While working with Cube.js framework we've seen a lot of diffrent ETL tools used by data engineers nowadays. Most of them require writing code. But there are some visual ETL you can try as well. We asked Dmitry Dorofeev, Head of R&D at Luxms Group, to tell us about his experience with comparing Apache NiFi and Streamsets.
***Our team at Luxms Inc. has recently faced a boring data integration problem: when some data is stored in Hadoop, some in Oracle, and a little bit is in Excel. The goal was to ETL all that data into Greenplum and finally provide some BI on top of it. We quickly found 2 mainstream open source ETL projects: Apache NiFi and Streamsets, and it seemed an easy task to choose one product out of the two. In no way was it easy. I know that better than anyone since I was responsible for the product evaluation and the final choice. In this post, I’d like to share my experience and maybe save days of your life.Spoiler: There is no silver bullet. Apache NiFi is not necessarily better than Streamsets, nor Streamsets better than NiFi. Everything has its pros and cons. This post is my personal experience with these tools as a novice user without any introductory training.
Dataflow ProgrammingProgrammers, analysts, and even managers often draw a box and arrow diagram to illustrate some flows. You can even use these boxes and arrows to create programs. We can track such attempts back to the 1960s when the Dataflow Programming paradigm was born in MIT.Today, we have tens of Dataflow Programming tools where you can visually assemble programs from boxes and arrows, writing zero lines of code. Some of them are open source and some are suitable for ETL.Yes, you don’t have to know any programming language. You just use ready-made “processors” represented with boxes, connect them with arrows, which represent exchange of data between “processors,” and that’s it.
There are three main types of boxes: sources, processors, and sinks. Think Extract for sources, Transform for processors, and Load for sinks.Almost anything can be a source, for example, files on the disk or AWS, JDBC query, Hadoop, web service, MQTT, RabbitMQ, Kafka, Twitter, or UDP socket.A processor can enhance, verify, filter, join, split, or adjust data. If ready-made processor boxes are not enough, you can code on Python, Shell, Groovy, or even Spark for data transformation.Sinks are basically the same as sources, but they are designed for writing data.
Apache NiFi vs StreamSetsWhen we faced yet another customer with complicated ETL requirements I decided to try visual dataflow tools. Visual might be attractive even if you use Singer, data build tool, or other handy open source ETL tools, right?Luckily, there are two open source visual tools with the web interface: Apache NiFi and StreamSets Data Collector (SDC). NiFi was donated by the NSA to the Apache Foundation in 2014 and current development and support is provided mostly by Hortonworks. SDC was started by a California-based startup in 2014 as an open source ETL project available on GitHub. The first release was published in June 2015.Both products are written in Java and distributed under the Apache 2.0 license.Here are some stats from GitHub for early 2018:
|First release year||2007||2015|
Architecture and featuresBoth tools encourage creation of long-running jobs which work with either streaming data or regular periodic batches. You can create manually managed jobs, but they might be tricky to set up. This is the greatest surprise and mind-shifting feature I personally had with these tools.
Apache NiFiApache NiFi has a well-thought-out architecture. Once data is fetched from external sources, it is represented as FlowFile inside Apache NiFi dataflows. FlowFile is basically original data with meta-information attached to it. You can easily process not only CSV or other record-based data, but also pictures, videos, audio, or any binary data.A processor usually will have 3 outputs:
- Failure. If a FlowFile cannot be processed correctly, the original FlowFile will be routed to this output.
- Original. Once an incoming FlowFile has been processed, the original FlowFile is routed to this output.
- Success. FlowFiles that are successfully processed will be routed to this relationship.
StreamSets Data CollectorThen, I tried Streamsets. Processors in Streamsets exchange records. That means that everything you ingest into Streamsets is converted automatically into the standard record-oriented format and all processors can handle it as a stream of records. There are no queues in between processors, at least, they are not represented visually, like we saw it in Apache NiFi.One nice thing about Streamsets is that it can process binary data. Some sources, such as Kafka Consumer, can read messages from the Kafka topic and pass them to other processors or external systems without parsing the structure of the binary message into the record format. This allows us to forward the efficient data to some other destination with minimum overhead. The more powerful option is the whole file data format, supported by several origins, including S3, directory, FTP, and more. With the whole file format, the file is not parsed, but file metadata and a reference to the content is sent along the pipeline. Processors can optionally act on the content – script evaluators and custom processors can get an input stream to the content. But in the default case, once the whole file record arrives at the destination, the data is streamed directly from its source.Even though there are some complaints about lack of binary data support in Streamsets, the whole file support has been there since version 184.108.40.206, released in September 2016. In Apache NiFi the same processor should have different versions of itself to handle different formats. One version for CSV, one for JSON, and another for Avro, for example. You might guess that it is not very user and developer friendly. This was addressed in Apache NiFi 1.5, where most processors are using the Avro format, so you should convert to Avro early and it will be almost the same experience as in Streamsets after that.
To change processor settings in Apache NiFi you must stop the processor, while in Streamsets you must stop the whole dataflow.That means that you always start your dataflow from the beginning after you make any changes in it with Streamsets. With Apache NiFi you have a chance to stop a misbehaving processor, fix it, and start again. Hopefully, queued FlowFiles will be sent to the fixed processor and you will not miss the data.But that doesn’t mean that Streamsets dataflows are harder to debug. Actually it is easier, you have a nice-looking live dashboard displaying a lot of statistics for every processor while your dataflow is running. Errors are cleanly presented as red numbers on the processor icon and you can see individual errors for every faulty record with a mouse click. You may even put record filters on the connections between processors to inspect records in question. Filters can be applied while your dataflow is running, so I used it as live debugging tool.Streamsets has 4 processor types:
- Origins: they get data from the external sources. You may have only one Origin Processor in your dataflow.
- Processors: data transformers.
- Destinations: they save data to the external systems or files.
- Executors: they process events, generated by other processors.
Apache NiFiThere is not much to say about the Apache NiFi UI. It feels spartan, and it is very easy to follow, thanks to the great architecture with minimum concepts. Probably the only drawback I discovered is that Apache NiFi will not autosize text fields for your long SQL queries, so you will have to manually resize popup text fields every time you want to edit it.
StreamsetsStreamsets has a more attractive UI, but it is not perfect as well.The first thing I quickly get annoyed with is the absence of Controller Services, especially for JDBC settings. You need to fill in all JDBC settings for every processor that reads data from the same JDBC source. There is just no user-friendly way to reuse such information.Before you can run your dataflow, Streamsets will check each processor inside your dataflow to make sure all processors are correctly configured. It sounds like a good thing, and it helped me sometimes, but other times – harmed me. In Apache NiFi you can have disconnected processors and I usually leave them so for debugging purposes. In Streamsets you can not do the same, since all the processors must be connected to make dataflow pass validation.Another annoyance is that you can not select several processors at once. At least I was unable to do so. Moving a dozen processors and reorganizing them one by one on the screen can make you mad.Streamsets has syntax highlighting for SQL which is a nice feature, but not always useful. Our data engineer creates heavy SQL queries which can easily be a hundred lines long. The syntax highlighting process becomes slow and that results in another annoyance. If you edit the last lines of the long SQL query the caret unexpectedly moves to the beginning and what you type appears on the first line.
- Clean and well-thought-out implementation of the dataflow programming concept
- Can handle binary data
- Data Provenance
- Spartan User Interface
- No live monitoring/debugging features with per-record statistics
- Live monitoring/debugging features with visual per-record statistics for every processor
- Sexy UI
- Well-suited for record-based data and streaming
- You need to stop the whole dataflow to edit a single processor configuration
- No reusable JDBC configuration for processors