Sophisticated throughput-optimized systems rely on deep latency-hiding. Schedulers see millions of atomic operations into the future, continuously rewriting the schedule globally to maximize locality and minimize resource contention based on real-time changes to workload, resource availability, and system behaviors.
In short, for each of the millions of in-flight operations (which might only map to a handful of user operations), it is trying to precisely optimize the concurrency, timing, and dependency sequencing such that when operations are executed every resource required is hot, uncontended, and available with high probability. When this works well it dramatically reduces the number of hidden stalls in execution. The schedule is constrained by tail latency requirements; a theoretically throughput-optimal schedule can defer execution indefinitely.
For an analytical database engine, an "atomic operation" is typically a query operation on a database page. A modern server can retire 100M ops/sec. While I am oversimplifying a bit, a 1 millisecond GC pause can blindly wreck the schedule for 100,000 operations in an unpredictable way. In these architectures we try to eliminate all context switches for the same reason which are 100x cheaper.
Practically, 1µs stall is a good heuristic for a noise floor. The schedulers have pretty wide concurrency on big systems, so the implied 100 operations are unlikely to have a dependency. Many stalls that are difficult to precisely control like cache line fills fit in here too.
If there was a GC that had a worst-case stall of 1µs then you could probably use it for these cases. Unfortunately, "low-latency" GCs tend to be more like 1000x that. I don't think there is any way of closing that gap short of putting a GC in hardware.