of the creation and consumption of API’s you can now quite literally with a single set of tools launch a server at several different cloud providers, geo locations, and even operating system varieties if so you wanted and if you’re clever with tools like Puppet, Chef, Cloud Formation, Cloud Foundry or others you can do it all from the comfort of your very own laptop in just a few minutes. You can quickly and relatively easily, historically speaking, compose masses of servers into useful services for nearly anything you can dream!

A third thing that’s changed is the raw power available via a command line or cloud console and in the newer implementations of older software architectures. You can now, in just a few moments, provision a server with 244 GiB memory and high speed 10 Gigabit Ethernet. And, that is just a building block to the real power. The real power comes as a result of massive improvements and capabilities in the arena of distributed computation, storage, and software defined networking. This allows you to provision dozens to thousands of these types of machines relatively on a whim. Frankly, not many people can even figure out what to do with all this power even if they do know how to provision it today. This has forced software architects and engineers to push forward much faster with zeal and learn how to write distributed applications and in many cases, the occasion is being met. So, raw power in both virtualized hardware and the software that can be deployed on it has come a very long way.

In summary, cloud computing had already been brewing for decades with its roots reaching far back in time. Grids, clusters and more were all precursors. However, it is striking how far things have come in just about five years. There has been unprecedented improvement and what feels like ever increasing speed of improvements. Good times indeed.

It's 2013! Things Break, Services Falter. Move Forward.

It's a New Year, I have the cloud, but I still have many of the same old Single Points of Failure.

It's known that a single point of failure (SPOF) is a risk. It's an Achilees heel so to speak. That goes for people, companies, planets, AMI's, AZ's, Regions, Countries, or beers in the fridge. Whatever processes you have to do your general day to day work should be able to deal with known SPOF's and be flexible enough to assimilate and adjust to newly found failure modes. But, and this is important, there is a substantial cost associated with eliminating certain SPOF's. Let's say you decided that you no longer are accepting of having Earth be an SPOF for your awesome blog. Well, in that case, you need a space program, and an interplanetary network that puts this desire out of reach unless you are NASA, Elon Musk, or Richard Branson. Admittedly, that is an extreme example but my point is that your tolerance for risk and downtime must be considered carefully for any technology for which you have implicit or assumed service level agreements with your users. Let's think about Netflix for a moment.

Netflix's service was severely impacted this last Christmas Eve by an outage affecting AWS ELB's in their US East region. Based on my arms length information about Netflix operations through what I've read that is public, in my opinion and far more than most organizations, Netflix understands this cost/benefit of utilizing AWS. They say themselves in a recent post:

"Our strategy so far has been to isolate regions, so that outages in the US or Europe do not impact each other."

"Netflix is designed to handle failure of all or part of a single availability zone in a region as we run across three zones and operate with no loss of functionality on two." Source: http://techblog.netflix.com/2012/12/a-closer-look-at-christmas-eve-outage.html

Netflix clearly understands the risk and still they have chosen to take it despite the known risks. They were completely at the mercy of AWS in this last outage since the failure was regional in nature and their systems do not allow for multi-regional failover within a country for a single user account or group of accounts YET; but they are working on it.

As an AWS client, they do have a reasonable expectation as a customer that the underlying primitives they use from AWS to compose their services will work reliably. In this case, that primitive was Elastic Load Balancers. Like an AMI is a virtual server, an ELB is something of a virtual load balancers. In VPC's ELB's can span AZ's but then, the ELB is an SPOF unless your service is capable of re-initializing an ELB dynamically when it ceases to serve its purpose and can then re-route traffic accordingly. This is non-trivial but can also likely be dealt with if you understand the various intricacies of geo aware anycast backed DNS services.

Someone asked me if the AWS outages of 2012 would make me re-think my plans for cloud computing in 2013. This does not change my cloud plans for 2013 in any way. But, to be clear, even though I really like AWS, AWS is not the cloud and the cloud is not AWS. AWS is a big and deeply important part of the cloud ecosystem. I'm quite thankful for all they've done to further the understanding of cloud around the world. They are likely to stay on top, from my point of view, for a long while. I and my teams deployed large amounts of AWS in 2012 supporting the services of numerous clients.

I don't think these outages will cause any meaningful pause in most cloud plans for 2013 for anyone who takes the time to understand these sorts of situations and doesn't just fall prey to FUD (Fear, Uncertainty, Doubt) and really is serious about moving to a cloud computing model will keep marching forward. It's not perfect but the benefits to business and technical agility far outweigh the risks and knowledge ramp up investment that is necessary to make full use of cloud computing.

Things break and outages happen. There are very few systems where this is not true and those systems have been designed specially to deal with an extreme need for continuous availability. Especially complex systems and systems deployed at a large scale like AWS can break in interesting ways. It's not so much that things break that is so bad. It is what is done next that matters to keep the same things from breaking again and again. AWS does a pretty good job on this front in my opinion. It performs, communicates, and adjusts far better than most hosting providers I have historical experience working with over the last 15 years or so. They have raise the bar substantially.

Regarding AWS's IaaS services. It is AWS's job to provide a reasonable SLA and maintain it. It is up to users of the services to provide their users with services that have a reasonable SLA and maintain. Decoupling the service from the server is at the heart of the accelerating innovation in hosting of internet connected services that began quite some time ago and now marches under the banner of cloud computing. Now, if you use their PaaS services, it's a bit of a different situation but that's the subject of a whole different discussion I suspect.

Supporting Information Blast from ProductionScale's blog past is contained within the following older posts of mine (in no particular order):

On Clouds and SPOF's, 4/2011

http://productionscale.com/home/2011/4/22/on-clouds-and-spofs-or-the-great-aws-outage-of-april-2011.html

Cloud Computing Impacts Technical Agility, 8/2010

http://productionscale.com/home/2010/10/8/cloud-computing-impacts-technical-agility.html

The Traits of a Modern IT Organization, 8/2008

http://productionscale.com/home/2008/8/25/the-traits-of-a-modern-it-organization.html

Thoughts on the Business Case for Cloud Computing, 4/2009

http://productionscale.com/home/2009/2/4/thoughts-on-the-business-case-for-cloud-computing.html

Get Your Head in the Clouds, 4/2008

http://productionscale.com/home/2008/4/24/cloud-computing-get-your-head-in-the-clouds.html

Why Should Businesses Bother with Cloud Computing, 3/2009

http://productionscale.com/home/2009/3/13/why-should-business-bother-with-cloud-computing.html

Building an Application upon Riak - Part 1

For the past few months some of my colleagues and I have been developing an application with Riak as the primary persistent data store.  This has been a very interesting journey from beginning to now.  I wanted to take a few minute and write a quick "off the top of my head" post about some of the things we learned along the way.  In writing this I realized that our journey breaks down into a handful of categories:

• Making the Decision
• Learning
• Operating
• Scaling
• Mistakes

We made the decision to use Riak around January of 2011 for our application.  We looked at HBase, Cassandra, Riak, MySQL, Postgres, MongoDB, Oracle, and a few others.  There were a lot of things we didn’t know about our application back then.  This is a very important point.

In any event, I’ll not bore you with all the details but we chose Riak.  We originally chose it because we felt it would be easy to manage as our data volume grew as well as because published benchmarks looked very promising, we wanted something based on the dynamo model, adjustable CAP properties per “bucket”, speed, our “schema”, data volume capacity plan, data model, and a few other things.

Some of the Stack Details

The primary programming language for our project is Scala.  There is no reasonable scala client at the moment that is kept up to date for Riak so we use the Java client.

We are running our application (a rather interesting business analytics platform if I do say so myself) on AWS using Ubuntu images.

We do all of our configuration management, cloud instance management, monitoring harnesses, maintenance, EC2 instance management, and much more with Opscode Chef.  But, that’s a whole other story.

We are currently running Riak 1.0.1 and will get to 1.0.2 soon.  We started on 0.12.0 I think it was... maybe 0.13.0.  I’ll have to go back and check.

On to some of the learning (and mistakes)

Up and Running - Getting started with Riak is very easy, very affordable, and covered well in the documentation.  Honestly, it couldn't be much easier.  But then... things get a bit more interesting.

REST ye not - Riak allows you to use a REST API over HTTP to interact with the data store.  This is really nice for getting started.  It’s really slow for actually building your applications.  This was one of the first easy buttons we de-commissioned.  We had to move to the protocol buffers interface for everything.  In hind sight this makes sense but we really did originally expect to get more out of the REST interface.  It was completely not usable in our case.

Balancing the Load - Riak doesn’t do much for you when it comes to load balancing your various types of requests.  We settled, courtesy of our crafty operations team on an on application node haproxy to shuttle requests to and from the various nodes.  Let me warn you.  This has worked for us but there be demons here!  The configuration details of running HA proxy to Riak are about as clear as mud and there isn’t much help to be found at the moment.  This was one of those moments over time that I really wished for the client to be a bit smarter.

Now, when nodes start dying, getting to busy, or whatever might come up you’ll be relying on your proxy (haproxy or otherwise) to handle this for you.  We don’t consider ourselves done at all on this point but we’ll get there.

Link Walking (err.. Ambling) - We modeled much of our early data relationships using link walking.  The learning?  S-L-O-W.  Had to remove it completely.  Play with it but don’t plan on using this in production out of the gate.  I think there is much potential here and we’ll be returning to this feature for some less latency sensitive work I perhaps.  Time will tell...

Watchoo Lookin’ for?! Riak Search - When we stared search was a separate project.  But, we knew we would have a use for search in our application.  So, we did everything we could to plan ahead for that fact.  But, by the time we were really getting all hot and heavy (post 1.0.0 deployment) we were finding our a few very interesting things about search.  It's VERY slow when you have a large result set.  It's just the nature of the way it's implemented.  If you think your search result set will return > 2000 items then think long and hard about using Riak's search functions for your primary search. This is, again, one of those things we’ve pulled back on quite a bit. But, the most important bits of learning were to:

• Keep Results Sets small
• Use Inline fields (this helped us a lot)
• Realize that searches run on ONE physical node and one vnode and WILL block (we didn’t really feel this until data really started growing from 100’s of 1000’s of “facets” to millions.

At this point, we are doing everything that we can to minimize the use of search in our application and where we do use it we’re limiting the result sets in various ways and using inline fields pretty successfully.  In any event, just remember Riak Search (stand alone or bundled post 1.0.0 is NOT a high performance search engine).  Again, this seems obvious now but we did design around a bit and had higher hopes.

 
OMG It’s broken what’s wrong - The error codes in the early version of Riak we used were useless to us and because we did not start w/ an enterprise support contract it was difficult sometimes to get help.  Thankfully, this has improved a lot over time.

Mailing List / IRC dosey-do - Dust off your IRC client and sub to the mailing list.  They are great and the Basho Team takes responding there very seriously.  We got help countless times this way.  Thanks team Basho!

I/O - It’s not easy to run Riak on AWS.  It loves I/O.  To be fair, they say this loud and clear so that’s my problem.   We originally tried fancy EBS setup to speed it up and make it persistent.  In the end we ditched all that and went ephemeral.  It was dramatically more stable for us overall.

Search Indexes (aka Pain) - Want to re-index?  Dump your data and reload.  Ouch.  Enough said.  We are working around this in a variety of ways but I have to believe this will change.

Basho Enterprise Support - Awesome.  These guys know their shit.  Once you become an enterprise customer they work very hard to help you.  For a real world production application you want Enterprise support via the licensing model.  Thanks again Basho!

The learning curve - It is a significant change for people to think in an eventually consistent distributed key value or distributed async application terms.  Having Riak under the hood means you NEED to think this way.  It requires a shifted mindset that, frankly, not a lot of people have today.  Build this fact into your dev cycle time or prepare to spend a lot of late nights.

Epiphany - One of the developers at work recently had an epiphany (or maybe we all had a group epiphany).  Riak is a distributed key value data store.  It is a VERY good one.  It’s not a search engine.  It’s not a relational database.  It’s not a graph database.  Etc.. etc..  Let me repeat.   Riak is an EXCELLENT distributed key value data store.  Use it as such.  Since we all had this revelation and adjusted things to take advantage of the fact life has been increasingly nice day by day.  Performance is up.  Throughput is up.  Things are scaling as expected.

In Summary - Reading back through this I felt it came off a bit negative.  That's not really fair though.  We're talking about nearly a year of learning.  I love Riak overall and I would definitely use it again.  It's not easy and you really need to make sure the context is correct (as with any database).  I think team Basho is just getting started but are off to a very strong start indeed.  I still believe Riak will really show it's stripes as we started to scale the application.  We have an excellent foundation upon which to build and our application is currently humming along and growing nicely.

I could not have even come close to getting where we are right now with the app we are working on without a good team as well.  You need a good devops-like team to build complex distributed web applications.

Lastly and this is the real summary, Riak is a very good key value data store.  The rest it can do is neat but for now, I'd recommend using it as a KV datastore.

I'm pretty open to the fact that even with several months of intense development and near ready product under our belt we also are only scratching the surface.

What I'll talk about next is the stack, the choices we've made for developing a distributed scala based app, and how those choices have played out.

Stop Staring at my Polyglot!

I received an interesting comment/question via my blog recently.  It went a bit like this... 

I’m developing a distributed cloud application but my developers are pushing back on me for having a polyglot database strategy.  What should I do?

I won’t get into exactly what it is they are doing since that would take several more pages.  This is something of a stream of thought post so apologies if it is a little rough around the edges.  The easiest way to answer is in the context of an application I’ve been working on for a while that has some similarities to what this person wants to build.  Everything I’m describing is part of an app I’ve been building with a client since earlier this year.

Typically you'll need a few layers of "data storage” for any distributed batch or real time application (cloud native application) which is what I understand that you are trying to build.

I consider anything that holds data that is for presentation, computation, or transformation part of the data storage architecture and I like to break it down by time in storage from least amount of time to most. 

Short or Very Short Term: Single node caches (like memcached) or volatile computer node memory
Mid Term: Queue's and IMDG's
Long Term  Durable Storage:  Dynamo and BigTable derivatives abound 

There are numerous database products that live in or even between those tiers these days; more than ever before.  By no means is what follows even close to an exhaustive list.  A quick list of the ones I have worked with in the last few months personally looks like: 

Short Term: Memcached, Redis, RabbitMQ, ZeroMQ, DRAM, APC, MongoDB
Mid-Term: Redis, GridGain, RabbitMQ, ZeroMQ, MongoDB
Long-Term: Riak, MongoDB, S3, Ceph, Swift, CloudFiles, EBS, HBase

Short-Term storage is ALL in memory, not persisted to disk, and not intended to be used for long periods of time.  Your application also has to be able to deal with the fact that this type of storage is essentially ephemeral.  If the node gets a KILL signal from some source or another your app needs to know how to deal with this gracefully.  In other words, storage here is not durable at all.

Mid-Term storage is used for longer running processes.  It benefits greatly from being distributed and having a higher degree of durability.  This is generally still where most of the work in done in main system memory (no disk I/O) but also where you might do complex calculations or data transformations on your way to your goal.  You do it here because it’s fast.  You do things here because they can be shared amongst lots of workers (like queue subscribers). 

Long-Term storage is used for exactly that, long term durable storage of important data that provides sufficient and reasonable interfaces from which to retrieve that data again when needed.  Preferably it’s possible to do things like map-reduce jobs so that you can iterate and retrieve what is necessary which you may then operate on at one of the higher levels up this stack.

You’ll see that I’ve put some of them in all or multiple categories which might seem odd until you understand how they work and match the technology to what ever you are trying to achieve from a business perspective.

I have a tendency to avoid things that require overly complex operational management issues for starting up projects because I like to try to get my TCO (Total Cost of Ownership) over time (3-5 years) as low as possible while achieving the project goals and SLA’s.  There are a couple of exceptions on the list above that do have more operational overhead (MongoDB and HBase) but they are good enough in the right context that you might want to learn and use them anyway.

Now, back to the question at hand.  Should I use one type of DB or many for the needs at hand.  In this case, I’ve told them that they should use as few as possible, possibly only one.  The reason for this was that in their case they will value speed, consistency, and lower cost of operations at this early stage of their project.  They are developing an interesting distributed system for cool reasons.  I recommended a choice to them that I think will help them get to their goals fast and cost effectively while allowing them down the line to break off pieces of the application later as and if needed.

Parting words are that it will, over time, be nearly unavoidable that this (and most) applications of a distributed nature end up being database polyglotoumous.  However, I do think it adds a lot of complexity and overhead and in the early stages of a project it's not usually necessary unless what you are doing is of great complexity in which case you might want to break that down anyway to something more manageable.

Scale Planning and the AKF Scale Cube

There are a lot of ways to draw diagrams for availability and scalability.  I use different ones for different purposes all the time.  However, when I was reading the Art of Scalability by AKF partners I ran across a nice compound diagram they call the AKF Scale Cube which helps simplify the explanation of the multi-dimensional nature of scalability issues in complex web application scenarios.  

I’ve been using this visualization model to help me explain how things fit together in both technical and business discussions.  It comes in very handy I must say.

Most recent I’ve been using it to describe a gnarly distributed application I’m working on for a client.  What follows is a generalized version of a functional use and some discussion of a compound view of ones I have created for clients of mine.

 


Some base-line definitions are in order if you haven’t read the book mentioned above.

X-Axis - Horizontal Scalability, Clones, Scale Out.  These are terms often associated with the X-Axis.  In the case of this graph, day you build a data processor then make 10 copies of it.  Well, that’s scaling to 10 units on the x-axis in this graph.  Depending on your application, this can help you increase your capacity; but not always!

Y-Axis - I’ve called this axis functional decomposition for a long time.  It can be thought of as breaking the application down from a monolithic single instance into discreet stand-alone parts.  I have some examples that are a bit of mix from various projects I’ve worked on in the past here.

Z-Axis - This is the tricky one for most folks.  This is what people might call sharding, partitioning, etc.  Keep in mind, I’m not only talking about a database here.  I’m talking about an entire complex multi-faceted distributed highly-available web application.

0,0,0 - The intersection of all three axis or 0,0,0.  This is what some would call an all-in-one server.  It’s often used for proof of concept for for launching without a care in the world for future growth needs.  There’s nothing wrong with it as long as you understand the limitations and technical debt associated with the approach.

Z-Axis Item Explanations

Client - Assuming this is a multi-tentant application you may want to shard your application by client such that each client or group of clients is assigned somehow to a specific cluster of nodes.

Geography - Assuming you’d like to have built in DRBC and your applciation is capable of surviving being split up into many pieces then you could end up sharding your application by data center and broader geographies such as city, state, country.

External Cloud - Using IaaS and PaaS resources outside of your own data centers.  For a refresher on IaaS and PaaS see the article I wrote, “Cloud Computing:  Get Your Head in the Clouds,” in 2008 that was heavily read over the years.

Internal Cloud - Using your own infrastructure resources BEHIND your own firewall to do whatever it is that your application does.  This doesn’t always have to be a cloud.  If you want to know what I think it takes to be a cloud then read the several articles I wrote over the years related to that topic.  I do set the bar pretty high though I’ve learned.

Purpose - You might want to simply partition along the Z-Axis by any generic purpose for various reasons.  I think of this a little bit like saying I want to put all widgets in data node 1 and all waggles in data node 2.  They’ll both fit in a single node but maybe I want to spread my risk around.  This one is a little nebulous but it can engender fun conversations about why things need to exist at all.

Shard Key - We see this all the time in traditional RDBMS style deployments and even in some of the newer tools in the NOSQL world.  It’s basically just some index of what node you put things one somewhere.  For those of you that had to deal with libraries before the internet you’ll remember the lovely card catalog.  It’s was nicely set up to help you figure out with shard of the library your book was close to.  Then, when you got there, good old dewey decimal system kicked in to take you the rest of the way.

Y-Axis Item Explanations

Data Processing - this could be some application that transforms data from one state to another.  For example, it might simply remove all the spaces in a document and replace them with dashes.  That’s a bit of a silly example, but just to make the point.

Data Aggregator - I’ve had to build project after project that needed one form or another of data aggregation.  So, just think of this as something that might consume and RSS feed and stick it in a database of some kind.

Distributed Calculation - I’ve been doing work and research with Map-Reduce, Actor Models, the Bulk Synchronous Parallel Model and more exotic instruments from past, present, and future.  This is simply something that does some kind of math or calculation of some sort.  For example, counting all the uses of the word onomatopoeia in 50TB of English essays by high school students across 100’s of of compute nodes.

Processor App - This is just a generic discreet application that processes something, like an API request for example.

Web App - This is an application, in my case, written in a modern MVC framework that has the job of interacting with web users and getting things done in the back-ground in various ways with various services.

Base Installation - I think of this as just shared code.  One of the developers I have been working with recently suggested that we extract a number of commonly used components from various application pieces on the Y-Axis and build a library of sorts.  Great suggestion in this case, so I stuck in on my general diagram to remind me in the future.

What’s interesting about all these conceptual applications is that if you create them correctly and with the correct architectural models that each item that lives on the X Axis will also be able to scale on the X and Z axis.  For example, you could have your web application running 5 X-Axis copies in 4 Y-Axis partitions; say external cloud, by client, purpose, and by geography per client.  So, you’d end up use four AWS Availability zones in 2 AWS locations running 96 application nodes in total.  Of course, your application has to be built correctly to take advantage of all of this distribution at every level.  But, that’s a topic for a later date I suppose.

In summary, this post was just to share some of my thinking around the use of a very nice visualization tool by the fine folks at AKF Partners.  So, a shout out to them for the nice tool and hopefully this helps people a bit understand how it can be used / thought of in a variety of ways.

Just remember, it's not one-size-fits all.  Your use, labels, and needs for such things will vary greatly depending on what you are trying to architect, develop, and deploy.