Revision lecture

Review of topics covered in lectures
Authors
Affiliation

Matt Graham

Centre for Advanced Research Computing

Anastasis Georgoulas

Data pipelines

Data pipelines are an abstraction for thinking about data analysis workflows.

flowchart LR
    access --> preprocess --> transform --> serve --> model:::highlight -->  publish
    classDef default fill:#00c0c3,stroke-width:0;
    classDef highlight fill:#f9c437;

Can be visualised as a directed acyclic graph (DAG) showing how data ‘flows’ between different stages of analysis.

Defining all stages of a data pipeline programmatically allows us to reproduce the analysis.

Published data artefacts should aim to be FAIR - findable, accessible, interoperable and reusable.

  • How can we visualize data pipelines?
  • What are the advantages of defining a data pipeline programmatically?
  • What issues should we consider when publishing the outputs of a data workflow?

Data version control

Data version control is a way of systematically keeping track of models and datasets.

gitGraph
    commit id: "a13fb"
    commit id: "cb28d"
    commit id: "2ef7c"
    branch develop
    commit id: "ff86a"
    checkout main
    commit id: "46cd3"
    merge develop
    commit id: "54eb9"

Builds on Git, a general version control system, which allows us to track changes in a directory of files.

DVC extends Git by

  • giving better support for working with large data files,
  • integration with remote data providers,
  • abstractions for thinking in terms of data workflows.
  • Why should we use version control?
  • How does data version control (DVC) differ from version control systems such as Git?
  • How does DVC represent data pipelines?

Databases

Databases are collections of data organized in a way that allows more efficient storage and retrieval.

erDiagram
     direction LR
     Person{
        text id
        text personal
        text family
     }
     Visited{
        integer id
        text site
        text dated
     }
     Site{
        text name
        real lat
        real long
     }
     Person ||--|| Visited : " "
     Visited ||--|| Site : " "

Relational databases store data in tables with columns with fixed names and types and relational constraints between primary and foreign keys across tables.

Structured Query Language (SQL) provides a standardized approach to extracting, aggregating and updating data in relational databases.

Non-relational databases offer an alternative approaches for working with less rigidly structured data.

  • What are the benefits of using a database to store and access data?
  • How can we specify the structure of a relational database?
  • How do we select data subsets of interest from a relational database?
  • What are the advantages and disadvantages of non-relational (‘NoSQL’) databases compared to relational databases?

Working with lots of data

When working with ‘big data’ the size of datasets and resources available limit the computations we can perform.

One option is to increase our available resources for a problem - for example by exploiting cloud compute services.

An alternative approach is to break down the problem in to smaller subtasks that can be distributed.

MapReduce offers one particular framework for efficiently breaking down tasks and combining results

Image credit: Ben Congdon
  • What are key issues affecting the performance of data analysis tasks?
  • How can we describe the scaling behaviour of the algorithms we use?
  • What are the benefits and drawbacks of using cloud services as an approach for scaling computations?
  • How can we break down and distribute large computational tasks?

Software frameworks

Logos of Hadoop, Spark and Dash

Apache Hadoop, Apache Spark and Dash are three open-source frameworks for large scale data processing.

Hadoop provides a distributed file system and implementation of MapReduce framework.

Spark allows processing large amounts of data in-memory and supports more general workflows than Hadoop’s MapReduce.

Dask is modern native Python framework for distributed data processing that supports a generic task-scheduling paradigm.

  • What are the key features of Hadoop, Spark and Dask and how do they differ?
  • Why would we use a distributed file system for data processing?
  • What is streaming data?
  • How does Dask abstract computations to allow them to be optimized and distributed?

General exam tips

  • You need to know how to apply concepts and tools covered in lectures to real data problems.
  • Make sure to justify and explain points you are making - need to see evidence of understanding material not just recall.
  • Read the exam rubric carefully - each question should be answered in a separate booklet.