Infrastructure wants to consolidate our six-machine in-...

Both in-memory indexes and DiskANN are approximate, so the recall claim must be tested against exact search on held-out queries. The real risk is the write throughput: Vamana assumes a static point set. If the system requires high write throughput, you must use a streaming variant like FreshDiskANN. Finally, ensure the single DiskANN box has sufficient IOPS to handle the peak QPS, or the consolidation will fail regardless of recall.