Apache Kafka is an open-source stream processing platform developed by the Apache Software Foundation written in Scala and Java. The Kafka event streaming platform is used by thousands of companies for high-performance data pipelines, streaming analytics, data integration, and mission-critical applications.
N/A
Elasticsearch
Score 8.6 out of 10
N/A
Elasticsearch is an enterprise search tool from Elastic in Mountain View, California.
For brokering messages, Confluent Kafka is well suited since it offers a managed solution ready to use. Scenarios where the solution is not very well suited are for example, where pricing is an issue. The solution costs quite a lot for basic usage (for example: for 3 clusters, pricing is above 100k$ a year).
Elasticsearch is really well suited for searching text (Natural Language Processing) and you can fine tune the searches and scoring very well. I like the ability to find Significant Terms in the Index, where you can find aggregations that are really relevant to a specific search. It also allows for queries to lead to new queries via aggregations which is great for navigating your data. It is less suited to doing more complex aggregations where slices of data are required to be processing using guassian normalizations. And doing searches which join different documents is very very hard, and requires serious thought on how to denormalize data.
Apache Kafka is able to handle a large number of I/Os (writes) using 3-4 cheap servers.
It scales very well over large workloads and can handle extreme-scale deployments (eg. Linkedin with 300 billion user events each day).
The same Kafka setup can be used as a messaging bus, storage system or a log aggregator making it easy to maintain as one system feeding multiple applications.
The Kafka Tool is a community-made Java application that looks and feels from the past century.
Logging can be confusing. This certainly shows when we have to do troubleshooting.
Hybrid scenarios - pub/sub, but there are services in and outside a Kubernetes cluster. Then there are a ~3 options, but only 2 (the harder ones) are production-safe.
Setting Java memory thresholds can be a pain for those not accustomed to things like Eden Space & Old Generation which can lead to over allocation, or more likely, under allocation. Apache Solr had a similar issue. It would be nice if the program would take an extra step and dogfood it's own advice by analyzing the system & processes to return a solid recommendation for that configuration. The proper configuration information is outlined in the documentation, it would be nice if that was automated.
The only health check that ElasticSearch reports back is a "red" status without any real solid information about what is going on, though its usually memory thresholds or disk I/O. I am currently on ElasticSearch 1.5 so that may have changed for newer versions. When the status goes "red", I as the administrator of the software, feel like I lose control of whats going on which should rarely happen. Something more verbose would eliminate that.
This is more of a critique of the ElasticStack in general. The whole top to bottom stack is starting to get feature creep with things that are better suited in other software and increasing the barrier for entry for people to get started with setting up a robust logging infrastructure. ElasticSearch as a storage search engine, is pretty streamlined, but I can see that the tools that comprise the ELK Stack are going to require a certification with constant study at some point. During major release for Logstash a while back, it literally took a month to learn a new language because Elastic completely changed the syntax. For a medium sized organization of only a couple of admins, that is a pretty high bar where time is money. They really should work on refining/automating the tools & search engine they have, instead of shoehorning/changing things on to an already rock solid foundation.
Kafka has suited our use case very well so far. Going forward we are planning to expand our platform manifold so the load on Kafka and our reliance on Kafka is going to increase only.
Apache Kafka is highly recommended to develop loosely coupled, real-time processing applications. Also, Apache Kafka provides property based configuration. Producer, Consumer and broker contain their own separate property file
To get started with Elasticsearch, you don't have to get very involved in configuring what really is an incredibly complex system under the hood. You simply install the package, run the service, and you're immediately able to begin using it. You don't need to learn any sort of query language to add data to Elasticsearch or perform some basic searching. If you're used to any sort of RESTful API, getting started with Elasticsearch is a breeze. If you've never interacted with a RESTful API directly, the journey may be a little more bumpy. Overall, though, it's incredibly simple to use for what it's doing under the covers.
Support for Apache Kafka (if willing to pay) is available from Confluent that includes the same time that created Kafka at Linkedin so they know this software in and out. Moreover, Apache Kafka is well known and best practices documents and deployment scenarios are easily available for download. For example, from eBay, Linkedin, Uber, and NYTimes.
We've only used it as an opensource tooling. We did not purchase any additional support to roll out the elasticsearch software. When rolling out the application on our platform we've used the documentation which was available online. During our test phases we did not experience any bugs or issues so we did not rely on support at all.
Apache Kafka is built for scale. From high throughput and real-time data streaming, it has a strong advantage over RabbitMQ with its low latency. This put Apache Kafka at the forefront as the platform of choice for large datasets messaging and ensuring scalability when data scale up tremendously. RabbitMQ however has its strengths in traditional messaging. Routing and message delivery reliability are the bedrock of RabbitMQ and this is where RabbitMQ excels. In my previous workplace, RabbitMQ was of choice as reliability matters more than scale. In two words. Apache Kafka for scale, RabbitMQ for reliability. And for cloud deployment and large dataset messaging in what I am doing now, Apache Kafka is the default choice.
Elasticsearch is the most well-known and supported free data platform that we identified. We are taking advantage of community knowledge and practices. In terms of flexibility and breadth of use cases no other competitor came close to Elasticsearch. We've tried Solr in the past be we encountered issues which were deal-breaking for us. MongoDB - it just did not pass our evaluation parameters as a main data platform. We still use it for smaller purposes, though.
Positive: bursts of traffic on special holidays are easy to handle because Kafka can absorb and buffer all the messages we need to process long enough to let an understaffed set of back-end services catch up on processing. Hard to put a number to it but we probably save $5k a month having fewer machines running.
Positive: makes decoupling the web and API services from the deeper back-end services easier by providing topics as an interface. This allowed us to split up our teams and have them develop independently of each other, speeding up software development.
Negative: our engineers have made mistakes such as accidentally dropping a few thousand messages due to the CLI being confusing to use, and as a result a customer lost some of their precious data. I'd say that was more our fault than Kafka's though.
I am not in finance and I suspect even if I was this would be hard to measure. But for sure, Elasticsearch has enabled us to have the most flexible data model in the industry for our customer's data, and in doing so we have attracted many many technical customers and got much of their $$$.
One problem with Elasticsearch is that because it runs on the JVM, there can be some stop-the-world JVM garbage collections happening that can take down nodes and reduce indexing speed. The solution for that tends to be "let's just upgrade the CPU on that machine". And before you know it you are paying $$$ because this'll happen with 40+ machines.
On the other hand, I do think that ES is more efficient than other systems and so it requires fewer nodes to keep it highly tolerant and available, so we probably saved some money that way.
