ETIS09 - Data Quality: Common Problems & Checks - Presentation

Data Quality:

Common Problems & Checks

David M. Walker

[email protected]

+44 (0) 7050 028 911 - http://www.datamgmt.com

Data Management & Warehousing

Date: 24 April 2009

Location: Zagreb, Croatia

24 April 2009 © 2009 Data Management & Warehousing Page 2

Agenda

•! Introduction

•! Common Problems

•! Automated Checking

•! Profiling Checks

•! Conclusions

Introduction

•! Data Quality problems are a SOURCE SYSTEM

issue and an ETL issue

–! They just manifest themselves in the data warehouse

•! Prevention is better than cure

–! Fixing the source system or the ETL is ALWAYS

cheaper and more effective than cleaning the data in the ETL or in the Data Warehouse itself

•! Data Quality is a continuous process

–! It is never finished and always needs to be monitored


The Impact of Poor Data Quality

•! Devalues the data warehouse

–!Discourages people from trusting or using the

system and therefore curtailing the life of the

data warehouse

•! Highlights failings in the source system

and/or the business process

–!Businesses would rather fix at any cost in the

data warehouse and pretend that there isn’t a

source system problem


Common Problems

•! 11 types of problem that account for the

most common problems

•! They usually reflect poor design and/or

implementation of systems

•! Most can be fixed or monitored and

managed to limit the impact


Referential Issues

•! Keys that are not unique

–!Systems that do not enforce the unique

primary key or flat files or spreadsheets

–!Also generated by ETL that creates a

surrogate key incorrectly

•! Referential Integrity Failures

–!Where referentially integrity is not enforced

values in the child table are created that are

not in the parent table


Data Type Issues

•! Format Errors

–!Typically in Date/Time type fields

–!02/04/2009 2nd April (UK) or 4th Feb (US)

•! Inappropriate Data Types

–!Storing Dates in Character Strings

20090624 as YYYYMMDD format string

–!But what about 20090230?


Data Model Issues

•! De-normalised tables

–! Commonly created for performance reasons

–! Inherently duplicates data

–! Often gets out of sync

•! Data/Column Retirement

–! Upgrade to system retires a column

–! ETL continues to use the old column

•! Poor Table/Column Naming

–! Don’t assume that a column does what it says

–! Don’t assume that a column is still being used for it’s original

purpose


Data Content Issues

•! Null Values

–!Systems that have many optional fields will

often have missing values

–!Null values allow rows to be silently omitted

from queries

•! Inappropriate Values

–!Databases allow special characters and/or

leading/trailing white space

–! “DataspaceQuality” != “DatatabQuality”


Data Feed Issues

•! Missing Data

–!Where a stream of files are loaded by ETL if

one is dropped it can go un-noticed

–!Common with CDR type loads in Telcos

•! Late Data

–!A short term data quality issue

–!Leaves users believing there is a problem

–!Produces inconsistent reporting over time


Automated Checking

•! Regularly run checks

•! Broad Coverage across systems

–! 100s and 1000s not 10s of queries

–! Run in a low priority loop in the background

•! Used against:

–! Sources

–! Data Staging

–! Data Warehouse

•! No Product Required

–! We often implement this as a controlling shell script and lots of

small scripts, one for each check


Trending

•! Absolute Trending

–! Track an expected value over time

•! e.g. Returned Mail is usually less than 500 items per day

•! If the value is <500 status is green, 501 to 1000 amber and >1000 red

•! Statistical Process Control (SPC) Trending

–! Track an expected value where the value changes over time

•! e.g. Telco CDRs – expect more as the company grows

•! Don’t want to be continuously changing the threshold

•! Compare current load to historical means

•! If current load within 2 Standard Deviations – Green,

3 Standard Deviations – Amber,

4 or more Standard Deviations - Red


Flow Control

•! Flow Control

–! ETL manipulates data

•! Joins, De-duplicates, Filters, Aggregates, etc

–! Use the formula:

Source Count - Filtered Count – DeDup Count – Target Count = 0

•! Trusted Source

–! Compare the result with a third system

–! e.g. Does the Count of Switch CDRs =

Count of those processed by the billing system Count of those processed in the DWH


Business Rules Based

•! Specific rules to match known business

rules

–!Account holders > 18 (Sys Date – DoB)

–!Account holders < 115

–!Credit Card numbers are 16 digits long

–!Number of accounts without a status

•! Result should yield Zero


Automated Checking - Ops

•! Managed by exception

–!Red given priority

–!Amber are always followed up

•! Massive number of checks

–!100’s are good

–!1000’s are better

•! Presentation

–!Alerts, RAG, Graphical, Numerical, etc.


Data Profiling Checks

•! Run manually because they need to be

interpreted by a human

•! Leads to new business rules being added

to the automated checks

•! Can be done with simple reporting tools or

commercial data profiling tools


Frequency Outliers

•! Count discreet values in a table and check

items with many more or less than normal

–!e.g. DoB 01-01-01 many times more common

than any other value indicates source default

and something that needs work

–!e.g. Count of SMS messages significantly

lower on a given day may equate to a genuine

system failure and therefore not a DQ

problem


Maximum & Minimum

•! Determine what a valid range for any

value should be

–!e.g. age between 18 and 115

–! Immediately finds individual data quality

issues that can be resolved

–!Allows an analyst to create new business

rules to prevent future problems


Sequential Keys

•! If a system has a sequential key:

Max Value – Min Value – Count = 0

•! If this is true – is it too perfect for an

operational system and therefore test data

•! If this is false – what has caused the gaps,

are the deletions intentional?


Data Types

•! Validation of mis-used data types before

loading

–!e.g. Dates in Character fields

–!Format: YYYYMMDD

–!Check MM between 01 and 12

–!Check DD between 01 and 31

–!Check MMDD does not include 0230, 0231

–!etc.


Skewed Pattern Profiling

•! Looking for specific patterns in data

–!e.g. UK National Insurance Numbers (?) have

the format AA 99 99 99 A

–!Pattern match all values looking for

exceptions

•! Number Lengths are a special case

–!e.g Credit Card Numbers are 16 digits long


Content Checking

•! Content Checking is the manual review of

character strings

•! Needs a good understanding of the nature

of the data

•! Often determines the need to do analysis

of other types


Nulls & White Space

•! Nulls

–! Fields that have large proportion of nulls are usually

not useful

–! Also common is default status of null

(e.g. Account us either closed or null)

•! White Space

–! Not Null fields with a single space

–! Tab instead of space

–! Leading/Trailing white space

–! Double White Space: “DavidSpaceSpaceWalker”


Punctuation & Control Chars

•! Punctuation

–!CSV files that are not properly quoted perform

field shifts

–!Address lines with extra commas

•! Control Characters

–!Data fields that contain ASCII character codes

0 to 31 and 127 to 159 are often ‘invisible’

when viewed in queries but cause failures

–!Also be aware of ‘code-page’ specifics


Problem Management Matrix


Continuous DQ Process


Quality is FREE …

… as long as you are prepared to

INVEST HEAVILY in it

Philip Crosby 1980

Especially true of

Data Quality


Data Quality:

Common Problems & Checks

David M. Walker

[email protected]

+44 (0) 7050 028 911 - http://www.datamgmt.com

Data Management & Warehousing

Thank You

Technology

ETIS09 - Data Quality: Common Problems & Checks - Presentation