vectordb-bench 1.0.21 fails optimize/compaction against Milvus 2.6.15

[FantasqueX]

Summary

vectordb-bench==1.0.21 appears incompatible with milvusdb/milvus:v2.6.15 during the Milvus HNSW benchmark build/load stage.

The benchmark successfully creates the collection and inserts 1M rows, but fails during the optimize/compaction step with:

Value out of range: 9223372036854775808

The final VectorDBBench result is marked failed (label: "x") with zero load timings.

Downgrading to vectordb-bench==1.0.20 avoids the issue in the same environment.

Environment

• Milvus image: milvusdb/milvus:v2.6.15
• vectordb-bench: 1.0.21
• pymilvus: 3.0.0
• Benchmark command:

vectordbbench milvushnsw \
  --uri http://localhost:19530 \
  --case-type Performance768D1M \
  --k 100 \
  --num-concurrency 128,160,192,224,256,320,384 \
  --concurrency-duration 30 \
  --concurrency-timeout 3600 \
  --num-shards 4 \
  --replica-number 1 \
  --m 16 \
  --ef-construction 128 \
  --ef-search 128 \
  --skip-search-serial \
  --skip-search-concurrent

Actual Behavior

The build/load phase fails after insertion during Milvus optimize/compaction:

Milvus optimizing before search
Milvus all persistent segments are sorted.
Milvus compact or list segments error: <MilvusException: (code=1, message=Unexpected error, message=<Value out of range: 9223372036854775808>)>
Milvus optimize error: 'ErrorCode' object is not callable
VectorDB optimize error: 'ErrorCode' object is not callable
Failed to run performance case, reason = 'ErrorCode' object is not callable

Result JSON:

{
  "metrics": {
    "insert_duration": 0.0,
    "optimize_duration": 0.0,
    "load_duration": 0.0
  },
  "label": "x"
}

Expected Behavior

The build/load stage should complete successfully and produce non-zero insert/optimize/load timings, as it does with vectordb-bench==1.0.20.

Suspected Cause

In vectordb-bench==1.0.21, the Milvus backend appears to call:

self.client.compact(self.collection_name, target_size=(2**63 - 1))

With pymilvus==3.0.0, MilvusClient.compact() treats integer target_size using the default unit "mb" and converts it via parse_target_size().

A quick repro:

from pymilvus.milvus_client.optimize_task import parse_target_size

parse_target_size("9223372036854775807mb")
# returns 9223372036854775808

That value is larger than signed int64 max and is rejected by Milvus:

Value out of range: 9223372036854775808

There also seems to be a secondary error handling issue:

e.code().name

raises:

'ErrorCode' object is not callable

which masks the original Milvus exception.

This issue is assisted with codex.

[XuanYang-cn]

@FantasqueX Thanks for the but report. This would be a bug of latest pymilvus upgrade. I'll fix it soon. In the mean time, you could avoid this by downgrade pymilvus to 2.6.12.

[XuanYang-cn]

@FantasqueX Should be fixed in the latest code.

[FantasqueX]

Confirm. v1.0.22 works with milvus v2.6.16 well.