Yes, I think the second. You have a pool of 100 http clients and a queue of one million requests and a queue to accept the responses as the clients complete, and a little machine that waits for capacity in the client queue to send the next request until there are no more requests. If the response is important to this process, your machine is also pulling from the response queue as available and computing whatever it needs from that, for example to decide whether to abort the rest of the requests. Any other use of the responses can be handled outside this loop.

The other way would work fine, but I think it's actually slightly more complicated and slower because you now have a queue of 10000 batches of 100 requests each and the machine has to watch for all one hundred clients to complete before sending off the next batch. Otherwise, it's the same situation.