This is difficult issue. It sounds as if you are going to have a handful of data collection schemes for your 100 data sources. This, combined with the desire to create a single, unified output format leads me to think Heirarchical. You will have some significant overlap in how you access the raw data per data source. All static web sites would have a url that you access. All FTP sites would have a remote server, login, password, and file path. All RDBMS will have similar credentials. In all cases, you will want to do the following:

fetch_raw_data

parse_raw_data

write_pased_data

This would make me go with something akin to:

DataCollector
DataCollector::Mechanized
DataCollector::Mechanized::WalMart
DataCollector::Mechanized::GeneralElectric
DataCollector::RDBMS
DataCollector::RDBMS::ExxonMobil
DataCollector::FTP
DataCollector::FTP::GeneralMotors
DataCollector::Scrape
DataCollector::Scrape::FordMotorCompany
DataCollector::Scrape::CiscoSystemsInc
. . .
etc.
[download]

Your driver program would then, unfortunately, need to know all of the DataCollector leaf classes or devise a method to dynamically load and run them. But for each of these classes, you could call the above mentioned methods. Those methods would make private method calls on down until you get to the ugly details in the individual implementation classes. These implementation classes would only need to know where it's going for data and how to pull the real data from raw data source. Up one level would be how to talk to the data source type, based on information in the implementation classes. Up in the top level is the detail of how to write out the data.

I hope this makes sense, isn't too vague, etc. Good luck.