Amazon Kinesis vs. Apache Spark Streaming vs. Google Cloud Dataflow

Amazon Kinesis is a great replacement for Kafka and it works better whenever the components of the solution are AWS based. Best if extended fan-out is not required, but still price-performance ratio is very good for simplifying maintenance.

I would go with a different option if the systems to be connected are legacy, for instance in the case of traditional messaging clients.

Incentivized

Apache Spark Streaming is a tool that we are using for almost a year and is excellent in managing batch processing. It is user-friendly. Using it, we can even process our massive data in fractions of seconds. Its pricing is its other plus point. Only its In-memory processing is its demerit as it occupies a large memory.

Incentivized

It is best in cases where you have batch as well as streaming data. Also in some cases where you have batch data right now and in future you will get streaming data. In those cases Dataflow is very good. Also in cases where most of your infra is on GCP. It might not be good when you already are on AWS or Azure. And also you want in-depth control over security and management. Then you can directly use Apache beam over Dataflow.

Incentivized

• Processing huge loads of data
• Integrating well with IoT Platform on Amazon
• Integration with overall AWS Ecosystem
• Scalability

Incentivized

• It is amazing in solving complicated transformative logic.
• It is straightforward to program.
• It is a very quick tool.
• It processes large data within a fraction of seconds.

Incentivized

• Streaming, Real time work load
• Batch processing
• Auto scaling
• flexible pricing

• Not a queue system, so little visibility into "backlog" if there is any
• Confusing terminology to make sure events aren't missed
• Sometimes didn't seem to trigger Lambda functions, or dropped events when a lot came in

Incentivized

• There must be more documentation.
• It is a profoundly complex tool.
• Its in-memory processing consumes massive memory.

Incentivized

• More templates for Bigquery and App Engine. There is only limited options for templates so the things we use can limit.
• I would like native connectors for Excel (XLSX) to reduce the need for custom wrappers in financial pipelines.
• Debugging Google Cloud Dataflow using only logs in Cloud Logging can be overwhelming sometimes, and it’s not always obvious which specific element in the flow caused a failure. IT uses a lot of time.

Incentivized

It really saved a lot of time and it's flexibility really can give you infra which is future-proof for most of the use cases may it be streaming or batch data. And with this you can avoid use of resource-heavy big data offerings.

Incentivized

The documentation was confusing and lacked examples. The streams suddenly stopped working with no explanation and there was no information in the logs. All these were more difficult when dealing with enhanced fan-out. In fact, we were about to abort the usage of Kinesis due to a misunderstanding with enhanced fan-out.

Incentivized

The main benefit was around set up - incredibly easy to just start using Kinesis. Kinesis is a real-time data processing platform, while Kafka is more of a message queue system. If you only need a message queue from a limited source, Kafka may do the job. More complex use cases, with low latency, higher volume of data, real time decisions and integration with multiple sources and destination at a decent price, Kinesis is better.

Incentivized

Apache Spark Streaming stands above all the huge data transformative tools because of its speed of processing which was quite slow in Presto as it takes a lot of our time in the data processing. Spark, comfortably provides integration with Jupyter like notebook environment. and Spark's combination with Jupyter and Python results in enhancing the speed .

Incentivized

Google Cloud Dataproc Cloud Datafusion

• Caused us to need to re-engineer some basic re-try logic
• Caused us to drop some content without knowing it
• Made monitoring much more difficult
• We eventually switched back to SQS because Kinesis is not the same as a Queue system

Incentivized

• Cost and time-effective tool for our business.
• We can integrate with Jupyter with many conveniences.
• Its high-speed data processing has proved beneficial for us.

Incentivized

• cost saving from managing our own data center for ETL servers
• consumption based pricing
• with auto scaling feature, we were able to expand components to support work load