As I’ve already stated, I’ve spent the last few weeks working on putting together log aggregation so that we know what our new service looks like in real time.
I’ve incorporated IIS logs, the application logs, machine statistics (memory, CPU, etc) and Windows Event logs into the log aggregator, and successfully used those events for analysis during load testing.
There was one piece missing though, which meant there was a hole in our ability to monitor how our service was actually operating in the face of actual usage.
The Elastic Load Balancer, or ELB, that sits in front of the publically accessible web service.
During load testing, I noticed that sometimes JMeter would record an error (specifically a 504, Gateway Timeout) but our dashboard in Kibana would show nothing. No errors, everything seemed fine.
It turned out that there was a default timeout on the ELB of 60 seconds, and at that point in the load testing, some requests were taking longer than that without causing any traffic over the connection. The ELB would terminate the connection, return a 504 to the client, but the request would still complete successfully (eventually) in the backend.
I needed to get eyes on the ELB.
Turning logging on for an ELB is fairly easy.
Just give it the S3 bucket you want it to log to, a prefix to use for entries made into the bucket and a time interval, and off it goes. All of this can be done through the CloudFormation template, which fits well into our strategy for environment setup (no manual tasks, automate all the things).
The only complex bit is setting up a bucket policy that sets the correct permissions to allow the ELB to write to the bucket, which is all pretty well documented. There is simply a well known ARN for what I assume is all Load Balancers in a region, and you setup a simple Put/Get/List policy to allow it to do its thing.
The only gotcha I ran into was when I included an underscore (_) in the prefix configuration setting for the ELB. The prefix setting is intended to make sure that the keys for files written into the bucket start with a common value. When I included an underscore, I got nothing but Access Denied errors. This was at the same time as I was setting up the bucket policy, so I assumed I had done that incorrectly. Turns out my bucket policy was flawless, and it was a completely unrelated (and unexpected) issue causing the Access Denied errors.
With that fixed though, the logs started flowing.
The ELB logs contain things like the ELB IP and port, where the request was forwarded to (IP and port again), the time to forward, process and respond to requests (3 separate entries, process is the time it takes for your server to do its thing), response codes, bytes transferred and other things. Very similar to IIS really, which is not unexpected.
Now all I had to do was get the information into our Log Aggregator.
Stashing Those Logs
I had been using Nxlog as my log processor. It was responsible for picking up files, processing them as necessary, enriching them with various pieces of information (hostname, component, application) and then shipping the results off via TCP to our log aggregator where Logstash was listening.
Nxlog is a fine product, but its scripting language is a hard to get a handle on, and the documentation is a bit sparse. Also it has no concept of decimal numbers, which meant that I had to convert some numbers to integers (like decimal seconds to milliseconds) via regular expressions. Altogether it got the job done, but I wasn’t particularly happy with it.
I thought that since I needed to do something a little bit more complicated (get files from S3 and process them) that I would use Logstash this time. Logstash as a log processor is a lot easier to distribute, configure and debug, which is nice. Its configuration is all in json, and is very easy to wrap your head around, and it has lots of component to accomplish various tasks like getting files from S3, parsing CSV lines, mutating fields to the correct type, etc. It even has a mutator (Logstash calls them filters) that allows you to execute arbitrary Ruby code for those times when you have to do something unusual.
Even better, Logstash is what's listening on the other end of the pipeline, so they play well together, and you only need to know 1 piece of software, instead of 2.
I built a similar distributable project to what I built for Nxlog, that creates a NuGet package that Octopus can deploy to get a copy of Logstash up and running on the target machine as a Windows Service. I won’t go into this in too much detail, but it was essentially the same thing that I did for Nxlog, except with different dependencies (JRE, Logstash, NSSM for service installation/configuration).
I added a small EC2 instance to our environment setup to act as a Log Processor, with the intent that it would immediately be used to process the ELB logs, but may also be used in the future to process other logs that don’t necessarily fit onto a specific machine (S3 access logs is the only one that comes to mind, but I’m sure there are more). The Logs Processor had an IAM role allowing it full control over the logs bucket that ELB was using (which was also created as part of the environment. Nice and clean, and no credentials stored anywhere.
I created a Logstash configuration to grab files from S3 and process them, and then deployed it to the Logs Processor.
Permission to Rage Requested
The current release version of Logstash (1.4.2) does not support the usage of IAM roles for the S3 input. If I wanted to use that input, I would have to enter the credentials manually into the config file. I could do this easily enough at deployment time (storing the credentials in Octopus, which is much better than in source control), but I would need to actually have a user setup that could access the bucket. As the bucket is created during environment creation, this would mean that the credentials would change every time the environment was recreated. We create temporary environments all the time, so this would mean a manual step editing Octopus every time you wanted to get something to work.
I contemplated using a small script during deployment time to grab some credentials from the IAM role on the machine and enter them into the config file, but those credentials expire and Logstash was running as a service, so at some stage it would just stop working and someone would have to do something to make it work again.
Luckily for me, the wonderful people behind Logstash (and specifically the S3 plugin) have developed a new version that allows the usage of IAM roles, and it was already in beta. Its a little unstable still (Release Candidate 2), but it was good enough for my purposes.
While doing some reading about Logstash and the new version I discovered that the file input was basically completely broken on Windows. The component that it was leveraging to get the unique identifier for files in order to record the position in the file that it was up to does not work in 1.4.2 and below, so you end up missing huge chunks of data when processing multiple files. This actually explained why I was having so much difficulty using the earlier version to process a large amount of IIS logs from a disconnected machine, and why there was holes in my data. Long story short, if you’re using the file input in Logstash and you’re on windows, get the latest release candidate.
I incorporated the 1.5 RC2 release into my deployment, but I still couldn’t get the S3 input to work.
Why Is It Always A Proxy
I hate proxies.
Not because of what they are. I think they actually do some useful things, like caching, obfuscating where requests are coming from when accessing the internet from within a network and preventing access to bad websites.
No I hate proxies because the support for them is always a pain in the ass. Every application seems to support proxies differently, if they support them at all. Some automatically read from the Internet Explorer registry setting for the proxy, some use the HTTP_PROXY environment variable, some have their own personal settings. This means that every time you want to use a piece of software in an environment that uses a proxy, you have to fight with it to get it to work.
Such was the case with the S3 input. The underlying Ruby based aws-sdk has support for proxies (as does the .NET one), with a Set-AwsProxy method.
I could not, for the life of me, figure out how to configure Logstash with a proxy for the AWS component though.
So , I was stuck. I had all the configuration in place to process the ELB logs, but I didn't have the logs themselves.
In the end I created a small Powershell script that uses the AWS Powershell Component to move all files from an S3 bucket to a local directory on a timer. I then installed that script as a Windows Service using NSSM. Finally I edited my Logstash configuration to process the local files instead. After tweaking my config to process the files correctly, everything started coming through into the Log Aggregator as expected, and I added the missing piece to our intelligence about the service.
I don’t like this solution, because it adds more moving parts than I think is strictly necessary, but sometimes you have to make compromises.
I’ve uploaded a repository with my deployable build for Logstash here, so hopefully someone else can benefit from the effort that I put into making it re-usable.
Setting up a deployment pipeline for this component saved me a lot of time throughout the development process, making redeploying my changes when I made a mistake or needed to change a dependency (like upgrading to Logstash 1.5 RC2) a breeze. I highly recommend spending that initial bit of effort in setting things up at the start so you can move quickly later.
In regards to the actual ELB logs, they don’t provide any groundbreaking information that IIS didn't already give us, except for the case where connections are terminated at the ELB due to inactivity. At least to my knowledge anyway, I suppose they will track if the underlying instances go offline, which will be good. The ELB entries come in a bit slower than the IIS ones (due to the delay before the log files are published from the ELB + the delay added by my own S3 downloader and Logstash file processor pair), but there’s not really much I can do about that.
I still hate proxies. Only because its easier to hate one thing than every application that doesn’t support them.