Fix(#590): use snake_case DiskANN build params so max_degree/search_list_size are applied

[idevasena]

Fixes #590 + Pull updated README.md from PR #588

Problem

For index_type: DISKANN, the max_degree and search_list_size values set in config.yaml (or via --max-degree / --search-list-size) had no effect on the built index. PR #497 fixed config parsing so the values reach create_index(),
but the index was still built with engine defaults.

Root cause

As identified in issue #590 (comment),
this is a key-name case mismatch, not a milvus.yaml server-config limitation. knowhere's DiskANN config declares snake_case build fields (src/index/diskann/diskann_config.h:87-99) and builds with
build_conf.max_degree.value() / build_conf.search_list_size.value() (src/index/diskann/diskann.cc:481-498). The benchmark's DISKANN branch passed CamelCase keys (MaxDegree, SearchListSize), which knowhere silently drops.

The reporter's build log (configured max_degree=72, search_list_size=137) confirms it: the snake_case fields were applied with defaults (56 / 100) while the CamelCase fields carried the configured values (72 / 137) but were ignored. HNSW and AISAQ were unaffected because their key names already match what the engine reads — the AISAQ branch was already snake_case.

Changes

File	Change
vdb_benchmark/vdbbench/load_vdb.py	DISKANN branch emits max_degree / search_list_size (snake_case), matching the AISAQ branch; help strings updated
vdb_benchmark/vdbbench/mpi_wrapper.py	Same fix in the multi-node _build_index_params() path
vdb_benchmark/vdbbench/benchmark/backends/milvus/backend.py	Emits snake_case; normalizes legacy CamelCase keys from older user configs (params.get("max_degree", params.get("MaxDegree", 64))) so existing configs keep working but now take effect
vdb_benchmark/vdbbench/benchmark/backends/milvus/__init__.py	DISKANN ParamDescriptor names renamed to snake_case so introspection advertises the working keys
vdb_benchmark/vdbbench/benchmark/configs/1m_diskann.yaml	Reference config switched to snake_case keys
vdb_benchmark/vdbbench/benchmark/backends/milvus/README.md	Parameter table updated; note added documenting legacy-key normalization
mlpstorage_py/cli/vectordb_args.py	Help strings refer to snake_case parameter names

Notes on scope:

• run_benchmark.py:_collect_index_params() intentionally still accepts both
  spellings; the backend normalizes CamelCase → snake_case, so this PR is
  backward compatible with existing configs.
• collection_mgr.py already reads both spellings for display; unchanged.

Impact on submissions

Any prior DISKANN run on the affected code built its index with knowhere defaults regardless of configured values, so DISKANN "tuning" results were not actually varying build parameters. After this fix, configured values take effect — DISKANN build times and recall/QPS characteristics may legitimately change versus earlier runs with the same config. This may warrant a note in the v3.0 changelog / Rules discussion.

Validation

1. Start Milvus and load with clearly non-default values:

   python3 vdb_benchmark/vdbbench/load_vdb.py \
     --index-type DISKANN --max-degree 72 --search-list-size 137 \
     --collection-name diskann_590_check --force ...

2. Confirm the applied build params in the Milvus/knowhere build log
   (grep -E "max_degree|search_list_size" <milvus log>): snake_case fields
   now show 72 / 137 instead of defaults 56 / 100, and no ignored
   CamelCase entries carry the configured values.

3. Confirm via index description:

   from pymilvus import Collection, connections
   connections.connect(host=..., port=...)
   print(Collection("diskann_590_check").index().params)
   # {'index_type': 'DISKANN', 'metric_type': 'COSINE',
   #  'params': {'max_degree': 72, 'search_list_size': 137}}

4. Backward-compat check: run the modular framework with an old-style config containing MaxDegree / SearchListSize and confirm the backend normalizes and applies them (same log signature as step 2).

5. Sanity: build-time sensitivity — with a fixed dataset, max_degree=16 vs 72 should now produce measurably different index build times and index file sizes, which was not the case before this fix.

6. Regression: HNSW and AISAQ load paths unchanged (M/efConstruction and inline_pq/max_degree/search_list_size still applied as before).

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅