Comparing Vector Databases on Kubernetes

Sergey Pronin Founder, Solanica Inc. · Maintainer, OpenEverest

Vectors: meaning as numbers

“puppy”

→ [0.12, −0.48, 0.91, …]

search by meaning,
not by keywords

Kubernetes is the default
for compute

70%

of organizations run stateful
workloads on Kubernetes

Every database is a
vector database now

PostgreSQL

MySQL

MariaDB

MongoDB

TiDB

Valkey

Redis

Cassandra

SQLite

ClickHouse

Oracle

SQL Server

Two ways to run vectors
on Kubernetes

General-purpose

vectors are a feature

pgvector
MariaDB
MongoDB
TiDB
Valkey

Dedicated

vectors are the product

Qdrant
Milvus
Weaviate
Chroma

HNSW: a map with highways

~1,000 of 10M vectors compared

10M vectors = 61 GB

before the HNSW graph

float32

61 GB halfvec

31 GB int8

15 GB 1-bit

1.9 GB

General-purpose

halfvec, bit

Dedicated

scalar, product, binary

To be fair: Tiger Data benchmarked 1-bit on both.

Scale up or scale out

General-purpose

1 × 512 GB

Dedicated

4 × 128 GB

Yes, but: Citus and TiDB shard too — as a whole SQL cluster.

Nearest neighbors + WHERE

General-purpose

10 → 0

filter after the search

Dedicated

10 → 10

filter during the search

Vector search eats CPU

General-purpose

OLTP vectors

Dedicated

OLTP

vectors

So… which one?

General-purpose

if vectors are a feature

under ~50M vectors
JOINs and transactions matter
one less database to run

Dedicated

if vectors are the product

100M+ vectors, scale-out
heavy filters, tight p99
isolated compute

What does production take?

Survive a pod loss

Scale out

Upgrade without downtime

Back up and restore

We tested it: 3–4 replicas · 100k vectors · 10 queries/s · Linode Kubernetes

Operator or Helm chart?

Milvus

Operator

maintained by Zilliz

Qdrant

Helm chart

official operator is Cloud-only

Weaviate

Helm chart

operator is WIP, no upgrade path

Chroma

Helm chart

single node only

Not in Qdrant's Helm chart

Production ready out of the box

Cloud only

Zero-downtime upgrades

Cloud only

Automatic shard rebalancing

Cloud only

Re-sharding

Cloud only

Full cluster backup and DR

Cloud only

No data loss on downscaling

Cloud only

Source: qdrant-helm README, chart 1.19.2

We tried to break all four

Milvus

Qdrant

Weaviate

Chroma

Kill a pod

●

○

Scale out

●

◐

○

Upgrade under load

●

○

Backup and restore

◐

●

○

● automatic ◐ manual steps ○ downtime or not possible

Add a node. Who moves the data?

Milvus

rebalanced itself in 40 s

Qdrant

4empty

move shards by hand

Weaviate

4empty

flag, restart, then by hand

Chroma: one pod, one outage

51 s

of failed queries on every
restart or upgrade

single node only
no auth since 1.0
no built-in backup

Milvus on

Built on the Milvus operator

every component in one view
pods and logs, no kubectl
scheduled backups to S3

Thank you

Try OpenEverest on your cluster

openeverest.io

Sergey Pronin · solanica.io

Comparing Vector Databases on Kubernetes

Vectors: meaning as numbers

Kubernetes is the defaultfor compute

Every database is avector database now

Two ways to run vectorson Kubernetes

HNSW: a map with highways

10M vectors = 61 GB

Scale up or scale out

Nearest neighbors + WHERE

Vector search eats CPU

So… which one?

What does production take?

Operator or Helm chart?

Not in Qdrant's Helm chart

We tried to break all four

Add a node. Who moves the data?

Chroma: one pod, one outage

Milvus on

Thank you

Kubernetes is the default
for compute

Every database is a
vector database now

Two ways to run vectors
on Kubernetes