Milvus Problems & Issues

Loading puts index files and raw field data of a collection into memory, so collections bigger than query node RAM fail to load or get the node OOM-killed. GitHub issues describe a query node killed for excessive memory and loaded collections showing 0 afterwards.

Official bottleneck guidance points to search parameters (ef, nprobe), filter selectivity, unindexed segments and queueing on query nodes. Users report sudden latency spikes even with query nodes about half used.

A third-party guide reports free-tier clusters auto-pause and the first query after a pause can time out, needing a retry after about 10 seconds. Not confirmed in official Zilliz pages we read.

Milvus does not make a collection searchable on create. You must create a vector index and call load_collection first, which surprises users coming from Chroma or Qdrant. The error is among the most commonly reported in setup guides.

The schema dim is fixed. Sending embeddings of another length (for example 768 into a 1536 collection) fails with a dimension mismatch, and the only fix is a new collection and re-ingest. Downstream tools such as LightRAG needed a patch for it.

Inserts sit in growing segments and results can be empty until flush or a stronger consistency level. Third-party guides list this as a top confusion for new users.

Standalone has a fixed data node pool and many parallel writers can exhaust it. Fix is fewer writers, batches of 10k+ rows, or cluster mode.

Official docs ask for 32 GB RAM minimum (128 GB recommended) and 8+ cores for a cluster, plus etcd, object storage and a WAL. Small teams often find this more than they want to operate.

Unlike some rivals, Milvus does not build the vector index for you in the quick path; create_index on the vector field is required before load. New users hit 'index not exist' and lose time.

Milvus relies on etcd and object storage. Docs warn slow etcd disks cause frequent leader elections; users report standalone crash loops on 'mvcc: database space exceeded', lost etcd keepalive, and 'failed to open WAL' on full disks.

Clients fail with 'Fail connecting to server on localhost:19530. Timeout' when the container is not up yet, the port is not published, or a firewall blocks gRPC 19530. Standalone can take a while to become healthy.

Dedicated is listed from $126/month (Standard) and $197/month (Enterprise), and cluster rates run from about $5 to $63 per million 768-dim vectors per month depending on tier. Serverless is billed by vCUs, which are hard to predict.

Several bug reports about load failures after power loss are open and tagged as needing more information, leaving self-hosters to debug alone. Paid support exists only through Zilliz plans.