Open ThreadDao opened 6 days ago
/assign @weiliu1031 /unassign
/assign @cqy123456 /unassign @weiliu1031
/assign @sunby
/assign @cqy123456
this seems to be an growing mmap issue
Growing insert is not locked. During the insert process, the vector chunks will be cleared (chunk number counter = 0)after the growing index is built. And the growing segment uses indexingrecord.SyncDataWithIndex to get whether the growing index has been successfully built. From the log, it can be seen that the chunks are cleared and the vector BF search at the same time, and there is a consistency problem in the access of indexingrecord.SyncDataWithIndex. SyncDataWithIndex = fasle ->jump to the BF search logic->SyncDataWithIndex = true -> try_remove_chunks -> BF search.
Is there an existing issue for this?
Environment
Current Behavior
milvus server
deploy a cluster with image 2.4-20241013-44564f04-amd64
test steps
results
Only one search requests failed: with
query failed: Assert "index < this->counter_" => index out of range, index=0, counter_=0 at /workspace/source/internal/core/src/mmap/ChunkVector.h:12
Expected Behavior
No response
Steps To Reproduce
Milvus Log
pods:
Anything else?
No response