Now, when we're tackling sentiment analysis using NLP technologies, we deal with unstructured data—customer chats, feedback on promotions or demos, and even media like images, audio, and video files. For processing such data, we rely on PySpark.
Beneath the surface, Spark functions as a compute engine with in-memory processing capabilities, enhancing performance through features like broadcasting and caching. It's become a crucial tool, widely adopted by 90% of companies for a decade or more.
AI libraries are the most valuable. They provide extensibility and usability. Spark has a lot of connectors, which is a very important and useful feature for AI. You need to connect a lot of points for AI, and you have to get data from those systems. Connectors are very wide in Spark. With a Spark cluster, you can get fast results, especially for AI.
The memory processing engine is the solution's most valuable aspect. It processes everything extremely fast, and it's in the cluster itself. It acts as a memory engine and is very effective in processing data correctly.
Manager - Data Science Competency at a tech services company with 201-500 employees
Feb 22, 2022
One of the key features is that Apache Spark is a distributed computing framework. You can help multiple slaves and distribute the workload between them.
Its scalability and speed are very valuable. You can scale it a lot. It is a great technology for big data. It is definitely better than a lot of earlier warehouse or pipeline solutions, such as Informatica.
Spark SQL is very compliant with normal SQL that we have been using over the years. This makes it easy to code in Spark. It is just like using normal SQL. You can use the APIs of Spark or you can directly write SQL code and run it. This is something that I feel is useful in Spark.
There could be enhancements in optimization techniques, as there are some limitations in this area that could be addressed to further refine Spark's performance.
The graphical user interface (UI) could be a bit more clear. It's very hard to figure out the execution logs and understand how long it takes to send everything. If an execution is lost, it's not so easy to understand why or where it went. I have to manually drill down on the data processes which takes a lot of time. Maybe there could be like a metrics monitor, or maybe the whole log analysis could be improved to make it easier to understand and navigate.
Its UI can be better. Maintaining the history server is a little cumbersome, and it should be improved. I had issues while looking at the historical tags, which sometimes created problems. You have to separately create a history server and run it. Such things can be made easier. Instead of separately installing the history server, it can be made a part of the whole setup so that whenever you set it up, it becomes available.
At times during the deployment process, the tool goes down, making it look less robust. To take care of the issues in the deployment process, users need to do manual interventions occasionally.
