This model kit is one of the most exciting but is also one of the most difficult I have ever attempted: Italeri 1:72 Junkers Ju-52. This Nazi-era transport plane has a unique three-motor arrangement which gives the aircraft a very distinctive look.
The quality of the kit itself is not up to my standard: many parts such as the wings, engine cowl and flaps did not really hold together. I had to take multiple attempts to glue the parts, because they fell off easily.
The actual aircraft featured in this kit is a Minesweeper variant. It has a large metal ring under the fuselage (see it here) which works to detonate magnetic mines in sea. In my opinion, the ring would make the plane look unattractive, so I chose not to install it on the model. Cockpit window frames are also difficult to paint, so I simply left them unpainted.
I like WW2 planes, but because of its difficulty, building this one was not really the kind of fun I was looking for.
I wrote this post to share with you how to install PySpark in MacOS. I was reformatting my laptop last week and I found it difficult to reinstall PySpark because the Apache Spark documentation did not mention much about MacOS. If you Google it, you would find there are quite many different ways of doing this, but I could assure you that what I wrote here would be the most straightforward way to install PySpark in a standalone setup in a MacOS.
Standalone setup is ideal if you need to write in PySpark but do not own or do not have access to a Hadoop / Spark cluster. If you only need to write PySpark programs locally in your laptop without actually running them in a cluster of machines, for example at home or in a coffee shop, then this is the way to go.
To install PySpark, follow these 7 easy steps below. This assumes that you are starting with a clean machine. Be warned, you will need a fast internet connection because the size of the software to download will be quite large.
Download and install Java (JDK) SE 8 if you haven’t done so. Beware: do not use other version. If you use newer version, you will have to downgrade.
Download and install Anaconda (http://anaconda.com). Pick Python 3.x instead of Python 2.x. Test your installation by creating and running a simple Python program before you proceed to the next step.
Download and install Homebrew (https://brew.sh). Homebrew is package manager for MacOS, we will need this to install Spark. Once Homebrew is installed, open a new Terminal and run this command to get core Apache Spark package:
brew install apache-spark
4. Once finished, go to your Spark directory /usr/local/Cellar/apache-spark as shown below (change the numbers with the actual version you have installed in your computer, in my case it is “2.4.5”) and then find and edit the bash_profile file in that directory:
cd /usr/local/Cellar/apache-spark/2.4.5 nano ~/.bash_profile
Add these new lines at the very bottom of the file, then save and close:
Python + Pandas are great for data analytics tasks and people love them. However, they have their own limitation: standard Python interpreter only runs in a single CPU core. When it comes to larger datasets, this weakness begins to take its toll as it prevents us to process data larger than what a single core CPU can handle.
Some people say Pandas is still good for files up to 10GB, but in the era of big data, this sounds small. And it is.
Apache Spark may be the answer for large data processing and is currently gaining a lot of traction, but using Spark would mean that you have to rewrite your Python programs, because Spark has completely different programming model.
Dask (https://dask.org) is an emerging framework built around Python ecosystem which offers running Python programs in multiple parallel processors. Similar to Spark, we can now run Python programs to process large amount of data. The contrasting difference is, you do not really need to rewrite your code as Dask is modelled to mimic Pandas programming style. Minimum change is needed to adopt Dask (at least, that’s the original goal).
In this post, I will show you how to create a simple Dask job.
For the purpose of demonstration, I have set up three Linux machines as Workers and another machine as Scheduler. For instructions on setting up the environment, the information in this Dask documentation can help you. It is pretty straightforward, I could set up my four machines in a matter of minutes.
In essence, you will need to install Dask in all machines, run Dask-Scheduler program in the Scheduler machine to listen incoming connections, and then run Dask-Worker programs in all Workers to talk to it. Once running, we can log on to the monitoring web app (called Dask Bokeh) where something like this will be shown:
At this point, Dask environment is ready to run programs. To test, I wrote a new Python program like below to activate a Dask Client (your interface to the multiprocessor capabilities), and then load a data file into the Workers:
This loads data from a CSV file located in a remote HTTP server, stored as a Dask DataFrame. In my example, the CSV file is a GPS tracking data with millions of records (around 7GB in size). There are ten fields in this file at my disposal, but I am only interested in the key column which is the IMEI, the unique GPS device identifier (Note: no, I did not take this data illegally from my workplace. Customer data remains protected, no law is violated).
As you can see in the code above, the operation is pretty much similar to Pandas DataFrame, read_csv() function is employed as usual. Take note on the persist() function which will ‘pin’ the DataFrame in all Worker’s main memory, leaving the data available for further process. We don’t have this in standard Pandas.
In this demonstration, a GroupBy() aggregate function is applied on the DataFrame to count how many records are there for each IMEI. Since the data is distributed over all Workers, the computation process will be done in all machines. Depending on the processing power, this can take several minutes. I was using four Virtual Machines on my 5-year-old Intel Xeon machine with 16GB RAM, and it took less than 3 minutes to complete.
The first few rows of the aggregation result are shown below (there are thousands of IMEIs in the source file):
The above example demonstrates basic data processing with Dask, but Dask has much more than that to offer. There is also Dask-ML which allows Sklearn libraries to be run on a parallel environment, all by modifying few lines of code. I have not had the chance to actually experiment with it to see if it really works, but Dask-ML certainly is an interesting option for those who need to run models on larger dataset but do not want – or do not have time – to convert to PySpark.
So how does Dask compare with Apache Spark? Spark is clearly more popular, the de facto standard for big data processing. Dask is young, Therefore one might ask, does Dask have a future by challenging Spark’s dominance? Well, for start, Dask is not meant to compete with Spark. Both have different use cases.
Furthermore, Dask receives funding from Anaconda Inc, one of the prominent players in the data science space, thus making the development pretty much active. The future seems quite bright for Dask.