Currently, from a high level point of view, the auto-merger is:

1. Constructing a <shard, nss> map (that is being optimized under SERVER-73378)
2. for each shard:
3. - For each namespace
4. ---- Issue an auto merger request (in parallel)
5. - Wait for previous merge requests to drain before throttling and issuing requests for next shard

In presence of thousands of collections, this may result in performance issues:

• Sending 100K network requests to the same shard in case of 100K collections with mergeable chunks on the same shard
• The commit of merges (for different collections) will serialize on the CSRS under the chunks lock