Written by: Paul Raff, Effective Data Consulting
“Hey - can you help me understand this data?” is an all-too-common question Data Scientists face in their day-to-day interactions with the business. In order to trust the data, you must understand the data, and to understand the data, you must understand it completely.
No Data Left Behind (NDLB) is a technique and a set of principles that all good data-driven institutions utilize, consisting of the following four themes:
- Don’t drop data - all data that comes into the system must be available somewhere for inspection and analysis.
- Use VIEWs or STORED PROCEDUREs or similar constructrs - clear, canonical ways in which the data should be accessed for consumption.
- Distinguish raw values from processed (“cooked”) values in your data.
- Be transparent in what you are presenting.
Good Case Study - Website Traffic
Website traffic - take Google for example - consists of a lot of different kinds of traffic:
- Normal website traffic from the people you care about analyzing.
- Bot/scraper traffic that’s fairly easy to identify but not helpful for most analyses.
- Test/internal traffic that’s used for quality control and testing, but is also not helpful for most analyses beyond the testing and QA teams.
Utilizing No Data Left Behind in this scenario results in a couple of key themes:
- The standard published view for analysis of this data automatically excludes bot/scraper and test/internal traffic, but allows for the data consumer to gain access to these slices of data through optional parameters.
- All analyses presented and shared declare and make transparent the data utilized to present this analysis.
Bad Case Study - NY Board of Elections
Kudos to the NY Board of Elections for testing out their elections systems, but they made a critical failure in keeping a clear separation between the ballots that count and their testing/dummy ballots. As a result, they are getting a lot of additional bad press and may get overhauled completely.
Even if you consider yourself a well-run data shop, it only takes one critical error to undo years of progress in your data systems and insights. To effectively leverage data, everyone must trust the data and fair or not, one error can bring into question everything else about your systems.
The NY Board of Elections failed on Principle #2 of NDLB - they did not have clear, canonical ways in which data was accessed and understood.
Principle 1 In Detail - Don’t Drop Data
Let’s go back to the website traffic example. Their data flow - starting from their raw ingest - could be simply conveyed like this:
How do we ensure that no data is dropped? Fundamentally it involves two key components:
- Define what your canonical/atomic unit of data is. For a website, it’s likely a pageview or a page impression.
- Ensure there’s a unique identifier available for your canonical/atomic unit of data.
We can then programmatically check NDLB via queries over our input and output datasets - if we name them
OUT_DATA, this check turns into something like the following:
SELECT input.pageview_id AS input_pageview_id, output.pageview_id AS output_pageview_id, COUNT(*) AS num_matches FROM IN_DATA AS input FULL OUTER JOIN OUT_DATA AS output ON input.pageview_id = output.pageview_id GROUP BY input_pageview_id, output_pageview_id HAVING input_pageview_id IS NULL OR output_pageview_id IS NULL OR num_matches != 1
You now have a clear-cut, simple way to ensure you’re not dropping data.
Principle 2 in Detail - Clear, Canonical Ways to Consume
Looking back at our data flow, there are three separate slices of data one may or may not want to consume:
- Good Traffic - this is the canonical set people should start with by default.
- Bot Traffic - primarily for use of the bot detection team.
- Test Traffic - primarily for use of the testing team.
There are a few possible solutions here in terms of how we can achieve a clear, canonical way to consume, but an important consideration here is that defaults matter - you must not make the data consumer do additional work to get the “standard” data.
One possibility: separate tables for these different slices:
WEBSITE_TRAFFIC WEBSITE_TRAFFIC_BOT WEBSITE_TRAFFIC_TEST
WEBSITE_TRAFFIC directly will get the Good Traffic. This is the canonical way. If separate analysis is necessary, the data consumer can do so by
UNIONing the relevant tables together depending on the scenario.
A separate possibility is to utilize a Stored Procedure or something similar where the parameters indicate what is being used. As with the previous example, it’s critical to ensure that the shortest/most intuitive path represents the most common analysis scenario for your data.
Principle 3 in Detail - Distinguish Raw Values from Cooked Values
The raw/cooked metaphor for data is wonderful - it cleanly describes how systems and entities receive data in “raw” form and there’s always some processing occurring that “cooks” the data into a cooked form.
Cooked data is great, but it helps to understand what’s actually happening under the hood. To address this, use a simple naming schema for your data to distinguish what’s raw and what’s cooked.
Sticking with our website traffic example, saw our raw data is held in a table called
RAW_DATA with the following schema:
RAW_DATA pageview_id INT page_url VARCHAR(100) user_id VARCHAR(20) pageview_time FLOAT
An important part of the data cooking process is standardization, and there are a few things that we could do here:
- Any strings can be normalized to remove excess spacing and standardize on capitalization.
- Times like
pageview_timecould have clearly incorrect values (negative 1 minute, anyone?). These issues should not be ignored, but the fact is this data already exists and there’s no going back.
A simple and effective way to abide by NDLB here is to add existing columns for the raw column counterparts to columns that were changed in some way. In our example we now have:
RAW_DATA pageview_id INT page_url VARCHAR(100) user_id VARCHAR(20) pageview_time FLOAT page_url_raw VARCHAR(100) user_id_raw VARCHAR(20) pageview_time_raw FLOAT
Done this way, we haven’t lost any information and we now better know and understand what has changed simply by looking at the instances where a column and its
_raw counterpart differ.
Principle 4 in Detail - Be Transparent
Nothing infuriates me more than a chart in an internal memo or presentation with no corresponding reference pointers. Rarely can I accurately reproduce the data being shown. It’s led me to think of the Data Scientist Consistency Check - ask 10 data scientists in your company to calculate a basic number, like the total # of pageviews of your website yesterday. The higher your number is, the more mature of an organization you are in!
Similar to the MVC Framework for websites, there is typically a clear separation between the analysis portion of a Data Scientist’s work and the presentation portion of it, especially to executives and non-technical people.
Luckily, Data Scientists are pretty standard in their workflow, specifically for analysis: notebooks. So the solution here is simple: provide a link. It doesn’t need to be flashy or in-your-face, but it needs to be there with the presentation. For example:
All you need is a link to the source of your analysis, and you’ve done the minimum possible to satisfy NDLB.
Abiding by No Data Left Behind is a high-ROI investment you can make to your data processes and systems. By ensuring that you are treating your data properly, you will reap dividends through faster insights, high-bandwidth data-driven communications, and very little back-and-forths around the data.
Trust is everything with data, and NDLB will give you that trust.
About the Author
Paul Raff is a PhD Mathematician and owner of Effective Data Consulting. He’s currently a Principal Data Scientist in the Clinical Analytics team at Aetna, a CVS Health Company. Prior to that he spent 10 years at Amazon as a supply chain researcher and at Microsoft as an expert in A/B testing and online experimentation.
Paul focuses on the big picture with data, which emphasizes the need to think of data holistically and utilizing simple and effective techniques for your data. He excels at architecting the appropriate use of data and solving today’s problems while preparing for tomorrow’s. Contact him to find out more about how he can help your business excel, either through freelance or consulting work.