Data plane inodixd
A Go binary with no runtime dependencies: S3 API, Reed–Solomon erasure coding, Raft metadata on Pebble, healing, scrubber, lifecycle, replication, encryption, notifications.
S3 object storage with RS erasure coding, per-division tenants, and offline licensing
S3 behavior follows the AWS documentation; when in doubt, we test against AWS.
Modules
- S3 API
- Buckets and objects, versioning, multipart, presigned (SigV4 and V2), Object Lock, tagging, ACLs, bucket policy, CORS, lifecycle, notifications, replication, STS AssumeRole/WebIdentity.
- Erasure coding
- Reed–Solomon per set (n ≤ 16), automatic parity, write quorum, per-object heal, scrubber, per-set heal.
- Metadata
- Multi-group Raft on Pebble: a system group plus per-bucket groups with automatic range splits, snapshots, and replica moves between nodes.
- Multi-tenancy
- Tenants with byte, bucket, user, and requests-per-second quotas; suspension; bucket prefixes; per-tenant console.
- Encryption
- SSE-S3, SSE-KMS (internal KEK or external backend), SSE-C, bulk re-encryption, mTLS between nodes; secrets are never logged.
- Metering
- 5-minute samples per bucket, daily and monthly aggregates, reconciliation; REST for reporting and internal chargeback.
- Replication
- Bucket replication to S3 targets, backlog and lag status, per-target bandwidth limits, bulk retry, dead-letter events.
- Alerting
- Disk offline, quorum, license, capacity, tenant quota → webhook, Slack, Teams, Telegram, email; silences and acknowledgements.
- Licensing
- Offline Ed25519 files with capacity and node limits, 30-day grace then read-only; optional license server.
- Operations
- Helm chart and Kubernetes operator, pool decommission, rolling upgrades, metadata backup/restore, encrypted support bundles, Prometheus metrics, hash-chained audit.
Erasure coding RS(16,12): data survives a full node failure
Every object is split into 12 data shards and 4 parity shards across 4 nodes × 4 disks. When a node goes down, any 12 shards still rebuild the whole object.
One object arrives
Your client sends a PUT exactly as it would to AWS S3. No special SDK.
Split into 12 data shards
The object is cut per 1 MiB block into equally sized data shards.
Plus 4 parity shards
Reed–Solomon computes parity. This setup is RS(16,12): 12 data, 4 parity.
Spread over 4 nodes × 4 disks
Every shard lands on a different disk; no two shards share a disk.
A node dies
Four shards vanish at once. Any twelve shards are enough to read the whole object.
Automatic heal
When a replacement node or disk comes online, missing shards are recomputed and written back.
The data plane keeps serving when the control plane is down
Storage nodes do not depend on the console. If the control plane goes down, your applications keep reading and writing.
Control plane console
Admin console with RBAC and MFA, tenants, metering, alerting, licensing, and encrypted support bundles; connects to many clusters over Admin gRPC.
Known differences from AWS S3
- Virtual-hosted style needs domain configuration on nodes or an Ingress; otherwise use path-style.
- SelectObjectContent (S3 Select), website hosting, server access logging, and transfer acceleration are not available yet.
- Any region is accepted in SigV4 because one cluster serves one region.
- Presigned POST (form upload) is SigV4 only.
Clean-room, permissive, local
- Built from scratch
- No MinIO or other AGPL code. Dependencies are MIT, BSD, Apache-2.0, and MPL-2.0 only.
- Native tenants and metering
- Released and tested together with the data plane as one version.
- Local support
- Team and contracts in Indonesian or English; operations guide in Indonesian.
Start with a 7-day trial on one node
Our team installs a 7-day trial on 1 node in your data center, with every S3 and console feature. The enterprise license is applied to the same cluster without reinstalling.