NoKV

Appearance

RustFS Backend ​

NoKV can store file bodies in RustFS through the S3-compatible object backend. RustFS supports single-node single-disk, single-node multi-disk, and multi-node multi-disk deployment modes. The single-node single-disk mode is useful for local testing and small non-critical deployments; it is not a high-availability production layout.

Start RustFS ​

Install the RustFS binary with Homebrew:

bash
brew tap rustfs/homebrew-tap
brew install rustfs
rustfs --version

Start a single-node local RustFS process:

bash
mkdir -p /tmp/rustfs-data
RUSTFS_ACCESS_KEY=rustfsadmin \
RUSTFS_SECRET_KEY=rustfsadmin \
rustfs server \
  --address 127.0.0.1:9000 \
  --console-enable \
  --console-address 127.0.0.1:9001 \
  --buffer-profile AiTraining \
  /tmp/rustfs-data

Create the default NoKV bucket with an S3-compatible client:

bash
AWS_ACCESS_KEY_ID=rustfsadmin \
AWS_SECRET_ACCESS_KEY=rustfsadmin \
aws --endpoint-url http://127.0.0.1:9000 \
  s3api create-bucket --bucket nokv

Use RustFS With NoKV ​

bash
cargo run --release -p nokvfs-cli --bin nokv-fs -- \
  init

Those are the CLI defaults for the RustFS backend: bucket nokv, endpoint http://127.0.0.1:9000, access key rustfsadmin, and secret key rustfsadmin.

The same object flags work for artifact publish, cat, and FUSE mount:

bash
cargo run --release -p nokvfs-cli --bin nokv-fs -- \
  --object-backend rustfs \
  --s3-bucket nokv \
  --s3-endpoint http://127.0.0.1:9000 \
  --s3-access-key-id rustfsadmin \
  --s3-secret-access-key rustfsadmin \
  put-artifact /runs/1/checkpoint.bin ./checkpoint.bin

Benchmark Against RustFS ​

bash
cargo run --release -p nokvfs-bench --bin nokv-fs-bench -- \
  --profile smoke \
  --workload checkpoint-publish \
  --object-backend rustfs \
  --object-concurrency 4 \
  --checkpoint-bytes 1048576 \
  --s3-bucket nokv \
  --s3-endpoint http://127.0.0.1:9000 \
  --s3-access-key-id rustfsadmin \
  --s3-secret-access-key rustfsadmin

mdtest-easy and mdtest-hard are metadata-only and do not exercise object storage. checkpoint-publish and training-read are the useful object-backed workloads for RustFS. The benchmark harness uses the deployable single-node metad service boundary by default. Use --block-cache off as a control run when measuring object backend latency instead of NoKV cache reuse.

References: