Durability, Scalability, Availability

When using a varnish/squid/nginx or any reverse proxy +cache / http accelerator on top of a modern web framework / web application that has it’s own caching layer you may run into some strange seemingly uncontrollable TTL’s on your content. I have come to rely heavily on varnish in the last year and squid before that. If your reverse proxy is setup defensively (and in my opinion it should be) then you probably tend to ignore somewhat the http headers coming out of the application layer and put an additional TTL on top of all of the (non-dynamic) content. A for instance: There is a certain editorial site that I maintain that I cache everything on the site in varnish for 30 mins… period. This means that once an editor pushes a publish button and the content is picked up into the cache it won’t change for a half an hour. OR WILL IT? Turns out that since the application layer also has a cache lifetime of something like 5 mins the object in varnish may or may not refresh for an hour or more. How? Well that’s where it gets difficult to explain. It all depends on when the change get’s published to the content with respect to the application servers cache expiry on the object AND varnish’s cache expiry on the object AND the time of the next request for the object after Varnish’s cache expiry has happened. HAH! Before we go any further a picture is needed.

As you can see from the graph there are two sawtooth looking lines representing the two types of cache’s in use on this site. I have made the assumption that an editor publishes a change to an object (say a homepage of a site) every half hour. Really they may publish 25 times within that half hour but let’s just keep it simple. So in the best case example an editor publishes a change. That publish event happens RIGHT BEFORE the cache in varnish AND the cache in the application server is invalidated for that object and the user request comes in RIGHT AFTER the cache is invalidated in both places so that the user request flows all the way upstream to the application server which re-renders the page with the new content, caches it and then sends it back up to varnish which also caches it and then sends it to the client. That is the BEST POSSIBLE case. That RARELY happens. What normally happens is that a user request comes in, hit’s the varnish cache and is served a moderately old object. I say moderate because more than likely on average the object is somewhere near the middle of it’s TTL. IF the varnish TTL has expired on an object what is next most likely is that the APP server is somewhere within the middle of the TTL for the object within IT’S cache and the app server will just serve the cached response and then varnish will do what it knows best and CACHE THE OBJECT FOR ANOTHER CACHE CYLE. That means ANOTHER 30 mins in this scenario. Hmmm editors are not to happy about this situation. Thing is. It get’s worse. What CAN happen is that editor publishes a change immediately AFTER the app server rolls over on it’s cache for the object, re-caches the old content and never sees the editors change. The re-cached old content in the appserver is then served again to varnish who could (as within the last example) roll over on it’s cache RIGHT at the end of the app servers cache cycle for the object BUT BEFORE the appserver’s cache actually invalidates. As explained in the last example the appserver serves varnish the old object, varnish re-caches the old object and then that old object is within the varnish cache for the extra cache cycle. SO in the worst case the editors story doesn’t get picked up for what is essentially the sum of TWO FULL CACHE CYCLES of both cache’s TTL’s.

All of this is VERY hard to explain to laypeople. Hence the picture above with the perty colorful stars.

I hope this helps someone out there. I know I will use it over and over again..

P.S. I do realize you can always just get rid of one of the application servers cache. Varnish is performing the same exact function so why bother. Fact is I simplified this example to html buffer caching in both layers so that you it would be easier to understand. What Is really going on is that the application layer is not caching the html buffer. It is actually caching the results of the queries from the database in memcache for a period of time. In addition there are “partials” strewn through out the app that are cached independently of context within the app which gives finer grained control over areas that change less frequently than varnish allows for. BTW I do realize varnish supports ESI allowing certain of these partials to be re-calculated at the app server rather than at varnish even when varnish serves some of the page from it’s cache. I hope to utilize ESI in the future but that will require an app re-write. ALSO I would love to have the application server send the cache invalidation requests upstream to varnish when it’s own cache cycles roll over on objects. Unfortunately the off the shelf app we are using is not that smart and again it would require a re-write to get that in there.

I have recently been doing some work with Camel routing library/engine for java that plays really nice with Spring and JMS via ActiveMQ. Camel makes it really easy to implement EIP (Enterprise Integration Patterns) into your Java app and ultimately allows you the beautiful flexibility describe in Gregor Hohpe’s work on enterprise integration patterns… great book btw!

Anyway I got to thinking how fun it would be to create a custom predicate that employs a machine learning algorithm (possibly based on a neural network) to do image recognition and route messages (images) based on it’s understanding of them. Given enough time and enough input this could really open up some awesome possibilities!

The most immediately applicable solution is moderation, allowing the AI to do the first pass, pipe positives to an administrative interface where an actual human being could determine if the image was actually naughty or if the algorithm flagged a false positive. The feedback on false positives could then be driven back into the pipeline thus tuning the algorithm further (false positives) or deleting the image from the servers and warning the contributing user (actual positives).

Another cool application would be to do something like google is doing where you could use the routing engine and the image recognition predicate to make a first guess at what the image is basically of and then query all sorts of different data stores/apis to find more information about the image. You wouldn’t need to do so much legwork query everything or storing a “master index” you could simply allow the algorithm to make an educated guess and say for example “this is a painting” or “this is a barcode” and then ONLY query the applicable datasources.

Of course this will be full of false positives. BUT if you make it fun and allow feedback back into the routing engine from the end users the algorithm will get smarter over time!

I have been using Varnish for quite some time and have always wished that there was some way for Varnish to know to serve “Stale” pages when the upstream application servers are swamped. There is actually a feature request for this on the Varnish Trac system here. NOTE: this feature should not really be necessary unless you have underestimated the ability of your application servers to handle your traffic. However even after proper capacity planning sometimes you get well… DUGG. We all know the “digg effect” (formerly referred to as the “slashdot effect”) and it’s repercussions (500, Guru meditation, Houston we have a problem!) There are many ways to skin a cat, but none would be as simple as this (considering we have an existing varnish setup). I should note that simply getting “Dugg” or “Slashdotted” normally wouldn’t take down a site with a proper reverse proxy setup based on Varnish. If your TTL is appropriate and you are using an appropriate GRACE value (for you Squid readers: “stale-while-revalidate“) then you will probably not saturate your app servers. Unfortunately if your content is good and your UI is right then maybe, just maybe a certain percentage of your new readers will stick around. And here is where it gets scary for the app servers. Maybe just, maybe your new readers will start to navigate in ways that your cache is not used to. Maybe they will start to hit those really OLD articles that haven’t been requested in months! If you think about your sites content vs it’s popularity you it will look something like this:

No matter what you do there will always be something that falls into those “long tails” if your traffic patterns shift suddenly you can very well start to make a lot more request to your upstream servers than you (or more importantly your reverse proxy) expected.

Back to the task at hand. What can I do while I wait for the Varnish team to put this feature through? EASY… use Squid! There are so many debates over which reverse proxy is currently the fastest, which one is easier to setup or integrate with legacy apps etc. I’m certainly NOT trying to get into that! In fact I will skirt the issue entirely saying this: when the features are right and you can afford to use it then why not? NOW don’t get me wrong. Afford can mean a lot of things. Take it as you will. I personally HATE using software, ANY software when I don’t have to. In fact I try to design my stacks as small as possible. As a general rule LESS SOFTWARE IS BETTER! It means less maintenance, less quality assurance… less hastle! However there are situations like the one I described above when you are put between a rock and a hard place. I can either:

A) Swap Varnish out completely and start using squid.

B) Augment my http acceleration layer with squid.

C) Buy more application servers and avoid the issue.

I wish, I wish, I wish C was always an option. Unfortunately not all client’s can afford to simply throw more money at the problem. If I had my choice I would scale horizontally off to the…horizon. SO I now get to choose between A and B. A is what my Sysadmin gut feeling (about never using more software than necessary) is telling me to do. BUT A also has the Test Engineer in me screaming “You will have to test everything all over again!”

Sooo here is another instance where the REAL WORLD comes crashing down on good systems engineering. C is the cheapest most cost effective solution. It could be said that maintaing another piece of software over time is going to be more costly than the upfront cost of swapping out Varnish entirely. But consider this…the Varnish feature that I was mentioning earlier… has already been assigned. It is only a matter of time before someone decides to pick it up and implement it. Hell I might even go ahead and do it if I can find the time. (BTW if your reading this month’s past the publish date of this post then you should definitely check that Trac ticket and see what has become of it.)

C it is. Now I am going to have to dust off my Squid skills and install that beast again. (Of course I couldn’t get through an article about varnish and squid with out some opinion…. Setting up Squid is not the easiest thing in the world!)

I had a bit of a quandary today as I reviewed some code today and I decided to write up this little post on the experience. In essence it came down to choosing between code readability and code reuse. I am a huge fan of both and they almost never conflict in fact they almost always exist in a symbiotic state. But not this time.

This title of this post has been buzzing around in my head recently as I have been doing a lot of performance tuning on a very large and HIGHLY trafficked web site of a major print publication. Anyway I was recently reading (reference coming soon! until then dig through my delicious links) a blog post on the topic of performance tuning and why it is silly how often people micro-tuning. Think of that in the same negative respect that you think of micro-managing. It’s a waste of time and resources. The moral: even though performance tuning is normally a good thing, after a while it looses it’s value. You spend your money on the low hanging fruit (AHEM more hardware!)

Back to today. The code that I was reviewing had a few very cryptic stanzas that included a call to a function that consumed a single integer based parameter which after thorough inspection simply ended up determining if the query it finally triggers against the database is ascending or descending. Now I honestly doubt this developer (no matter how Jr) actually thought they would be increasing the performance of the application by including this integer parameter. In fact I am sure that they assumed that they were doing a good thing by reusing code (which I am a HUGE proponent of normally!) but when it came time for another developer to get into the code to augment it slightly it took exponentially more time then it would have if the original developer had simply a) created the database queries inline or b) created two separate nearly identical functions. Instead (fyi we are using an ORM that follows the ActiveRecord Pattern) the object has one single method getNextArticle(int foo) that returns both the previous AND the next article! How confusing! Anyway the best solution that required the least new untested code at this point was to wrap that function with two new functions that map to the integer values. Now we have:


/*Note that the following function has a different method signature than the original and thus can coexist due to the method overloading feature of java*/
getNextArticle(){
return this.getNextArticle(1);
}



getPreviousArticle(){
return this.getNextArticle(2);
}


I can already hear someone out there saying: “That’s silly why not re-factor the whole object so that it has:

getNextArticle()

getPreviousArticle()

wrapping some new function that is called with a String paramater like ASC or DESC.” Well to put it simply, time. Regression testing takes time and we would have to do a heck of a lot more of it on this project that unfortunately does not yet have 100% unit test coverage (to my DISDAIN) and does not have an automated testing setup. So the moral of the story is this. Even though code reuse is almost always a GOOD sign of proper software development practices sometimes it can lead to poor readability which then becomes more of a maintenance problem than the code reuse actually solves. In retrospect it would have been great to have found this code before it went through thorough user acceptance testing. However due to the accelerated nature of meeting client demands and working on unrealistic project schedules we were not able to do enough peer code reviews of code nor were we able to create a team large enough to do pair programming.

AHH how the real world always ruins principles, techniques, patterns and paradigms that work so well on paper!