Post Archive

2023

Back to Top ↑

2022

Deploy Databricks on Azure with Terraform

This is a minimal example for deploying Databricks service on Azure. The smallest number of nodes in the cluster will be 1, and maximum 5. Node type will be the smallest one, and Spark version the latest one with long term support. You will be able to log in automatically with your SSO user. Auto-termination is set to 20 minutes.

Databricks logo

Amazon review dataset translations for sentiment analysis

This repository contains translations of a 150 000 randomly selected entries from Amazon dataset originally created by Julian McAuley and Jianmo Ni, containing over 20GB of data. The goal is to create smaller datasets for sentiment analysis on languages other than English, for which there are many publicly available datasets already. Translation is performed with Microsoft Azure’s Translator cloud service.

AWS Spanish Translation

Java OCP exam code snippets

These are my preparation exercises for the OCP exam. Each example is self-contained, executable code snippet. Usually, my examples are based on meaningful illustration and clear functionality, as in this case of BiPredicate lambda function:

        // Map
        Map<Integer, String> henryFondaMovies = new HashMap<>();
        henryFondaMovies.put(1962, "The Longest Day");
        henryFondaMovies.put(1940, "The Grapes of Wrath");
        henryFondaMovies.put(1964, "Fail-Safe");
        henryFondaMovies.put(1957, "12 Angry Men");
        henryFondaMovies.put(1937, "You Only Live Once");
        henryFondaMovies.put(1938, "The Mad Miss Manton");

        BiPredicate<Integer, String> yearFilter = ((year, title) -> year > 1950);
        BiPredicate<Integer, String> titleFilterNumeric = ((year, title) -> title.matches(".*\\d+.*"));
        BiPredicate<Integer, String> titleFilterThe = ((year, title) -> title.contains("The"));

        // this will print only movies that are released after the 1950
        for (Map.Entry<Integer, String> entry : henryFondaMovies.entrySet())
            if (yearFilter.test(entry.getKey(), entry.getValue()))
                System.out.println("after 1950: " + entry.getValue());

        // this will print only movies that contain number in the title
        for (Map.Entry<Integer, String> entry : henryFondaMovies.entrySet())
            if (titleFilterNumeric.test(entry.getKey(), entry.getValue()))
                System.out.println("number in title: " + entry.getValue());

        // this will print only movies that contain "The" in the title
        for (Map.Entry<Integer, String> entry : henryFondaMovies.entrySet())
            if (titleFilterThe.test(entry.getKey(), entry.getValue()))
                System.out.println("the in title: " + entry.getValue());

Computer Science courses

List of Computer Science university courses freely available on YouTube

CS50 2020 at Harvard
CS50 2020 at Harvard

MIT 6.006 Introduction to Algorithms 2020
MIT 6.006 Introduction to Algorithms 2020

Back to Top ↑

2021

Questions and answers portal on Spanish - nuestras-preguntas.net

www.nuestras-preguntas.net was my demo project for questions and answers platform on Spanish language with full text search capability. Application was completely configurable, so switch to another language could be easily accomplished by providing translation for existing fields. It was running in production for nine months, with two million question and answer pairs. CICD pipeline was built in Jenkins, and application was running in Docker container on Digital Ocean hosting service. Whole system had four containers: SpringBoot application, ElasticSearch database, LetsEncrypt certbot, and Nginx proxy. They were orchestrated with Docker compose. VM instance had 8GB of RAM, and ElasticSearch was by far the largest consumer of resources. My architectural decisions were mainly guided by cost reduction, and comparable cloud solutions would be orders of magnitude more expensive.

Applications unique feature was anonymous usage, meaning registration with email was not necessary to submit a question or answer, or vote. I developed a custom encryption protocol mostly based on AES algorithm so traffic between client and server could not be reverse-engineered and automated. To defend against abuse, there were also throttling services and honey-pot traps against bots. Anti-scraping mechanisms included rotating HTML schema (without an effect on user experience or functionality), in combination with detection of unusually high amount of requests that triggered responses with poisoned content.

alt text

Back to Top ↑

2020

Test Pyramid

Martin Fowler on testing pyramid.

Java 8 code examples for the OCA exam

After doing a couple of mockup-exams, I realised that it is necessary to be familiar with a code that does compile, just as much as with code that does NOT compile. Following examples include both forms of code. Incorrect statements are commented out, and can be easily transformed back to code that throws compile-time errors.

The examples are not ordered in any way. They are just grouped by similarity (eg. arrays, flowcontrol, lambdas, exceptions, etc.)

One code example is not necessarily designed to explain only one thing; it can illustrate several syntactic or programming principles of Java language.

One example is not enough to explain or understand a principle: therefore, code contains seemingly redundant lines of code that nonetheless help in acquiring the pattern behind the particular case.

Back to Top ↑

2019

Top 500 companies in Croatian IT - Superset

First image is a dashboard composed of all other figures. Second figure is a bar chart of companies grouped by city and activity. There is a clear domination of capital city (Zagreb) where are most of the companies registered. However, there is also a clear domination of 6201 code, which denotes ‘Computer programming’. Interestingly, other cities have companies only in this domain, and nothing in counseling(6202), wireless communication (6120), data analysis (6311), equipment management (6203), or even uncategorized activity (6209). Third image represents profit allocation with a pie chart. Roughly two thirds annual profits are from computer programming. Fourth image is a word cloud that counts employees in each city and presents them in adequate ratio. Clearly, Zagreb as a capital city with the most companies has the largest word size. Fifth is a chord diagram that connects company location with activity. Again, presence of Zagreb and programming is easily notceable. Sixth image is a bubble chart. Each data point represents one company. X axis is mapped to a value of ‘Capital and reserves’. Y axis is mapped to the annual revenue. Scales are logarithmically adjusted. Bubble size is linked to the number of employees in the company, while bubble color is linked to the activity type. Presence of red coloer indicated ‘Computer programming’ category. Big purple circle in the middle is Croatian Telecom (Hrvatski Telekom). Big bluish circle in the bottom right corner is VipNet. Since SuperSet reads these numbers as strings, ratios are not correct, and the same figure should be done properly with another tool.

alt text

Top 500 companies in Croatian IT - matplotlib

First five images are presenting relation of company’s capital to the annual profit, while simultaneously showing relations of company size and type of main economic activity. X axis is mapped to annual profit, Y axis is mapped to capital and reserves. Circle size is correlated to the number of employees, while circle color represents type of work in the IT sector, such as computer programming, counseling(6202), wireless communication (6120), data analysis (6311), equipment management (6203), or uncategorized activity (6209). First image clearly illustrates extremely dominating position of T-Com company. Following images are presenting the same data with different levels of zoom. Last figure is a bar chart, revealing Zagreb as a software center, and computer programming as a sole activity of companies in different cities.

alt text

Compass and MongoDB with the Zagreb Surveillance Cameras dataset

This is a demonstration of MongoDB’s Compass visualization capabilities with geographical data. The repository contains an original csv, a preprocessed csv, a gawk script for transforming csv to json, and finally json prepared for importing in MongoDB. The list is not updated, and new cameras have been installed after the publication of the dataset in 2016. The first image is a classical, ‘neutral’ view, encompassing all geolocations. Other images are zooming into particular part of the city while shifting angle.

alt text

alt text

Visualizing Geolocations in Italy using ELK stack

This is a demonstration of ELK stack geo-mapping capabilities.

alt text

The original Dataset

The presented data are a subset of geo points in a zip file for Europe (https://s3.amazonaws.com/data.openaddresses.io/openaddr-collected-europe.zip) available through the OpenAddresses project. I used only a couple of files related to Italy (ferrara.csv, bologna.csv, statewide.csv, etc.). Circle size and color are not related to any other features in a dataset; they simply indicate a quantity of points in a certain area.

Created With

Visual Exploration of Croatian Literature Translations

This is a demonstration of ELK stack’s visualization capabilities. First figure is a smoothened area chart of translations through years, spiking in 2008, just before the economic crisis kicked in. Second visualization is a donut chart of translated authors, with Ivo Andrić being the most translated author (14.71%). Third image represents a bar chart of top Croatian literature translation publishers. Fourth figure is a donut chart of translations by language. German language dominates with 21.15%, followed by English (11.09%), Slovakian (10.62%), and Slovenian (9.04%). Figure five is identical to the previous figure, with added tabular info. Sixth figure represents an area figure of the number of translations grouped by country. Germany leads with over 300 published items. Seventh visualization is a pie chart of the most commonly translated titles. First place is taken by “Na Drini ćuprija” (9.1%), written by Ivo Andrić.

alt text

Figures from Steven Pinker’s book The Better Angels of Our Nature

Here are several visualizations about violence rates created with R.

Several percentages in a figure about a share of violent deaths are referencing only male population; to simplify, I neglected that information.

Third figure illustrates the ‘Recivilization of the 1990s’ thesis, a period of violence decline. Pinker claims it is due to the increased incarceration: “The most effective was also the crudest: putting more men behind bars for longer stretches of time.”; and increase in the police force: “In a stroke of political genius, President Bill Clinton undercut his conservative opponents in 1994 by supporting legislation that promised to add 100,000 officers to the nation’s police forces.”, among other things.

Fourth figure is adopted to the time scale of century instead of mid year, and the conflicts are ordered.

alt text

Steven Pinker’s book: The Better Angels of Our Nature

Croatian Language Dataset

This is a dataset of sentences in Croatian language for anyone interested in Natural Language Processing. The dataset is Spark’s dataframe in a snappy compressed ORC format. Most of the sentences are from the Croatian Wikipedia dump and OpenSubtitles project. There are still entries that do not belong here, whether in a form of misspelled or grammatically incorrect sentences, logs, different languages, or weird tags. However, those should be minimal. The goal of this project was to create a reusable dataset with standardised use of Croatian language for various purposes.

Size

  • 14.7 million entries
  • 840 mb uncompressed size
  • 460 mb snappy compressed ORC

alt text

Netflix Recommendation Model

alt text

Netflix prize was an open competition for the best collaborative filtering algorithm, which started in 2006. BellKor’s Pragmatic Chaos team from AT&T Labs won the prize back in 2009. This Spark application will use Spark’s 2.4 built-in ALS algorithm to create a recommendation model for the data set from the competition.

Back to Top ↑

2018

Back to Top ↑

2017

Back to Top ↑

2015

Back to Top ↑