Apache Kafka Reviews

Currently, I work for an observability company. We stream customer data into our cloud, digest the information, enrich it, transform it, save it, and use on-the-fly aggregation with Kafka. Previously, I worked for a security company doing normal detection using streaming with Kafka. 

I also worked for a company with a data platform based on Kafka, where we ingested clickstream data and enriched it before streaming.

The most valuable feature of Kafka is the Kafka Streams client. Unlike other systems like Flink or Spark Streaming, you don't need a separate engine to do real-time transformations and analytics. The amount of data that can be streamed into the platform and the scalability are also significant benefits.

Kafka requires fine-tuning to find the best architecture, number of nodes, and partitions for your use case. It’s a trial-and-error process with no one-size-fits-all solution. Issues may arise until it’s appropriately tuned. 

While it can scale out efficiently, scaling down is more challenging, making deleting data or reducing activity harder.

I have been working with the Kafka product for more than ten years.

Since Kafka is written in Java, it's not as stable as it should be on the JVM. The stability depends on fine-tuning the system to find the best architecture for your use case. However, the replication factor helps avoid data loss despite the stability issues.

Kafka's architecture allows for scalability by adding nodes and partitions to topics. However, it's not as effective in scaling in, making reducing activity and deleting data harder. 

Scalability can be managed both manually and automatically to meet demands.

I used to work with Spark Streaming and Flink, however, not in the past year.

If you are unfamiliar with Kafka, setting up the cluster can be quite difficult. You need to understand the architecture and components and compute the data volume upfront. For experienced individuals, the setup is less difficult yet still requires preparation.

From a time-saving perspective, onboarding new customers is straightforward, requiring them merely to stream their data into our platform.

We use Apache Kafka, which is open-source, so we don't have fees. I can't comment on ownership costs as I am not responsible for that domain.

Apart from Kafka, I have experience working with Spark Streaming and Flink.

When implementing Kafka, it's important to plan the cluster size upfront to ensure easy scalability. Adding or removing nodes can disrupt the clusters, so proper sizing and planning are key. 

I would rate Kafka as a solution as a nine.

Kafka is used as a streaming platform where multiple producers and consumers exchange high load and high volume of messages asynchronously without affecting each other's performance. 

It serves as an industry-standard platform for such operations. Kafka is also integrated into data system architecture for applications like monitoring events on platforms like LinkedIn to enable further analytical insights.

Kafka makes data streaming asynchronous and decouples the reliance of events on consumers. 

It was the first of its kind to provide a streaming pipeline, setting a new component in the tech architecture and ecosystem. It allows continuous messaging without impacting performance.

Confluent has improved aspects like documentation and cloud support, yet Kafka's reliance on older architectures like ZooKeeper in previous versions is a limitation. 

Its language and architecture could be further improved to solve issues in consensus algorithms, as Red Panda does.

I have been working with Kafka for about a year and feel comfortable using it.

I have not had any issues in terms of performance; however, there may be performance issues due to Java's garbage collector, which can cause memory issues if bloated.

While I have not tried setting Kafka up on Docker containers, it is possible. I have only run a single-node broker for Kafka.

I also use Redpanda, which is similar to Kafka in features, however, they differ in internal workings affecting performance and resource usage.

The setup process is straightforward as per the documentation. It involves unpacking zip files with the necessary packages, ensuring Java and JVM are installed. 

Previously, Kafka relied on ZooKeeper, requiring two configuration files. However, with the newer KRAP version, the setup does not need ZooKeeper, which simplifies the process.

Apache Kafka was part of a college curriculum, and I set IP up myself. I found setting it up manageable.

I definitely recommend Kafka, as it is the industry standard for streaming platforms. While Red Panda is similar, Kafka remains the stronger choice in the market for its established support and usage in big companies.

I'd rate the solution nine out of ten.

We are always using Apache Kafka for our real-time scenarios. It helps us detect anomalies and attacks on our website through machine learning models.

We are managing our data by topics. Splitting topics is more effective for us. Apache Kafka is very fast and stable. It offers scalability with ease and also integrates well with our tools. Fault tolerance is a good feature, and it also has high throughput rates.

Config management can be better. We are always trying to find the best configs, which is a challenge.

I have been working with Apache Kafka for more than four years. It has been used since the beginning of our department, maybe six years.

It is very stable and meets our needs consistently.

If there is latency, our Kubernetes admin includes our Kafka nodes to increase scalability. Kafka provides flexibility and integrates easily with Kubernetes.

Before Apache Airflow, I used Cron Tab. However, Apache Airflow makes it easy to follow and manage tasks, and data science departments can easily build their models or pipelines using it.

I would rate Apache Kafka nine out of ten.

I was planning to use the tool for real-time analysis in terms of data processing and real-time analytics workflows. The real-time IoT data comes through with a few challenges, and that is for one time, so it is more like a Kafka topic. I want to actually use multiple Kafka topics where one of them can be directly fed into the data pipeline, another one can be fed into the real-time alert system, and the next one can be fed into machine learning.

The most valuable features of the solution revolve around areas like the latency part, where the tool offers very little latency and the sequencing part. The sequencing part actually helps to aggregate things in a way that I don't need to write another function or kind of sequence it, and I write an aggregate function to figure out the maximum value in the last ten samples.

One complexity that I faced with the tool stems from the fact that since it is not kind of a stand-alone application, it won't integrate with native cloud, like AWS or Azure. Apache Kafka has another mask on it, so if users can have a direct service, like Grafana, that can actually be used as a stand-alone tool with Grafana cloud, or you can use a mix of AWS and Grafana, so there is not much difference with it. I expect Apache Kafka to have Grafana's same nature.

The product's support and the cloud integration capabilities are areas of concern where improvements are required.

I have been using Apache Kafka for a year.

Stability-wise, I rate the solution an eight out of ten.

Scalability-wise, I rate the solution an eight out of ten.

Around four people in my company use the product.

I did not interact much with the product technical support team. I did not have dedicated support that responded to all my queries since I was using the product's free version. I rate the support a seven out of ten.

Neutral

I have worked with Databricks. I use Databricks and Apache Kafka simultaneously.

The product's deployment phase is neither complex nor straightforward. As the software has evolved a lot, users can actually keep it even simpler by opting for a plug-and-play model.

The solution is deployed on an on-premises model.

The solution can be deployed in two or three days.

I was involved with the tool's installation process.

I cannot comment on the tool's ROI since I did not use it for production purposes.

I was using the product's free version.

I did not come across any scenarios involving fault tolerance because when it comes to the issue data consistency issues, like missing or incorrect value of data are actually part of the system where the data is being fed. Nevertheless here, when it comes to the missing values, I never tried the option, especially whenever a value is missing, that can allow one to impute the value with another parameter.

Speaking about if I incorporated any emerging data stream streaming trends in Apache Kafka workflows, for example, utilization of AI, I would say that I use it as a local system, so if I have an EC2 server where I kind of read the sample and then populate the regression and reintegration model on top of it, but that is done locally and not on the cloud.

I recommend the product to those who plan to use it. I like Kafka and Flink, and I want to actually create a system in AWS mainly for real-time streaming so that I don't need to worry about multiple data copies.

Considering the improvements needed in the product's support, and the cloud integration capabilities, while looking at the simplicity during the installation phase, I rate the tool a seven out of ten.

We have different use cases for different clients. For example, a banking sector client wants to use Apache Kafka for fraud detection and messaging back to the client. For instance, if there's a fraudulent swipe of a credit or debit card, we stream near real-time data (usually three to five minutes old) into the platform. Then, we use a look-up model to relay the data to the messaging queue to the customers. This is just one use case.

We also have data science engineers who use Kafka to feed on the data (usually within the last five to seven minutes of transactions) to detect any fraudulent transactions for internal consumption of the bank. This is not for relaying back to the customer. This client is based in the Middle East, and they have some interesting use cases.

We are still in the cluster database phase. Based on the use cases captured during the advisory phase, there will be a mix of 40 to 60% of users. 40% will be internal data science and IT teams, while 60% will be end users. So, the total number of users we have seen is 25. Out of these, around 15 business users will make decisions based on reports generated by Kafka analytic data. The remaining users are internal, who analyze this data daily to identify more use cases from a predictive and AI perspective for the future banking domain.

Moreover, our current client is an enterprise business. It is a globally renowned bank that has entered Saudi.

One of the major features that we are currently exploring, which is coming from my previous experience as well, is a multiple PubSub model kind of architecture. We have one hub data, which is the IBM DB2 system that banks use for their daytime transaction tracking from OLTP systems. We want to use the data on different platforms. So, we are trying to use Kafka in a model where it will be a publisher onto multiple messaging queues. These different messaging queues belong to different business units, where we are segregating the data lake we are building into different domains. For example, HR data is becoming too sensitive, so they don't want to give it to any other businesses. We are working on a common publisher and multiple subscriber model, which I feel is much more easily implementable using Kafka.

The other part that we are trying to implement, and which is in its very near center stages, is to see if we can make it future-ready. Right now, in the Middle East, there are not many cloud subscribers like DCP AWS and Azure. It is all on-premise. But it'll be there just in the next two or three years. So, we are trying to see if we can have these Kafka models working from a future perspective wherein instead of dumping some of the data into a data lake, we can directly dump it into solutions like DCP BigQuery for real-time analytics. This is just for the use cases, which are for real-time analytics. This data will definitely also be there in the data lake as that is the intention of keeping it. 

But using Kafka, we are trying to see if we can make these subscribers ready to use these DCP BigQuery platforms for real-time analytics. It's still in the remittance stages, but those are still use cases.

One of the major areas for improvement, which I have to check out, is their pulling mechanism. Sometimes, when the data volume is too huge, and I have a pulling period of, let's say, one minute, there can be issues due to technical glitches, data anomalies, or platform-related issues such as cluster restarts. These polling periods tend to stop messaging use, and the restart ability part needs to be improved, especially when data volumes are too high. 

If there are obstructions due to technical glitches or platform issues, sometimes we have to manually clean up or clear the queue before it eventually gets sealed. It doesn't mean it doesn't get restarted on its own, but it takes too much time to catch up. At that point, one year ago, I couldn't find a solution to make it more agile in terms of catching up quickly and showing that it is real-time in case of any downtime. 

This was one area where I couldn't find a solution when I connected with Cloudera and Apache. One of our messaging tools was sending a couple of million records. We found it tough when there were any cluster downtimes or issues with the subscribers consuming data. 

For future releases, one feature I would like to see is a more robust solution in terms of restart ability. It should be able to handle platform issues and data issues and restart seamlessly. It should not cause a cascading effect if there is any downtime.

Another feature that would be helpful is if they could add monitoring features as they have for their other services. A UI where I can monitor the capacity of the managed queue and resources I need to utilize more to make it ready for future data volumes. It would be great to have analytics on the overall performance of Kafka to plan for data volumes and messaging use. Currently, we plan the cluster resources manually based on data volumes for Kafka. If they can have a UI for resource planning based on data volume, that could be a great addition.

I have been using Apache Kafka for five years. In the current project, we're setting up a cluster. We'll be doing the service installations next week. It's a private cloud-based implementation, and I'm leading the end-to-end implementation. In my previous project, we mainly used Kafka for streaming real-time SAP data into the analytics platform for a technology client.

But for the current banking sector client, we're setting up a 58-node cluster and reserving six nodes for Kafka because we have a lot of streaming use cases.

I would rate the stability of Apache Kafka a six out of ten due to the polling period and high data volumes; there is a catch-up problem. If there is a five-minute downtime, it can have a cascading effect. 

When it comes to data availability, that is, how available the data is on the messaging queue, I would rate it a little less due to the coding mechanism and data getting stuck when the data volumes are high.  From the data availability perspective, I would rate it between six to seven. The major reason for this is the ransomware data that comes in day in and day out. If the resources are not allocated correctly to the Kafka messaging queue, sometimes it gets stuck. And once it is stuck, it can have a steady effect on catching up to the real-time data. Only because of this issue, I rate it between six to seven. 

However, from an overall data security perspective and ensuring that the data is consistent across the system, I would rate it around nine. If your PubSub model is written correctly, you can be assured that the data will not be lost. It will either be in the messaging queues or your landing tables or staging tables, but it will not be lost, at least if you have written it correctly.

I would rate the scalability of Apache Kafka somewhere around seven out of ten. I'm not going on the higher side because a lot of manual work is involved in upgrading Kafka. You have to estimate the overall capacity, not just in data but also in other use cases running on the same cluster. In the cloud, it is easier as you don't have to worry about the turnaround time of your cluster setup. But in an on-premise setup, you need to add more nodes, RAM resources, and storage based on the increasing data volume. 

Also, it would help if you ensured that the streaming use cases running on Kafka are not impacting other use cases like batch or archival use cases. Because of this manual activity of overall estimation, I will still keep it somewhere around six to seven. But regarding scalability in terms of horizontal or vertical scalability from the data perspective, I feel more comfortable with Kafka compared to other available streaming solutions.

I have noticed a drastic improvement in the last five to seven months. Last year, the turnaround time for certain cases was around 14 days unless you escalated. Issues used to take two, three, or even four times to follow up on. Even though the solutions provided were often resource upgradation solutions, which I felt were not always the best. 

However, in the last month, I have seen Cloudera coming through with two to three days turnaround times, even for low-severity issues. I'm unsure if it is region-specific, but I assume it should be, as they have region-specific teams. Sometimes, however, you cannot always depend on them. The type of solutions provided are sometimes like hidden trials. When you work for bigger enterprises, you cannot always go with them because there is a cost associated.

For example, in one of my recent cases, we implemented the engine policy, a security setup on our cluster. We were stuck at some point and asked for technical support. They provided a solution that was just a patchwork. When we did our analysis and went to the bank security team to review it, we found out that their solution was inadequate. They told us to set up role-based access control through Ranger, where AD users should be synced with the Ranger, and access control policies should be set up. However, they provided only for the local range level if you have Linux users. That is not a solution because, at the enterprise level, everything is integrated with AD and authenticated by AD.

I would rate Cloudera's support on a scale of five to six. But from the turnaround time, there has been an improvement from last year. We used to wait two to three days for critical solutions, but now it is much better. I used to work for the US region in my previous project, and the turnaround time was not as expected. It all depends on the licensing, and if you have a premium license from Cloudera, they assign a professional services guide to your project, and you get better support. If you do not have a premium license, you have to go through the process of rating the cases and wait for their support team to come. Overall, it is not at par with them if you compare it with solutions like Azure DCP and others.

Neutral

The initial two months were for capacity estimation, where we worked with the client's different business teams to understand the data volumes and use cases. Then, the next four to five months went into procurement, where we had to work with infrastructure teams and vendors to understand the servers and networks required for the cluster. 

The actual cluster setup took us two months, and it was a little longer due to a shortage of expertise on the client's networking team. We had to handle everything ourselves since it was an on-premise setup with physical servers and network connections. Currently, we are in the security review phase, and once it completes, we will start implementing various use cases like task and batch processing, archival, etc.

If you see my experience from the Apache Kafka implementation and clusters as a perspective, I will rate the setup somewhere between seven to eight out of ten.

We deployed the solution On-premises because it is a Middle East client. That is where my admin experience in the last two years has been too much. So even if I move to the cloud, it'll be much easier because I have seen cluster implementation from scratch and how it is done. So I have been involved in the very first stage of working with Cloudera on the sizing and then working on the actual infrastructure networking team on the implementation, working with the network team on doing all the network structuring, then setting up the cluster ourselves. I have a team of around seventeen people who am I getting here. So yeah, from that perspective, it is there, but what we are implementing is Cloudera private cloud-based as a solution, which is a future-ready solution for the cloud. 

So once cloud services enter with least, especially Saudi Arabia, for example, the CPaaW as an Azure, our cluster will always be ready to be upgraded to the cloud because there's a private cloud-based solution on which we have the cluster. We can anytime add the cloud-native hosts and nodes onto our cluster. Also, at the same time, because it's a banking client, it has some restrictions in terms of geographies and all for the data to decide. The physical cluster provides a solution from the future AD perspective that once TCP, Azure, and AWS set up their data centers in Saudi Arabia, we can have some of our data nodes in the bank data center, plus we can have other sets of nodes or VMs in the cloud service provider's data center.

I have seen that the ROI is very good when implemented correctly and used for a period of time. I have seen, from a POC perspective, data getting churned in a couple of months, and the amount of insights generated was overwhelming. 

I have also seen some critical decisions taken based on that data at an enterprise level, which earlier used to take years. Because of the time it used to take, the intention of doing those analytics used to lose its flavor. But now, people can make decisions in a few months based on this streaming analytics use case through Kafka. And they see, if those decisions had been taken earlier, they could have quickly gone on for four to five percent of their year-on-year profits. 

However, too many things are involved because of the overall use case perspective, data perspective, underlying cluster sizing, and sourcing. One has to think from a holistic point of view, not just from a business point of view.

I have experience in private cluster implementation. When you use Apache Kafka with Cloudera, the pricing is included in your Cloudera license. The pricing is based on the number of nodes, the storage cost, and other components. As part of this license, Kafka is one of the solutions offered. When you compare it with OnCloud, if you don't have a good volume of data and use cases, your benefits realization will not be there, as the initial cost of setting up the cluster and bringing up the license can be as much as $760k for a small cluster of ten to twenty nodes. You need at least 20-30 GBs of data and use cases before utilizing and profiting from the Teradata license and cloud data. Kafka is just one piece of it. 

When it comes to the cloud, the pricing also goes at the solution level so that you can compare it at the Kafka level. Still, I don't have much information on that from where I am currently implementing the solution. After we did the cost-benefit analysis, we only opted for the solution. We realized that by bringing in Cloudera along with Kafka, we would be able to replace two or three existing systems, including Teradata, Oracle, Informatica, and IBM Datastage. Only then were we able to realize the benefit for the bank. Otherwise, Cloudera would be much more expensive, especially in the short term. With distributed computing, the concept of Delta Lake is coming in, and IDBMS systems like Teradata and distributed systems like data lakes will coexist. Not all use cases will be solved, but cloud solutions like Azure come as a package, and you need not worry about having different physical systems in your enterprise to take care of. That's where I think the cost-benefit analysis from a data perspective becomes too important.

At the end of the day, we bring in big data systems only when the data volumes are high. When the data volumes are low, the cost-benefit analysis can easily show that systems like Oracle or Teradata can run it just fine.

From an architecture and solution design perspective, I would say that before going for streaming solutions, we should analyze the data, which might be old, and decide if it's a streaming use case or not. Often, people think it's a streaming use case, but when they perform analytics on top of it, they realize they can't do a month-to-date or year-to-date analysis. So, it's essential to think again from the data basics perspective before going to Kafka.

Overall, from the product and solution perspective, I would rate it a nine based on my personal use of data.

Apache Kafka is a messaging solution where you have topics to pass on your information. You can send messages to multiple topics.

We need to manage limited resources. Additionally, we can send or push messages through a specified port. This is a significant feature because, unlike traditional queues, Kafka uses a cluster of nodes, making it easy to integrate with various algorithms. This clustering is an advantage and a key feature of Kafka, providing good interaction and scalability.

For example, when you want to send a message to inform all your clients about a new feature, you can publish that message to a single topic in Apache Kafka. This allows all clients subscribed to that topic to receive the message. On the other hand, if you need to send billing information to a specific customer, you can publish that message on a topic dedicated to that customer. This message can then be sent as an SMS to the customer, allowing them to view it on their mobile device.

Apache Kafka is different in its design. If you have topics around the front end of clusters in the facility, it is scalable. The software is scalable to handle and process data. However, it might not be suitable for handling specific types of images or media files. Other than that, it should handle the rest of the data processing needs.

There are no multiple versions, which simplifies the process of granting access with Kaspersky. Every message is accurately delivered. However, Kafka does not support sending messages directly. You need to publish messages finalization. If you want to resend a message, you must resend it manually. Kafka does not automatically handle this. Another thing is the need for a redo option if an issue occurs. If a message is not sent properly, it can be retransmitted within the core system. You should enable the gateway in your program for it to function correctly. Messages will not be delivered or refreshed unless you enable the direct replay option in the product settings.

I have been using Apache Kafka since 2020-21

The initial setup of Apache Kafka is challenging and requires experience. Each message should always receive a response, so prioritizing traffic is essential. Furthermore, the client or consumer must always be in sync, or the message will not be processed.

One pair of nodes is sufficient for the system. If our other system requires more than five nodes, it might not be feasible. Currently, other components are functioning as expected. The Kafka setup won't take much time.

When using Apache Kafka, it’s important to manage different environments carefully to avoid confusion. For instance, you can configure different client applications for producing and consuming messages. Ensure that the configurations for each environment (development, testing, production, etc.) are separated. This includes managing source code and data appropriately to maintain security and efficiency. Proper management of Kafka assets and operations phases is crucial for a smooth workflow.

I recommend Apache Kafka since it is extremely fast, stable and has been used for a very long time. We haven't encountered any major issues or concerns regarding its performance and customer service.

Overall, I rate the solution a nine out of ten.

I have previous professional experience using Kafka to implement a system related to gathering software events in one centralized location. 

One example of how Kafka has been beneficial is in the context of financial trading. When a trade is executed, it generates an event. I used Kafka to create an application that captures these events and stores them in a topic, allowing for efficient processing in real time.

Regarding the most valuable feature in Kafka, I would say it's scalability. The publisher-subscriber pattern and low latency are also essential features that greatly piqued my interest.

Maintaining and configuring Apache Kafka can be challenging, especially when you want to fine-tune its behavior. It involves configuring traffic partitioning, understanding retention times, and dealing with various variables. Monitoring and optimizing its behavior can also be difficult.

Perhaps a more straightforward approach could be using messaging queues instead of the publish-subscribe pattern. Some solutions may not require the complex features of Apache Kafka, and a messaging queue with Kafka's capabilities might provide a more complete messaging solution for events and messages.

I have been using Apache Kafka for the past 10 years. 

The stability may improve if the configuration and management aspects become less challenging.

It depends on the configuration., but scalability is one of the best features of Kafka. I would rate it nine out of ten.

Support can vary depending on whether you're using the open source version or a paid one. Our version, the paid console version, offers highly available support, and you can find a wealth of information and assistance from various providers online. However, when I used MSA on AWS, I encountered limited support for it.

Neutral

Despite the challenges we faced with configuration and management, I believe the return on investment is safeguarded.

The cost can vary depending on the provider and the specific flavor or version you use. I'm not very knowledgeable about the pricing details.

I believe that when working with Kafka Apache, it's essential to have a specialist who thoroughly understands and can optimize all the available variables within the solution to achieve the desired behavior.

I would rate it an eight out of ten.

We use Kafka for Elastic Stack and Kafka SCRAM login.

I have many users of Apache Kafka. It's like a subject to study in enterprises. However, we have not decided if the systems should generalize Apache Kafka for every application and every IT system.

We use Kafka for mapping and ThoughtSpot data from one IT system source to the destination. We also prefer it to exchange data from our internal IT systems.

Kafka is a new method we opted to apply to our need for data exchange. Also, we use the solution's integration capabilities.

Irovement-wise, I would like the solution to have more integration capabilities. Also, the solution's setup, which is currently complex, should be made easier.

I have experience with Apache Kafka.

It is a stable solution.

A lot of my experience indicates that Apache Kafka is scalable. We can have ten or even fifty hundred users on the solution. So, it's possible because we are a big enterprise.

I have experience with Apache Kafka's technical support.

The solution's initial setup process was complex. The deployment process took three or four years.

Right now, I can't deliver the planning process required for deployment.

For deployment and maintenance, we have a manager and an operational person. However, I can't give an exact count of the people required for deployment and maintenance.

To be able to recommend Kafka to others, especially considering every context, we will have to set a benchmark and compare Kafka with other tools.

I rate the overall solution a seven out of ten.