Chained CTEs in Redshift - How do I know which DIST KEY the CTE will inherit? - view

Question

I have a view in Redshift which consists of lots of CTEs that are joined (chained) between each other. Inside these CTEs there are joins between multiple tables. If I then Join to a CTE that has a join of multiple tables inside where does the SORT KEY and DIST KEY for the Join from? How does Redshift decide which table in the join in the CTE, the CTE should inherit it's DIST KEY and SORT KEY from? If at all?
For example, tbl1 has a DIST KEY on tbl_key, tbl2 has a DIST KEY on tbl_id, tbl3 has DIST KEY on tbl_key.
First, I create a CTE which is the join of tbl1 and tbl2.
With cte1 as (
Select tbl1.col1, tbl2.col2
From tbl1
Join tbl2 on tbl1.job_no = tbl2.job_id )
Second, I create a CTE that joins to the first CTE
With cte2 as (
Select cte1.*, tbl3.col3
From cte1
Join tbl3 using (tbl_key))
Now my question is, does CTE1 have a DIST KEY on tbl1's DIST KEY of tbl_key or tbl2's DIST KEY of tbl_id? or both? or neither?

Answer 1

In Redshift, CTEs are just used to simplify the reading of sql. They are processed just the same as subqueries. i.e. they are not made physical and therefore do not have their own dist/sort key.
You could rewrite your code as
Select cte1.*, tbl3.col3
From (Select tbl1.col1, tbl2.col2
From tbl1
Join tbl2 on tbl1.job_no = tbl2.job_id
) as cte1
Join tbl3 using (tbl_key)
which can be simplified further as
Select tbl1.col1, tbl2.col2, tbl3.col3
from tbl1
join tbl2 on tbl1.job_no = tbl2.job_id
join tbl3 using (tbl_key)
If you are able to choose your dist/sort keys then you should consider which tables are the biggest and prioritise those accordingly.
for example if tbl1 and tbl2 are large then it may make sense to have them distributed as you described.
However, if tbl2 and tbl3 are both large, it may make sense to distribute both on tbl_key.

Answer 2

When you issue a query Redshift will compile and optimize that query as it sees fit to achieve the best performance and be logically equivalent. Your CTEs look like subqueries to the compile / optimization process and the order in which the joins are performed may have no relation to how you wrote the query.
Redshift makes these optimization choices based on the table metadata that is created / updated by ANALYZE. If you want Redshift to make smart choices on how to join your tables together you will want your table metadata to be up to date. The query plan (including join order and data distribution) is set at query compile, it is not dynamically determined during execution.
One of the choices Redshift makes is how the intermediate data of the query is distributed (your question) but remember that these intermediate results can be for a modified join order. To see what order that Redshift plans to join your tables look at the EXPLAIN plan for the query. The more tables you are joining and the more complex your query, the more choices Redshift has and the less likely it is that the EXPLAIN plan will join in the order you specified. I've worked on clients' queries with dozens of joins and many nested levels of subquery and the EXPLAIN plan is often very different than the original query as written.
So Redshift is trying to make smart choices about the join order and intermediate result distribution. For example it will usually join small tables to large tables first and keep the distribution of the large table. But here large and small are based on post WHERE clause filtering and the guesses Redshift can make based on metadata. The further join is away from the source table metadata (deep into the join tree) the more uncertain Redshift is about what the incoming and outgoing data of the join will look like.
Here the EXPLAIN plan can give you hints about what Redshift is "thinking" - if you see a DIST INNER join Redshift is moving the data of one table (or intermediate result set) to match the other. If you DIST BOTH then Redshift is redistributing both sets of data to some new distribution (usually one of the join on columns). It does this to avoid having only 1 slice with data and all others with nothing to do as this would be very inefficient.
To sum up to see what Redshift is planning to do with your joins look at the EXPLAIN plan. You can also infer some info about intermediate result distribution from the explain plan but is doesn't provide a complete map of what it plans to do.