INT8 quantization doesn’t reduce disk/RSS; db size grows vs. FP32

[tae898]

Summary
INT8 quantization speeds ingest but doesn’t shrink footprint. For MSMARCO-1M:

• Ingest time: ~1h07 (INT8) vs. ~1h51–2h00 (FP32).
• Recall: ~0.9072 vs. ~0.8994 (comparable).
• Footprint: db size ~10.6 GB (INT8) vs. ~9.6 GB (FP32); RSS still ~9.5 GB.

It appears quantized vectors are not reducing storage/memory—likely due to duplicated storage (graph + bucket) or storing float copies alongside int8.

Repro

1. Build vector index with quantization=INT8, storeVectorsInGraph=false.
2. Ingest MSMARCO-1M, run one search to trigger graph build.
3. Measure db size and RSS during/after build.

Expected
INT8 should materially reduce disk and memory vs. FP32.

Actual
Disk increases and RSS remains high.

Possible causes

• Vectors still serialized as float somewhere (graph or doc fetch).
• Duplication persists despite quantization (graph + bucket both storing vectors).

Help
I can investigate storage paths and submit a fix once root cause is confirmed.

[lvca]

Is this fixed after the latest fix?

[tae898]

yeah this problem is still there. I'll look into it.