Conditions in Join Clauses (3.20+)


(lloyd tabb) #1

Release 3.20 (still weeks away when writing this) will come with the ability to add conditions in Join clauses.

The problem

Tables in many MPP database have a single sort key, that key is often time. Joining to these tables presents a challenge.

The common solution is to add a date condition in the join predicate. Until now that has been difficult in Looker.

For example, you might wish to join orders and emails for a given user over time and compute a conversion rate. Both orders and emails might be indexed on time and otherwise difficult or slow to join.

In SQL

You might write the following query

SELECT 
    users.id
    , COUNT(DISTINCT order.id)
FROM users
LEFT JOIN orders 
   ON orders.user_id=users.id 
      AND orders.created_time  BETWEEN '2015-01-01 00:00:00' AND '2015-01-31 00:00:00'
WHERE
  users.created_at BETWEEN '2015-01-01 00:00:00' AND '2015-01-07 00:00:00'
GROUP BY 1

In this case, we are looking at the number of orders that happened in january by the users created in the first week of january.

Since the orders table only has an index on time, we need to add a timeframe in order to avoid scanning the entire orders table. Adding:

 AND orders.created_time  BETWEEN '2015-01-01 00:00:00' AND '2015-01-31 00:00:00'

helps the query optimizer figure out how to pull out a subset of the data and join it.

In LookML

In release 3.20, these kinds of query can now be simply expressed in LookML.

The following example starts from the user base view then joins orders. In the explore, you must set two filters, the timeframe of the user creation and the order timeframe you wish to examine. In a non-MPP world, the SQL optimizer could probably use user_id as a key to limit scanning the orders table, but in an MPP world, you would need to set this manually.

- connection: red_look

- scoping: true           # for backward compatibility

- explore: users
  always_filter:
    users.created_date: 30 days
    orders.order_date_filter: 30 days
  joins:
  - join: orders
    relationship: one_to_many
    sql_on: |
       ${orders.user_id}=${users.id}
       AND {% condition orders.order_date_filter %} orders.created_at {% endcondition %}
      
- view: orders
  fields:
  - dimension: id
    primary_key: true
  
  - dimension: user_id
   
  - filter: order_date_filter
    type: date
  
  # You can still group by whatever time frame you like.
  - dimension_group: created
    type: time
    timeframes: [time, date, week, month]
    sql: ${TABLE}.created_at
  
  - measure: count
    type: count    
    
- view: users
  fields:
  - dimension: id
    primary_key: true
  
  - dimension_group: created
    type: time
    timeframes: [time, date, week, month]
    sql: ${TABLE}.created_at
  
  - measure: count
    type: count

Running the following query:

Yields the following SQL

SELECT 
	COUNT(*) AS "users.count",
	COUNT(DISTINCT orders.id) AS "orders.count"
FROM users
LEFT JOIN orders ON orders.user_id=users.id
AND (( orders.created_at ) >= (CONVERT_TIMEZONE('America/Los_Angeles', 'UTC', timestamp '2015-03-01')) AND ( orders.created_at ) < (CONVERT_TIMEZONE('America/Los_Angeles', 'UTC', timestamp '2015-04-01')))


WHERE 
	((users.created_at) >= (CONVERT_TIMEZONE('America/Los_Angeles', 'UTC', timestamp '2015-03-01')) AND (users.created_at) < (CONVERT_TIMEZONE('America/Los_Angeles', 'UTC', timestamp '2015-03-07')))
ORDER BY 2 DESC
LIMIT 500

This is just a simple example, there are many ways to use this feature to optimize queries.


Filter Out Rows in a Join
Have filters on joined view not change measures on explore view?
Looker 3.20 Release Notes
(Brett Sauve) #2

If you’re not familiar with the {% condition orders.order_date_filter %} orders.created_at {% endcondition %} syntax, it is called a “templated filter”. You can read about them here.


(Vlad Dubovskiy) #3

This will change our lives in a few tough modeling cases.


(Zev Lebowitz) #4

Will this work in the “sql:” join parameter as well?


(lloyd tabb) #5

Yes, It should.


(Aaron Bostick) #6

This is very exciting. Will {% parameter } also work in this case in addition to {% condition } because I have a few use cases where parameter works much better (one date filter against multiple date columns)?


(lloyd tabb) #7

Yep! It should though I haven’t tried. Play with it on http://learnbeta.looker.com . You should have an account.


(Ken Yeoh) #8

Is there any way to make this apply for all fields in a table? (Move all statements in the WHERE to the ON part of a join)

We have a situation where we’re doing a FULL OUTER JOIN on two tables, and are interested in the results from both sides, but are also interested in the NULLs. (Think a usage log and customer lookup, and wanting to see customers who have NOT used X).

Right now, I’m thinking I have to write a condition for each filterable field on both tables.
Any other ways to accomplish this?


(lloyd tabb) #9

This is an unusual pattern, probably difficult for business users to understand at any rate. Suppose you have a users table and you want to know users that never did something. There is a transaction table with a type string field that they might have done.

Almost all SQL dialects support EXISTS and that coupled with {% condition %} blocks can give you most of what you need.

The following snippit will filter the users table on users that never had a transaction of type of filtered by ‘user_did_not’. You can make other filters and include it here.

The pattern is a little odd, I’d recommend building a special purpose explore.

- view: users
  fields:
  - dimension: id

  - filter: user_did_not
    sql:  |
       NOT EXISTS (
         SELECT * FROM user_transactions
         WHERE ${user.id} = user_transctions.user_id
           AND {% condition user_did_not %} user_transactions.type {% endcondition %} )

(brettg) #10

An additional performance consideration here:

If using a join condition that is on dates, avoiding timestamp casting on the filter can lead to considerable performance gains in certain cases. You can include the datatype parameter to avoid this casting. Similar to the time dimension group, looker defaults to assume datetime.

  - filter: order_date_filter
    type: date
    datatype: date