Maybe inconsistent handling of implicit timestamp<->bigint conversion

[nyh]

This is a spinoff of https://github.com/scylladb/scylla-dtest/pull/3247, where we noticed that SELECT mintimeuuid(writetime(x)) works in Cassandra, but on Scylla it fails saying:

A WRITETIME produces a org.apache.cassandra.db.marshal.LongType value, which doesn't match the type: org.apache.cassandra.db.marshal.TimestampType of arg0(system.mintimeuuid)"

It's actually nice that mintimeuuid(writetime(x)) doesn't work - writetime(x) returns a USING TIMESTAMP value, typically in microseconds, while mintimeuuid() expects a timestamp-type value, which is in milliseconds, and the two don't match, so it's good we don't accept it although Cassandra does. We can consider this a Cassandra bug, and so far so good.

But something still worries me... If g is a bigint (64-bit) number, the following select does work on both Scylla and Cassandra:

SELECT mintimeuuid(g)

And yet, the following select does not work in Scylla (it does work on Cassandra):

SELECT mintimeuuid(writetime(x)) 

How can this be? both g and writetime(x) have the same type - bigint. Either bigint can be implicitly converted to timestamp, or it can't. So how come this conversion works for the expression g but not for writetime(x)? Maybe we have a bug in the implicit conversion code?

Again, I'm not too worried about this specific example, where I even like the fact that the implicit conversion didn't work. But I'm worried that it indicates a bug that we did something about implicit conversions inconsistently or wrongly, and it might bite us in the future in other scenarios.

CC @avikivity @cvybhu

[cvybhu]

The code that prepares expressions is often inconsistent in how it checks type compatibility - sometimes it's a direct type comparison, and sometimes test_assignment (which allows implicit conversions). I hope to unify this soon, because it leads to inconsistencies like this.

Although this is selector code, so maybe the issue is somewhere there.

[nyh]

I found another example of an implicit conversion that Scylla does and Cassandra does:
If you have a counter column c, you can run SELECT bigintasblob(c) FROM tbl WHERE p=1.
In Cassandra this doesn't work - bigintasblob() insists on getting an actual bigint, not a counter, even though they are the same size.

Suprisingly, the opposite direction is rejected in Scylla not just in Cassandra - SELECT counterasblob(i) where i is a bigint column.

So I don't understand when Scylla allows, and when it doesn't, these implicit conversions of function arguments.

[nyh]

Interestingly, the issue #5552 was about mintimeuuid(writetime(g)) being accepted but then crashing Scylla. Commit 1078c86 by @bhalevy fixed it by detecting out-of-range timestamps (as it turns out in #17035 this forgot to exclude negative timestamps), but this detection is no longer needed to solve the original problem of mintimeuuid(writetime(g)), which is now rejected.