There are many senior people or folks with families/etc that don't have the time or patience for this ridiculous cargo cult hazing ritual masquerading as a hiring technique.
I would say your assertion that "leetcode is a proxy for IQ" is a bold statement to make without some bold sources to back it up.
I've personally met a number of people that are great at these small scoped problems but suffer at broader systems engineering.
Companies that get a lot of applicants care more about false positives than false negatives. So they're fine not hiring a lot of qualified people, as long as there's a low chance of accidentally hiring an unqualified person. Apparently LeetCode-style problems work for that type of screening.
However, if you are getting resumes on the scale that the larger, well known tech companies do, then the individual attention cannot be paid on each resume. At that point, one looks more at a quick filter to try to remove the risky hires.
If you've got 1000 resumes and 80% of them are bad hires, and after an online assessment that takes it down to 200 resumes of which only 50% would prove to be bad hires, that is a significantly improved pool to consider.
Yes, it is possible that a great candidate was removed in that filter, but its also possible that one still remains in that pool.
Startups may be cargo cutting big tech interview processes, but the process has value for big tech companies and any others that have more resumes than they can reasonably deal with.
95% of programming is data structures and algorithms because that's where all the performance gains are usually made. The other 5% is knowledge of the hardware and micro-optimisations like avoiding branch mispredicts and understanding the compiler.
Hashing in databases is very similar to how hash-tables work as is sharding.
"95% of programming is data structures and algorithms" - it's just not. 95 percent of programming is gluing together libraries, refactoring code and reorganising what you have to solve new business requirements.
Your focus on performance (albeit important) above all else makes me question how much real world experience you have building large scale systems. (Unless you're a systems/embedded developer).
But this is the crux of the issue: using metrics that don't map well to what we're trying to optimize for. Perhaps your trying to optimize for finding great performance focused, low level software devs? If so that's different that my goal.
Some of these questions are perfectly valid during LC coding rounds. Others would be more appropriate in a system design round.
Leetcode can never compete with or allow a candidate to demonstrate real multi-dimensional or multi-systems thinking. You may be able to write some bass ackwards arcane algorithm and prove you’re “31337” but can you make that work at scale with a data store that you have no information about? How’s it going to get the data to operate on in the first place? Will your data retrieval query cripple a prod database? Read from replica? How far behind can that data be without making your fancy algorithm useless? How will you parallelize that work set? In-request or background job? Multi-threaded or single thread? What about changing state in multi threaded contexts? Database indexes? Schema design? How will you handle failures in your algorithm? How will you handle malformed data? And most importantly, does what you’re doing benefit the user, or just your epeen?
Leetcode is bullshit. It’ll never let somebody show you they’re good at anything other than solving bullshit problems. And you’ll lose everybody who CAN work within the larger context by chasing candidates who’s chief skill is solving bullshit problems.