Building:

Use compiler directly:

$ g++ -Wall -std=c++17 -Isrc -O3 -c *.cpp src/*.cpp  
$ g++ -Wall -std=c++17 -O3 -o data-explorer *.o  

OR
use CMake + GCC/Clang to compile project and tests (from some IDE or cmd).

Execution:

$ data-explorer sample.txt

Usage:

<operation> <aggregation> <grouping>

Usage Example:

$ avg score movie_name

Example output:

avg score GROUPED BY movie_name  
ender's_game 8  
pulp_fiction 6  
inception 8  
Operation time = 0.000009s  

Files:

main.cpp - main file :)
Column.[h|cpp] - Abstract and base class for column inheritance hierarchy.
IntegerColumn.[h|cpp] - Class for storing data and performing operations on integer type columns.
StringColumn.[h|cpp] - Class for storing data and performing operations on string type columns.
DataLoader.h - Abstract and base class for data loaders.
FileDataLoader.[h|cpp] - File data loader. Loads from file headers, types and finally data.
Dataset.[h|cpp] - Representation of data. Contains info about headers, column types and stores Column class objects.
Operation.[h|cpp] - Stores enum OperationType. Also "math" is done here using some templates.
Query.h - Trivial structure for storing which query user requested.
UserInterface.[h|cpp] - Functionalities related to interaction with user.

Additional info

As speed is most important expectations from task there was some optimization performed. Ones with biggest impact:

1. Used std::unordered_map instead of std::map.
2. Used std::vectors to store data and passed by const reference.
3. Storing strings as mapped values (std::string <-> unsigned int) and usage of indexes for operations (performance and significant memory optimisation).
4. Minimized copying.

Potential further optimizations:

1. Usage of dynamic C-style arrays for storage. To introduce those input file need to be passed 2 times (first to check number of rows).
2. Usage of C-style array + index instead of maps (if applicable and worth doing).

Potential options for scalability:

1. Usage for multithreading by introducing threads.
2. Usage of MPI (make sense with more sophisticated calculations).
3. GPU calculations (in case of more complex calculations).

I'm not fully happy about:

1. Template usage.
2. Allowing accessing private field data of Column subclasses from outside (performance reasons).