Matthew Kudija

Home About Notes Reading Notes

SQL Code Snippets

Create & Update Tables

Create schema:

-- create schema
CREATE SCHEMA my_schema;

Create table:

-- create table
DROP TABLE IF EXISTS schema.table;
CREATE TABLE schema.table
    AS (
        SELECT * FROM table
    );

Create or replace table:

-- create or replace table
CREATE OR REPLACE TABLE schema.table AS (
    SELECT *
    FROM differentschema.table
)

Update table:

DELETE
FROM table
WHERE source = 'a';

INSERT INTO table
SELECT * FROM source_table

Grant usage:

GRANT USAGE ON SCHEMA <name> TO user;
GRANT SELECT ON ALL TABLES IN SCHEMA <name TO user;

GRANT USAGE ON SCHEMA <name> TO GROUP readonly;
GRANT SELECT ON ALL TABLES IN SCHEMA <name> TO GROUP readonly;

GRANT USAGE ON [schema] TO mode_analytics;
GRANT SELECT ON [schema.table] TO mode_analytics;

Search for Column Names

SELECT *  
FROM INFORMATION_SCHEMA.COLUMNS  
WHERE COLUMN_NAME LIKE '%search%'  
ORDER BY TABLE_NAME;

Big Query:

SELECT DISTINCT * 
FROM `gcp-hopper-etldata-production.warehouse.INFORMATION_SCHEMA.COLUMNS` WHERE column_name LIKE '%session_id%'

Get Tables in Schema

select t.table_name
from information_schema.tables t
where t.table_schema = 'schema_name'  -- put schema name here
      and t.table_type = 'BASE TABLE'
order by t.table_name;

Joins

on

FROM information_schema.tables t
    JOIN information_schema.columns c 
           on c.table_name = t.table_name 
           and c.table_schema = t.table_schema

using

FROM information_schema.tables t
    JOIN information_schema.columns c USING(table_name, table_schema)

Find tables with colName column

select t.table_schema,
       t.table_name
from information_schema.tables t
inner join information_schema.columns c 
           on c.table_name = t.table_name 
           and c.table_schema = t.table_schema
where c.column_name = 'colName'
      and t.table_schema not in ('information_schema', 'pg_catalog')
      and t.table_type = 'BASE TABLE'
order by t.table_schema;

Unique items in column

SELECT 
    t.device
FROM 
    database.table t
GROUP BY t.device

	device
1	"iPhone10,1"
2	"iPhone10,2"
3	"iPhone10,3"
4	"iPhone10,4"
5	"iPhone10,5"

value_count of column

SELECT 
    value_count_col
    ,COUNT(id_col) AS value_count
FROM 
    database.table t
GROUP BY 1
ORDER BY 2 DESC

SELECT 
    t.device
    ,COUNT(t.distance) AS value_count
FROM 
    database.table t
GROUP BY t.device
ORDER BY value_count DESC

	device	value_count
1	iPhone10,5	23832
2	iPhone10,2	14553
3	iPhone10,4	3487
4	iPhone10,3	3078
5	iPhone10,1	23

np.where equivalent for cases

CASE
    WHEN t.score BETWEEN 0  AND 24  THEN '0-24'
    WHEN t.score BETWEEN 25 AND 49  THEN '25-49'
    WHEN t.score BETWEEN 50 AND 74  THEN '50-74'
    WHEN t.score BETWEEN 75 AND 100 THEN '75-100'
    ELSE 'OTHER'
END AS score_group

Select only first/last record using RANK() OVER PARTITION BY

In this example, there may be many records for a given user_id, but we only want to select the most recent record for each user_id. Alternatively, we could select the first record by using ASC:

SELECT
    ...
    ...
    FROM(
    SELECT
        ...
    ...
        ,RANK() OVER (PARTITION BY t.user_id ORDER BY t.created_at DESC) AS RANK
        FROM
            database.table AS t
    )
WHERE RANK = 1

Cumulative Value SUM() OVER PARTITION BY

In this value, LTV is a monthly value that we want to show cumulative amount:

SELECT
    ...
    , SUM(ltv_prod)  
      OVER (PARTITION BY policy_effective_month  
      ORDER BY calendar_month ASC ROWS UNBOUNDED PRECEDING) AS ltv_cumulative
    FROM
            database.table AS t
    )
WHERE RANK = 1

Round

ROUND rounds the number, and if you don't want the additional .000 you can CAST as an integer:

CAST(ROUND(col_name, 0) AS INT) AS col_name_rounded

Decile

Use NTILE function (link):

ntile(20) OVER (PARTITION BY state ORDER BY score) AS score_ntile_state,

To get actual deciles you can use a CASE statement:

SELECT 
    score_decile, 
    COUNT(score_decile) AS count
FROM (
    SELECT
        *,
        CASE
            WHEN score < 10 THEN '0-9'
            WHEN score > 9 AND score < 20 THEN '10-19'
            WHEN score > 19 AND score < 30 THEN '20-29'
            WHEN score > 29 AND score < 40 THEN '30-39'
            WHEN score > 39 AND score < 50 THEN '40-49'
            WHEN score > 49 AND score < 60 THEN '50-59'
            WHEN score > 59 AND score < 70 THEN '60-69'
            WHEN score > 69 AND score < 80 THEN '70-79'
            WHEN score > 79 AND score < 90 THEN '80-89'
            WHEN score > 89 THEN '90-100'
        END as score_decile
    FROM database.table
)
GROUP BY score_decile
ORDER BY score_decile

A quick trick is to ROUND with -1 to approximate deciles if score is in the range from 0-100. This is not exact as your buckets will be 0-5, 6-15...86-95, 96-100 so the first and last buckets will be small, but if you are just comparing between two distributions and care less about the absolute distribution, this may be an easy trick.

SELECT 
    score_decile_approx, 
    COUNT(score_decile_approx) AS count
FROM (
    SELECT
        *,
        ROUND(score, -1) AS score_decile_approx
    FROM database.table
)
GROUP BY score_decile_approx
ORDER BY score_decile_approx

Finally, we may want the rate rather than the count, which we can get with the RATIO_TO_REPORT(count) OVER () AS rate command. This gives rate for each row as a fraction of the sum of count count column.

SELECT 
    score_decile_approx, 
    COUNT(score_decile_approx) AS count,
    RATIO_TO_REPORT(count) OVER () AS rate
FROM (
    SELECT
        *,
        ROUND(unmapped_score, -1) AS score_decile_approx
    FROM database.table

)
GROUP BY score_decile_approx
ORDER BY score_decile_approx

Equal Size Buckets

WITH buckets AS (
    SELECT
        value
        , NTILE(5) OVER (ORDER BY value ASC) AS value_bucket
    FROM table
)

SELECT 
    value_bucket
    , MIN(value) AS bucket_lower
    , MAX(value) AS bucket_upper
FROM buckets
GROUP BY 1
ORDER BY 1

see also Bigquery RANGE_BUCKET() to assign values rather than case

Ratio over groupby

Above we use the RATIO_TO_REPORT command to get a ratio rather than count. If we want a time series showing percent, and it is groupe by time (week, month, etc.), this is how you get a rate for each group rather than the whole table:

    RATIO_TO_REPORT(account_count) OVER(PARTITION BY DATE_TRUNC('month', account_timestamp))

or

    DATE_TRUNC('quarter', account_timestamp) AS x,
    RATIO_TO_REPORT(account_count) OVER(PARTITION BY x)

Number of items and most recent in table

SELECT 
    COUNT(*), 
    MAX(created_at) 
FROM 
    database.table

Rolling Mean

SELECT 
    DATE_TRUNC('day', profile_timestamp) AS x,
    COUNT(DISTINCT account_id) AS new_user_count,
    AVG(new_user_count) OVER (ORDER BY x ROWS BETWEEN 4 PRECEDING AND CURRENT ROW) AS count_rolling
FROM 
    database.table

Percent Change

Display the % symbol:

    CONCAT(ROUND((edw_PLE_months_30_4 - edw_PLE_months_30) / edw_PLE_months_30 * 100, 2),'\%') AS percent_diff

Percent of Total

round(sum(payout_usd)) as payouts,
round(sum(payout_usd) /  sum(sum(payout_usd)) over (),3) as pct_of_total,

or

earned_premium/SUM(earned_premium) OVER () as percent_of_total

or

RATIO_TO_REPORT(earned_premium) OVER () AS percent_of_total

Optimize Query

If you run the query with EXPLAIN on top it will give you the query plan and how costly each step is.

Pivot Table

You can pivot, but you need to manually specify the column names:

WITH pre AS (  
    SELECT state  
         , amount  
         , month  
    FROM state_mgt.planned_rate  
)  
SELECT *  
FROM pre PIVOT(SUM(amount) FOR MONTH IN ('2022-02-28','2022-03-30','2022-04-30'));

You can use this snippet to get the column names:

select string_agg(concat("'",month,"'")) from (select distinct month from pre)

Most Recent Month-Start Date (Prior Month Start)

select date_trunc(date_sub(current_date(), interval 1 month), month)

Most Recent Month-End Date (Prior Month End)

select last_day(date_sub(current_date(), interval 1 month))

Month Start Date

SELECT DATE_TRUNC(CURRENT_DATE,MONTH) AS calendar_month

View Tables in Schema

-- view tables in schema
SELECT t.table_name
FROM information_schema.tables t
WHERE t.table_schema = 'schema' -- put schema name here
    AND t.table_type = 'BASE TABLE'
ORDER BY t.table_name;

Lag (Offset column by n rows)

SELECT
    col
    LAG(col, 1) OVER (ORDER BY calendar_month ASC) AS col_lag_7

Development Months and Terms

dev_mo starts at 0, term starts at 1:

, DATEDIFF('month', policy_effective_month, calendar_month) AS dev_month  
, dev_month / 6 + 1 AS term

Dates

Last 7 days of data

SELECT 
    *
FROM 
    database.table t
WHERE
    t.created_at > GETDATE() - INTERVAL '7 days'

DATE_DIFF

DATE_DIFF('month', start_month, calendar_month) AS age

DATE_TRUNC

To transform a timestamp into weekly or daily etc. data use DATE_TRUNC(). Available dateparts are listed here.

GCP/BQ

DATE_TRUNC(timestamp, WEEK) AS week

SELECT
    DATE_TRUNC('day', timestamp) AS day

or

SELECT
    DATE_TRUNC('week', timestamp) AS week

or

SELECT
    DATE_TRUNC('month', timestamp) AS month

or

SELECT
    DATE_TRUNC('quarter', timestamp) AS quarter

Get month offset from current date (also end of month)

• CURRENT_DATE to get current date
• DATEADD to offset by a number of months
• LAST_DAY to get last day of month

policy_inception_month = LAST_DAY(DATEADD(MM,-6, CURRENT_DATE))

or

WHERE prediction_date = LAST_DAY(DATE_ADD('month', -1, CURRENT_DATE))

Current Date

SELECT LEFT(GETDATE(),10)

Select * Except Columns

SELECT
joined.*,
air.* EXCEPT(purchase_date,purchase_week,purchase_month),
hotel.* EXCEPT(purchase_date,purchase_week,purchase_month),

GCP example: flex_dashboard_july2022

Array of Dates

SELECT m AS calendar_month
FROM UNNEST(GENERATE_DATE_ARRAY('2023-01-01', CURRENT_DATE(), INTERVAL 1 MONTH)) AS m

Z-score Anomaly Detection

ref: Simple Anomaly Detection Using Plain SQL | Haki Benita

WITH data AS (
    SELECT
    event_date
    , dim1
    , dim2
    , value
  FROM table
  GROUP BY 1,2,3
)

, averages AS (
  SELECT 
    *
    , ROUND(AVG(value) OVER (PARTITION BY dim1, dim2 ORDER BY event_date ROWS BETWEEN 21 PRECEDING AND 1 PRECEDING)) AS value_avg
    , ROUND(STDDEV(value) OVER (PARTITION BY dim1, dim2 ORDER BY event_date ROWS BETWEEN 21 PRECEDING AND 1 PRECEDING),3) AS value_std
  FROM data
)

, thresholds AS (
  SELECT
    3.0 AS z_thresh -- zscore anomaly threshold 
)

, zscores AS (
  SELECT 
    *
    , (value - value_avg) / NULLIF(value_std, 0) as value_zscore
    , (exercise_count - count_avg) / NULLIF(count_std, 0) as count_zscore
    , (cpe - cpe_avg) / NULLIF(cpe_std, 0) as cpe_zscore
    , (SELECT z_thresh FROM thresholds) AS z_thresh
  FROM averages
)

, bounds AS (
  SELECT 
    *
    , value_avg + value_std * z_thresh AS value_bound_upper
    , value_avg - value_std * z_thresh AS value_bound_lower
  FROM zscores
)

, anomalies AS (
  SELECT 
    *
    , CASE WHEN value_zscore < -z_thresh OR value_zscore > z_thresh THEN 1 ELSE 0 END AS value_anomaly
  FROM bounds
)

SELECT *
FROM anomalies
WHERE event_date BETWEEN DATE_ADD(CURRENT_DATE(), INTERVAL -120 DAY) AND CURRENT_DATE()
ORDER BY 1 DESC,2,3

SELECT EXCEPT

SELECT
    * EXCEPT(col1, col2)

30 Days Ago

DATE_ADD(CURRENT_DATE(), INTERVAL -30 DAY)

Last day of prior month

date_sub(date_trunc(current_date(), month), interval 1 day)

Monthly Run-Rate

with run_rate as (
  select 
    extract(day from current_date()) - 1 as t_0,
    extract(day from last_day(current_date())) as t_1,
    extract(day from last_day(current_date())) / (extract(day from current_date()) - 1) as rr,
)

select
  last_day(date(action_timestamp)) as action_month,
  sum(value) as value,
  sum(value)*(select rr from run_rate) as value_rr,
from `table`
where date(action_timestamp) < current_date()
  and last_day(date(action_timestamp)) = last_day(current_date())
group by 1

Loop over Date Array

Consider using a cross join below instead of a loop...

drop table `table-name`;

create table `table-name` (
ref_date date,
today_flag string,
another_col string,
);

------------------------------------------------------------------------------

for ref_date in (
select 
m as ref_date
from unnest(generate_date_array(current_date(), date(current_date() + interval 7 day), interval 1 day)) AS m
)

do insert into `table-name`
(
  select 
    ref_date.ref_date,
    if(ref_date.ref_date = current_date(),'today','not today') as today_flag,
    'this' as another_col
);

end for;

------------------------------------------------------------------------------

select * from `table-name` order by 1;

Cross Join to Loop

Instead of a loop, use a cross join, which finds all the unique combinations between two tables.

with dates as (
  select 
    m as ref_date
  from unnest(generate_date_array(current_date(), date(current_date() + interval 3 day), interval 1 day)) AS m
)

, items as (
  select 
    m as row_no
  from unnest(generate_array(1,3)) AS m
)

select *
from dates, items
-- from dates cross join items 
order by 1,2

Resources

• SQLBolt - Learn SQL
• SQLZOO
• SQL Query Order of Execution
• Database Design for Mere Mortals
• The SQL Murder Mystery
  • Congrats, you found the brains behind the murder! Everyone in SQL City hails you as the greatest SQL detective of all time. Time to break out the champagne!
• spyql: Query data on the command line with SQL-like SELECTs powered by Python expressions
• The Data Warehouse Toolkit

---

Created: 2019-06-25-Tue
Updated: 2025-03-07-Fri