Error converting NVARCHAR to Float using a CAST in T-SQL - sql-server

Question

I have a table which has values in the r_version_label column like:
*CURRENT*, *LATEST*, 0.1, 0.2, 0.3, *0.8.5,* 1.0, 1.1
The CURRENT, LATEST and legacy version numbers such as 0.8.5 I can ignore.
I am writing SQL as below:
WITH cte_version_label AS
(
SELECT DISTINCT r_version_label
FROM pharma_document_rp
WHERE r_version_label LIKE '%.%'
AND r_version_label NOT LIKE '%.%.%'
)
SELECT *
FROM cte_version_label
WHERE CAST(r_version_label AS float) = 0.1
But I am getting:
Msg 8114, Level 16, State 5, Line 1
Error converting data type nvarchar to float.
I can however do this:
WITH cte_version_label AS
(
SELECT DISTINCT r_version_label
FROM pharma_document_rp
WHERE r_version_label LIKE '%.%'
AND r_version_label NOT LIKE '%.%.%'
)
SELECT CAST(r_version_label AS float)
FROM cte_version_label
Which returns all the right values without error.
So why can't I cast in the WHERE clause, but can in the SELECT clause? Obviously there is not really a CAST issue as I am removing the offending items, otherwise the SELECT CAST would not work.
The issue is, I need to run a python script reading in version numbers from Excel and then look these up in the table. Excel converts 1.0 into 1. So I need the whole query to operate using "floats" not the string type version stored in the database.

Answer 1

John's suggestion to use try_convert is definitely a better option.
But in response to WHY the second query works and the first doesn't, have a look at the execution plans.
On my instance (SQL 2017 Enterprise) this is the Estimated execution plan of the first query (can't use the actual because the query errors out).
Have a look at the predicate used in the first node. It's trying to do the CAST (internally using CONVERT) in the first operation on your whole table. When that hits something like 0.8.5 it bails.
Now let's look at the execution plan for your second query that works (this one is the Actual execution plan).
Notice the predicate in the first node - it's just your string filter. The CAST does not happen until later down the execution chain, in the Compute Scalar node, AFTER values that offend the CAST have already been filtered out.