Howdy! I'm Matt Mazur and I'm a data analyst who has worked at several startups to help them use data to grow their businesses. This guide is an attempt to document my preferences for formatting SQL in the hope that it may be of some use to others. If you or your team do not already have a SQL style guide, this may serve as a good starting point which you can adopt and update based on your preferences.
Also, I'm a strong believer in having Strong Opinions, Weakly Held so if you disagree with any of this, drop me a note, I'd love to discuss it.
If you're interested in this topic, you may also enjoy my LookML Style Guide or my blog where I write about analytics and data analysis.
Simplified Chinese version here: 中文版
Here's a non-trivial query to give you an idea of what this style guide looks like in the practice:
with hubspot_interest as (
select
email,
timestamp_millis(property_beacon_interest) as expressed_interest_at
from hubspot.contact
where
property_beacon_interest is not null
),
support_interest as (
select
conversation.email,
conversation.created_at as expressed_interest_at
from helpscout.conversation
inner join helpscout.conversation_tag on conversation.id = conversation_tag.conversation_id
where
conversation_tag.tag = 'beacon-interest'
),
combined_interest as (
select * from hubspot_interest
union all
select * from support_interest
),
first_interest as (
select
email,
min(expressed_interest_at) as expressed_interest_at
from combined_interest
group by email
)
select * from first_interest
It's just as readable as uppercase SQL and you won't have to constantly be holding down a shift key.
-- Good
select * from users
-- Bad
SELECT * FROM users
-- Bad
Select * From users
When selecting columns, always put each column name on its own line and never on the same line as select
. For multiple columns, it's easier to read when each column is on its own line. And for single columns, it's easier to add additional columns without any reformatting (which you would have to do if the single column name was on the same line as the select
).
-- Good
select
id
from users
-- Good
select
id,
email
from users
-- Bad
select id
from users
-- Bad
select id, email
from users
When selecting *
it's fine to include the *
next to the select
and also fine to include the from
on the same line, assuming no additional complexity like where
conditions:
-- Good
select * from users
-- Good too
select *
from users
-- Bad
select * from users where email = '[email protected]'
Similarly, conditions should always be spread across multiple lines to maximize readability and make them easier to add to. Operators should be placed at the end of each line:
-- Good
select *
from users
where
email = '[email protected]'
-- Good
select *
from users
where
email like '%@domain.com' and
created_at >= '2021-10-08'
-- Bad
select *
from users
where email = '[email protected]'
-- Bad
select *
from users
where
email like '%@domain.com' and created_at >= '2021-10-08'
-- Bad
select *
from users
where
email like '%@domain.com'
and created_at >= '2021-10-08'
Some IDEs have the ability to automatically format SQL so that the spaces after the SQL keywords are vertically aligned. This is cumbersome to do by hand (and in my opinion harder to read anyway) so I recommend just left aligning all of the keywords:
-- Good
select
id,
email
from users
where
email like '%@gmail.com'
-- Bad
select id, email
from users
where email like '%@gmail.com'
Some SQL dialects like BigQuery support using double quotes, but for most dialects double quotes will wind up referring to column names. For that reason, single quotes are preferable:
-- Good
select *
from users
where
email = '[email protected]'
-- Bad
select *
from users
where
email = "[email protected]"
If your SQL dialect supports double quoted strings and you prefer them, just make sure to be consistent and not switch between single and double quotes.
Simply because !=
reads like "not equal" which is closer to how we'd say it out loud.
-- Good
select
count(*) as paying_users_count
from users
where
plan_name != 'free'
-- Good
select
id,
email
from users
-- Bad
select
id
, email
from users
While the commas-first style does have some practical advantages (it's easier to spot missing commas and results in cleaner diffs), I'm just not a huge fan of how they look so prefer commas-last.
-- Good
select *
from users
where
id in (1, 2)
-- Bad
select *
from users
where
id in ( 1, 2 )
-- Good
select *
from users
where
email in (
'[email protected]',
'[email protected]',
'[email protected]',
'[email protected]'
)
-- Good
select *
from users
-- Good
select *
from visit_logs
-- Bad
select *
from user
-- Bad
select *
from visitLog
-- Good
select
id,
email,
timestamp_trunc(created_at, month) as signup_month
from users
-- Bad
select
id,
email,
timestamp_trunc(created_at, month) as SignupMonth
from users
- Boolean fields should be prefixed with
is_
,has_
, ordoes_
. For example,is_customer
,has_unsubscribed
, etc. - Date-only fields should be suffixed with
_date
. For example,report_date
. - Date+time fields should be suffixed with
_at
. For example,created_at
,posted_at
, etc.
Put the primary key first, followed by foreign keys, then by all other columns. If the table has any system columns (created_at
, updated_at
, is_deleted
, etc.), put those last.
-- Good
select
id,
name,
created_at
from users
-- Bad
select
created_at,
name,
id,
from users
Better to be explicit so that the join type is crystal clear:
-- Good
select
users.email,
sum(charges.amount) as total_revenue
from users
inner join charges on users.id = charges.user_id
-- Bad
select
users.email,
sum(charges.amount) as total_revenue
from users
join charges on users.id = charges.user_id
By doing it this way it makes it easier to determine if your join is going to cause the results to fan out:
-- Good
select
...
from users
left join charges on users.id = charges.user_id
-- primary_key = foreign_key --> one-to-many --> fanout
select
...
from charges
left join users on charges.user_id = users.id
-- foreign_key = primary_key --> many-to-one --> no fanout
-- Bad
select
...
from users
left join charges on charges.user_id = users.id
-- Good
select
users.email,
sum(charges.amount) as total_revenue
from users
inner join charges on users.id = charges.user_id
group by email
-- Bad
select
users.email,
sum(charges.amount) as total_revenue
from users
inner join charges
on users.id = charges.user_id
group by email
When you have mutliple join conditions, place each one on their own indented line:
-- Good
select
users.email,
sum(charges.amount) as total_revenue
from users
inner join charges on
users.id = charges.user_id and
refunded = false
group by email
It can be tempting to abbreviate table names like users
to u
and charges
to c
, but it winds up making the SQL less readable:
-- Good
select
users.email,
sum(charges.amount) as total_revenue
from users
inner join charges on users.id = charges.user_id
-- Bad
select
u.email,
sum(c.amount) as total_revenue
from users u
inner join charges c on u.id = c.user_id
Most of the time you'll want to type out the full table name.
There are two exceptions:
If you you need to join to a table more than once in the same query and need to distinguish each version of it, aliases are necessary.
Also, if you're working with long or ambiguous table names, it can be useful to alias them (but still use meaningful names):
-- Good: Meaningful table aliases
select
companies.com_name,
beacons.created_at
from stg_mysql_helpscout__helpscout_companies companies
inner join stg_mysql_helpscout__helpscout_beacons_v2 beacons on companies.com_id = beacons.com_id
-- OK: No table aliases
select
stg_mysql_helpscout__helpscout_companies.com_name,
stg_mysql_helpscout__helpscout_beacons_v2.created_at
from stg_mysql_helpscout__helpscout_companies
inner join stg_mysql_helpscout__helpscout_beacons_v2 on stg_mysql_helpscout__helpscout_companies.com_id = stg_mysql_helpscout__helpscout_beacons_v2.com_id
-- Bad: Unclear table aliases
select
c.com_name,
b.created_at
from stg_mysql_helpscout__helpscout_companies c
inner join stg_mysql_helpscout__helpscout_beacons_v2 b on c.com_id = b.com_id
When there are no join involved, there's no ambiguity around which table the columns came from so you can leave the table name out:
-- Good
select
id,
name
from companies
-- Bad
select
companies.id,
companies.name
from companies
But when there are joins involved, it's better to be explicit so it's clear where the columns originated:
-- Good
select
users.email,
sum(charges.amount) as total_revenue
from users
inner join charges on users.id = charges.user_id
-- Bad
select
email,
sum(amount) as total_revenue
from users
inner join charges on users.id = charges.user_id
-- Good
select
count(*) as total_users
from users
-- Bad
select
count(*)
from users
-- Good
select
timestamp_millis(property_beacon_interest) as expressed_interest_at
from hubspot.contact
where
property_beacon_interest is not null
-- Bad
select
timestamp_millis(property_beacon_interest)
from hubspot.contact
where
property_beacon_interest is not null
-- Good
select *
from customers
where
is_cancelled = true
-- Good
select *
from customers
where
is_cancelled = false
-- Bad
select *
from customers
where
is_cancelled
-- Bad
select *
from customers
where
not is_cancelled
-- Good
select
id,
email,
timestamp_trunc(created_at, month) as signup_month
from users
-- Bad
select
id,
email,
timestamp_trunc(created_at, month) signup_month
from users
I prefer grouping by name, but grouping by numbers is also fine.
-- Good
select
user_id,
count(*) as total_charges
from charges
group by user_id
-- Good
select
user_id,
count(*) as total_charges
from charges
group by 1
-- Bad
select
timestamp_trunc(created_at, month) as signup_month,
vertical,
count(*) as users_count
from users
group by 1, vertical
-- Good
select
timestamp_trunc(com_created_at, year) as signup_year,
count(*) as total_companies
from companies
group by signup_year
-- Bad
select
timestamp_trunc(com_created_at, year) as signup_year,
count(*) as total_companies
from companies
group by timestamp_trunc(com_created_at, year)
-- Good
select
timestamp_trunc(com_created_at, year) as signup_year,
count(*) as total_companies
from companies
group by signup_year
-- Bad
select
count(*) as total_companies,
timestamp_trunc(com_created_at, year) as signup_year
from mysql_helpscout.helpscout_companies
group by signup_year
Each when
should be on its own line (nothing on the case
line) and should be indented one level deeper than the case
line. The then
can be on the same line or on its own line below it, just aim to be consistent.
-- Good
select
case
when event_name = 'viewed_homepage' then 'Homepage'
when event_name = 'viewed_editor' then 'Editor'
else 'Other'
end as page_name
from events
-- Good too
select
case
when event_name = 'viewed_homepage'
then 'Homepage'
when event_name = 'viewed_editor'
then 'Editor'
else 'Other'
end as page_name
from events
-- Bad
select
case when event_name = 'viewed_homepage' then 'Homepage'
when event_name = 'viewed_editor' then 'Editor'
else 'Other'
end as page_name
from events
Avoid subqueries; CTEs will make your queries easier to read and reason about.
When using CTEs, pad the query with new lines.
If you use any CTEs, always select *
from the last CTE at the end. That way you can quickly inspect the output of other CTEs used in the query to debug the results.
Closing CTE parentheses should use the same indentation level as with
and the CTE names.
-- Good
with ordered_details as (
select
user_id,
name,
row_number() over (partition by user_id order by date_updated desc) as details_rank
from billingdaddy.billing_stored_details
),
first_updates as (
select
user_id,
name
from ordered_details
where
details_rank = 1
)
select * from first_updates
-- Bad
select
user_id,
name
from (
select
user_id,
name,
row_number() over (partition by user_id order by date_updated desc) as details_rank
from billingdaddy.billing_stored_details
) ranked
where
details_rank = 1
-- Good
with ordered_details as (
-- Bad
with d1 as (
Leave it all on its own line:
-- Good
select
user_id,
name,
row_number() over (partition by user_id order by date_updated desc) as details_rank
from billingdaddy.billing_stored_details
-- Okay
select
user_id,
name,
row_number() over (
partition by user_id
order by date_updated desc
) as details_rank
from billingdaddy.billing_stored_details
This style guide was inspired in part by:
Hat-tip to Peter Butler, Dan Wyman, Simon Ouderkirk, Alex Cano, Adam Stone, Brian Kim, and Claire Carroll for providing feedback on this guide.