This isn't quite true - these comparisons don't evaluate to false, they evaluate to NULL.
I intuitively think of NULL as "unable to compute," a generalized NaN.
In the databases I've used, such as MSSQL[1], they evaluate to UNKNOWN.
[1]: https://docs.microsoft.com/en-us/sql/t-sql/language-elements...
create table #nulltest (a bit null);
insert into #nulltest(a) values (0),(1);
insert into #nulltest(a) values (null != null);However, that doesn't change things, I can't get any of these to parse:
insert into #nulltest(a) values (1 <> 0);
insert into #nulltest(a) values (null <> null);
insert into #nulltest(a) values (true and true);
insert into #nulltest(a) values (null and null);
The point of me asking is that it was my belief contrary to what was suggested, UNKNOWN is not the same as NULL and, crucially, does not evaluate as NULL.
Sql does not have this problem.
If every type has it's own equality operation (e.g. Number.eq) then I have no issue with it being weird in unexpected types (e.g. Number.eq(NULL, NULL) =false)
For example in JavaScript almost every use of === should be replaced by Object.is.
On the same note, the original intent was for NaN to be insidious and propagate into every derived calculation or comparison, thus making it easy to tell if a mathematical function had failed, without having to add conditionals at every step within it. Which makes a lot of sense. However, not being able to know exactly where the NaN was produced is kind of shit. Now, if NaN had a stack trace property like Error (exception) objects...it would make this behavior much more useful. Makes me wonder if the prototype of NaN can be extended, hm.
IMO NaN should have never existed, like signed zeros, (I am more agnostic about [signed] Infinity) I think that they were implemented in IEEE 754 because the encoding had space left for them.
I not an expert in any way on this this is literally a personal opinion.
Which large parts of SQL treat as equivalent to false.
WITH vars AS (
SELECT CAST(null AS INT) as X
)
SELECT 'yes this is a tautology' as "is this is a tautology"
FROM vars
WHERE X = X
The truth, imho, is that SQL has an archaic type system. In this day and age it should be offering algebraic types where we can implement our own "empty" values and the like that may not use three-value logic as appropriate, so that I can say that for example End Date is blank instead of either unknown or Date.MaxValue.
I don't see it that way. Null cast to boolean is always false. Null equality is undefined. There is no contradiction there.
It's not the `=` operator that is special or significant here. It is the `WHERE` clause semantics, which are defined as returning rows when the boolean expression is true. This treats both false and null cases as the other possible outcome, when rows are not returned. Not understanding this distinction is also where people write buggy queries using WHERE NOT ..., because like equality the NOT operator will return NULL for a NULL input (so NULL is not like false and does not negate to true).
SELECT 'hidden by null result' WHERE CAST(NULL AS boolean);
SELECT 'also hidden by null result' WHERE NOT CAST(NULL AS boolean);
SELECT 'shown by true result' WHERE CAST(NULL AS boolean) IS NULL;
SELECT 'hidden by false result' WHERE NOT (CAST(NULL AS boolean) IS NULL);
The existing GROUP BY behavior is different than WHERE because it does not have a special rule for handling true values. It just returns a group for each distinct value.
The prior (new default) UNIQUE constraint behavior made sense when you think of NULL as this bottom value and meaning "unknown", while also thinking that a uniqueness constraint is only applied to complete keys. The check is effectively deferred, and the partial or incomplete tuple might become valid and unique once its missing part(s) are filled in with real values. The new optional behavior makes more sense if you think that the uniqueness constraint enforces distinct records even with partial keys, just as the GROUP BY clause will determine distinct groups.
There is also a similar issue with FOREIGN KEY constraints. The default MATCH SIMPLE behavior is consistent with the default UNIQUE constraint behavior. An incomplete foreign key is essentially ignored and not subject to integrity checks. The MATCH FULL behavior is almost a style checker to say that a compound foreign key should be all null or all non-null, while still ignoring incomplete foreign keys. As far as I know, the MATCH PARTIAL case is not implemented, but I think it might complement this new UNIQUE NULLS NOT DISTINCT behavior?
One can imagine a further branching point in behaviors for foreign keys and I'm not sure how MATCH PARTIAL will be defined. One mode might be easily implementable with the new index method, if we want partial foreign keys to be able to reference partial keys in the referred table, i.e. the partial key with NULLs is considered an actual row key. But we might also imagine another partial matching mode where the the non-null components of a compound foreign key have to all exist in at least one key in the referred table, while ignoring the null fields as incompletely formed references. This would be an improvement on MATCH SIMPLE which allows the non-null columns in a foreign key to hold values which could not possibly match the referring table no matter which values are filled in for the remaining null fields of the foreign key. However, this enforcement would not be easily implemented without a powerset of indexes for all partial combinations allowed in the referencing tables!
But there is a clear winner here. Tabs won, almost all code files use spaces. Of course everyone uses the tab key on their keyboard, and the IDE makes spaces for you; no one is going to manually press space 4 times.
Is the debate really unsettled?
MSSQL considers nulls to be equal, everyone else considers them to be distinct (aside from a few DBs which just refuse to create UNIQUE indexes on nullable columns), and Oracle is apparently huffling glue in a corner.
> As an example of the irreconciliable weirdness of NULL, consider that "NULL = NULL" is false, and so is "NULL != NULL", while rows with NULL still group together in GROUP BY.
Group by is the exception here. And arguably it is so because the alternative is never useful, or close enough as to not be worth it.
Sounds pretty unsettled to me.
Overall I remember running into this but finding it fairly simple to add a couple specific checks around the condition where it came up in our code.
Oracle introduced VARCHAR2 in the 90s(?) to mean “Oracle semantics” as opposed to VARCHAR which is documented as “may change to standard SQL semantics in a future version”. However I don’t think they’ll ever materialize that distinction.
SQL is old enough and this debate so unsettled still that I think it should be clear there isn't a categorically right behavior here anymore than there is a clear winner between tabs and spaces.
As an example of the irreconciliable weirdness of NULL, consider that "NULL = NULL" is false, and so is "NULL != NULL", while rows with NULL still group together in GROUP BY.
I appreciate you giving folks the option.