[jsd1982]

The obvious solution to me is to implement streaming/buffered processing of the content while downloading it instead of downloading the entire content into memory to be processed in a single contiguous byte[].

Buffered IO would have the CPU processing acting as a natural backpressure mechanism to prevent downloading too much content and also prevent unbounded memory allocation. Each CPU worker only needs to allocate a single small buffer for what it processes and it can refill that same buffer with the next IO request. Your memory usage becomes entirely predictable and will only scale with how many concurrent threads you can actually execute at once.

Also, no matter how you artificially rate limit the virtual thread scheduling (e.g. via semaphore), if you still insist on downloading the entire content into memory before starting processing then obviously you cannot process any single piece of content larger than what can fit into available memory at any given time.