Daniel Molnar, The Data Janitor about talks/tv music articles translations

Cyclic or acyclic product usage flows stored in Redshift? No probs, visualize them with ease.

A second chapter in abusing Redshift with window functions came suddenly on a Tuesday evening when my dear colleague, Jenny, prepared to shine light on how our users change products.

First let’s find the ones who had more than one product during their lifetime.

with multi_product_users as
(
 select user_id 
 from subscriptions 
 group by 1
 having count(distinct product_id) > 1
),

Let’s get the starting date for each product usage — we have a separate master table for production descriptions named products. Let’s inject a date smaller than any real date for each user plus a row that will end up at the end of our ordered list.

usage as
(
 select p.name, u.user_id, s.started_at
 from products as p, multi_product_users as u, subscriptions as s
 where p.id = s.product_id
 and u.user_id = s.user_id
 union all
 select '', user_id, '1970–01–01' from multi_product_users
 union all
 select '', 999999999999, ''
 order by 2, 3
),

Construct these lovely strings of ‘first product [$] second product’ with the window function ‘lag’.

concatenated as 
(
 select user_id,
 started_at,
 name || ' [$] ' || lag(name, 1)
 over (order by user_id, started_at) as lag
 from usage
),

Count all ‘first product [$] second product’ pairs where neither product name is an empty string — here we end up with edges of a graph weighed.

counted as
(
 select lag, count(*)
 from concatenated
 where split_part(lag, ' [$] ', 1) != ''
 and split_part(lag, ' [$] ', 2) != ''
 group by 1
 order by 2 asc
)

If we push the weight in between the two product names we end up with a dump that we can transform to a Sankey Diagram with a copy-paste.

select replace(lag, '$', count) as sankey
from counted;

The only caveat with the Sankey Diagram is that your product flow must be acyclic.

How to handle cyclic product flows? You can pretty easily end up with a GDF-format graph descriptor that Gephi can load if you concatenate the two product names with a comma and leave the weigh in the end.