Assign a variable a dynamic value in SQL in Databricks / Spark

Question

I feel like I must be missing something obvious here, but I can't seem to dynamically set a variable value in Spark SQL.

Let's say I have two tables, tableSrc and tableBuilder, and I'm creating tableDest.

I've been trying variants on

SET myVar FLOAT = NULL

SELECT
    myVar = avg(myCol)
FROM tableSrc;

CREATE TABLE tableDest(
    refKey INT,
    derivedValue FLOAT
);


INSERT INTO tableDest
    SELECT
        refKey,
        neededValue * myVar AS `derivedValue`
    FROM tableBuilder

Doing this in T-SQL is trivial, in a surprising win for Microsoft (DECLARE...SELECT). Spark, however, throws

Error in SQL statement: ParseException: mismatched input 'SELECT' expecting <EOF>(line 53, pos 0)

but I can't seem to assign a derived value to a variable for reuse. I tried a few variants, but the closest I got was assigning a variable to a string of a select statement.

Please note that this is being adapted from a fully functional script in T-SQL, and so I'd just as soon not split out the dozen or so SQL variables to compute all those variables with Python spark queries just to insert {var1}, {var2}, etc in a multi hundred line f-string. I know how to do this, but it will be messy, difficult, harder to read, slower to migrate, and worse to maintain and would like to avoid this if at all possible.

Answer 1

The SET command used is for spark.conf get/set, not a variable for SQL queries

For SQL queries you should use widgets:

https://docs.databricks.com/notebooks/widgets.html

But, there is a way of using spark.conf parameters on SQL:

%python spark.conf.set('personal.foo','bar')

Then you can use:

$sql select * from table where column = '${personal.foo}';

The trick part is that you have to use a "dot" (or other special character) on the name of the spark.conf, or SQL cells will expect you to provide value to the $variable on run time (It looks like a bug to me, i believe rounding with {} should be enough)

Answer 2

You are missing a semi-colon at the end of the variable assignment.

SET myVar FLOAT = NULL;
...

Hope it helps :)

Answer 3

I've circled around this issue for a long time. Finally, I've found a workaround using @Ronieri Marques solution plus some pyspark functions. I'll try to provide a full working code below:

first I create a sample table:

%sql
create table if not exists calendar
as 
select '2021-01-01' as date
union
select '2021-01-02' as date
union
select '2021-01-03' as date

%sql 
-- just to show the max and min dates
select max(date), min(date) from calendar

Combining sqlContext + toJSON it is possible to dynamically assign a value to the variable, in this case I use a query:

%python
result = sqlContext.sql("select max(date), min(date) from calendar").toJSON()
spark.conf.set('date.end'    , result.first()[14:24])
spark.conf.set('date.start'  , result.first()[39:49])

Finally it will be possible to use the variables inside a SQL query:

%sql 
select * from calendar where date > '${date.start}' and date < '${date.end}'

Note that the substring result.first()[14:24] and result.first()[39:49] are necessary because the value of result.first() is {"max(date)":"2021-01-03","min(date)":"2021-01-01"} so we need to "tailor" the final result picking up only the values we need.

Probably the code can be polished but right now it is the only working solution I've managed to implement.

I hope this solution could be useful for someone.