It is a common requirement to export the data in Elasticsearch for users in a common format such as .csv. An example of this is exporting syslog data for audits. The easiest way to complete this task I have found is to use python as the language is accessible and the Elasticsearch packages are very well implemented.

In this post we will be adapting the full script found here.

POST logs/_doc
{
  "host": "172.16.6.38",
  "@timestamp": "2020-04-10T01:03:46.184Z",
  "message": "this is a test log"
}

2. Note that when there is a failure to write to the file, it will write the message to a array to print back.

3. At the end of the script, all the failed messages will be re-printed to the user

Solving a 2 year old problem

Logs are usually sent via UDP traffic and most commonly available as a syslog message:

• UDP: Best effort process-to-process based communication
• TCP: Reliable host-to-host based communication

These logs in many products (especially SIEM – Security information and event management) have their sources identified using the source-IP of the packet instead of the content of the message (Often the content of the log does not contain source identification information, this is a result of poor logging design), as such in more complex topologies with log caching or staged log propagation, the source of the logs cannot differentiated by the final system.

input {
  generator { message => "Hello world!" count => 1 }
}
filter {
  mutate {
   add_field => {
      "extra_field" => "this is the test field"
      "src_host" => "3.3.3.3"
   }
   update => {"message" => "this should be the new message"}
  }
}
output {
  spoof {
    dest_host => "<REPLACE WITH YOUR DESTINATION IP>"
    dest_port => "<REPLACE WITH YOUR DESTINATION PORT>"
    src_host => "%{src_host}"
    src_port => "2222"
    dest_mac =>  "<REPLACE WITH YOUR DESTINATION MAC ADDRESS>"
    src_mac => "<REPLACE WITH YOUR MAC ADDRESS>"
    message => "%{message}"
    interface => "ens32"
  }
}

Summary

When running Logstash in large scale environments it can be quite difficult to troubleshoot performance specifically when dealing with UDP packets.

The issue could occur at multiple layers, in order of dependent layers of concern:

• Infrastructure
• Logstash Application
• Pipeline

The following steps assume installation of Logstash on a Linux machine (CentOS 7.4) but similar steps can be used for other machines.

journalctl -u logstash.service

1. Docker
2. Elasticsearch
3. Logstash

Why Log to Elasticsearch?

Elasticsearch is a fantastic tool for logging as it allows for logs to be viewed as just another time-series piece of data. This is important for any organizations’ journey through the evolution of data.

This evolution can be outlined as the following:

1. Collection: Central collection of logs with required indexing
2. Shallow Analysis: The real-time detection of specific data for event based actions
3. Deep Analysis: The study of trends using ML/AI for pattern driven actions

Data that is not purposely collected for this journey will simply be bits wondering through the abyss of computing purgatory without a meaningful destiny! In this article we will be discussing using Docker to scale out your Logstash deployment.

The Challenges

If you have ever used Logstash (LS) to push logs to Elasticsearch (ES) here are a number of different challenges you may encounter:

• Synchronization: of versions when upgrading ES and LS
• Binding: to port 514 on a Linux host as it is a reserved port
• High availability: with 100% always-on even during upgrades
• Configurations management: across Logstash nodes

When looking at solutions, the approach I take is:

• Maintainability
• Reliability
• Scalability

The Solution

Using Docker, a generic infrastructure can be deployed due to the abstraction of containers and underlying OS (Besides the difference between Windows and Linux hosts).

Docker solves the challenges inherent in the LS deployment:

• Synchronization: Service parallelism is managed by Docker and new LS versions can be deployed just by editing the Docker Compose file
• Port bindings: The docker service is able to bind on Linux hosts port 514 which can be forwarded to the input port exposed by your pipeline
• High availability: Using replica settings, you can ensure that a certain amount of LS deployments will always be deployed as long as there is always one host available. This makes workload requirements after load testing purely a calculation as follows:

1. Max required logs per LS instance = (Maximum estimated Logs)/(N Node)
2. LS container size = LoadTestNode(Max required logs per LS instance)
3. Node Size = LS container size * 1 + (Max N Node Loss)

I.e. Let’s say you have 1M logs required to be logged per day, and have a requirement to have 3 virtual machines for a maximum of 1 virtual machine loss.

1.  333,333 = 1M / 3
2. Assume 2 CPU, 4 GB RAM
3. 4 CPU, 8 GB = (2 CPU , 4 GB RAM) * 2

Why not deploy straight onto OS 3 Logstashes sized at 4 CPU and 8 GB RAM?

• Worker number does not scale automatically when you increase the CPU count and RAM, as such even with more CPU and RAM you cannot ensure that you will effectively be able to continue ingesting the same amount of logs in a “Node down” situation
• Upgrading the Logstashes become a OS level activity with associated headache
• Restrictive auto-scaling capability
• No central stout view of Logstashes

Architecture

Let’s take a look at how this architecture looks,

When a node goes down the resulting environment looks like:

A added bonus to this deployment is if you wanted to ship logs from Logstash to Elasticsearch for central and real-time monitoring of the logs its as simple as adding Filebeats in the docker-compose.

What does the docker-compose look like?

version: '3.3'

1. Estimate the Virtual Machines (VM) sizes and LS sizes based on estimated ingestion of logs and required redundancy
2. Deploy VM and install docker 
3. Create a docker swarm
4. Write a logstash.yml and either include the pipeline.yml or if you are using ES configure centralised pipeline management.
5. Copy the config to all VM’s to the same file location OR create a shared file system (that is also HA) and store the files there to centrally manage config
6. Start your docker stack after customizing the docker-compose file shared in this article
7. Set up a external load balancer pointing at all your virtual machines
8. Enjoy yummy logs!

The BUT!

As with most good things, there is a caveat. With Docker you add another layer of complexity however I would argue that as the docker images for Logstash are managed and maintained by Elasticsearch, it reduces the implementation headaches.

In saying this I found one big issue with routing UDP traffic within Docker.

This issue will cause you to lose a proportional number of logs after container re-deployments!!!

• What is your current logging deployment?
• Have any questions or comments?
• Are interested seeing a end-to-end video of this deployment?
• Comment below!

Disclaimer: This article only represents my personal opinion and should not be considered professional advice. Healthy dose of skeptism is recommended.

If you are currently developing using C# (Particularly .NET Core 2.0+) here are some shortcuts I hope will be able to save you time I wish I could have back.

There is official documentation for C# Elasticsearch development however I found the examples to be quite lacking. I do recommend going through the documentation anyway especially for the NEST client as it is essential to understand Elasticsearch with C#.

1. Low Level Client

“The low level client, ElasticLowLevelClient, is a low level, dependency free client that has no opinions about how you build and represent your requests and responses.”

– ElasticSearch Official Documentation

Unfortunately the low level client in particular has very sparse documentation especially examples. The following was discovered through googling and painstaking testing.

1.1. Using  JObjects in Elasticsearch

JObjects are quite popular way to work with JSON objects in .NET, as such it may be required to parse JObjects to Elasticsearch, this may be a result of one of the following:

• Definition of the object is inherited from a different system and only parsed to Elasticsearch via your application (i.e. micro-service)
• Too lazy to strongly define each object as it is unnecessary

The JObject cannot be used as the generic for indexing as you will receive this error:

Instead use “BytesResponse” as the <T> Class

1.2. Running a “Bool” query

The examples given by the Elasticsearch documentation does not give an example of a bool query using the low-level client. Why is the “Bool” query particularly difficult? Using Query DSL in C#, “bool” will automatically resolve to the class and therefore will throw a error:

Not very anonymous type friendly… the solution to this one is quite simple, add a ‘@’ character in-front of the bool.

1.3. Defining Anonymous Arrays

This one seems a-bit obvious but if you want to define an array for use with DSL, use the anonymous typed Array (Example can be seen in figure 4) new Object[].

1.4. Accessing nested fields in searches

Nested fields in Elasticsearch are stored as a full path, . delimited string. This creates a problem when trying to query that field specifically as it creates a invalid type for anonymous types.

The solution is to define a Dictionary and use the dictionary in the anonymous type.

The Dictionary can be passed by the anonymous type and will successfully query the Nested field in Elasticsearch.

2. NEST Client

“The high level client, ElasticClient, provides a strongly typed query DSL that maps one-to-one with the Elasticsearch query DSL.”

–ElasticSearch Official Documentation

The NEST documentation is much more comprehensive, the only issue I found was using keyword Term searches.

2.1. Using Keyword Fields

All string fields are mapped by default to both text and keyword, the documentation can be found here. Issue is that in the strong typed object used in the Elastic Mapping there is no “.keyword” field to reference therefore a error is thrown.

Example:

For the Object:

public class SampleObject
{
public string TextField { get; set; }
}

Searching would look like this

Unfortunately the .Keyword field does not exist, the solution is using the .Suffix function using property name inference. This is documented in the docs however it is not immediately apparent that is how you access “keyword”.

I hope this post was helpful and saved you some time. If you have any tips of your own please comment below!