PADION
dnsd · Internal DNS Control Plane
health-aware DNS · stock CoreDNS · no-DB · multi-node replication
Health-checks per zone publish only live backends as A records, and static records · per-domain forwards · SOA/cache are managed via Web UI/REST. Untouched stock CoreDNS reads the output and responds — no code linkage, so version independence. Runs without a DB, from a single config file.
DNS Control
Internal stacks (like a lakehouse) run many stateless services — gateways, coordinators, workers — scattered across nodes. To stop traffic to dead instances, names must point only to "live places," but static DNS or manual management cannot keep up with health changes. Putting a heavy LB in front of every service is overkill.
PADION dnsd health-checks backends and publishes only live ones as DNS A records, managing static records · forward · SOA in one console. Actual responses come from untouched stock CoreDNS reading zone files, so CoreDNS upgrades are decoupled — and a single config file without a DB is the source of truth.
Per-service TCP/HTTP/HTTPS health-checks (path · rise/fall · timeout). Publish only live backends in the A set. Choose fail-open or fail-closed behavior when all are down. Debounce coalesces flapping.
dnsd generates CoreDNS configuration; CoreDNS reads it automatically. No code linkage → CoreDNS version independence. Operable in a single container as well.
Per-source domain: different upstream · policy (random/round_robin/sequential) · health_check · max_fails. Domain-specific cache (per-rule). Loop detection on every forward block (prevents resolv.conf cycles).
Single writer · multi reader. Manual promote on leader failure · split-brain prevention · peer status · sync history.
Every change goes to append-only audit logs (rotation + retention + time-range query). Standard metrics exposed · CoreDNS apply verification · SBOM provided.
Web UI login (session + forced default password change) + bearer token (constant-time comparison) for REST / Replication. Optional HTTPS (TLS). A single config file is the source of truth (no DB) · changes only persisted after validation (all-or-nothing) · monotonic serial guarantee.
Define a service once — health changes flow into the CoreDNS configuration automatically, so only live backends respond.
01 · Define
Register zone · service · backend · forward via Web UI/REST
02 · Probe
Periodic TCP/HTTP/HTTPS probes → determine live backends
03 · Render
Only live ones written to CoreDNS configuration safely (atomic update)
04 · Serve
Reload and respond · changes replicate to followers
CoreDNS · File + Reload
file + reload)forward + cache + loopREST API + Web UI
Manage zones · services · records · forward · SOA·cache · replication via Web UI (login session) and REST API. Automation uses a Bearer token. All mutations require authentication.
Deploy · Host / Docker / k8s
ID/password session auth (PBKDF2 hash). The initial password must be changed on first login.
Every mutation requires a session or Bearer token. Tokens are compared in constant time (timing-attack prevention).
Login / logout / failures logged with user ID + client IP.
All changes append-only · periodic rotation · retention months · time-range query.
Serve admin / UI over HTTPS with a public certificate. Tokens and sessions are never returned in responses.
Safe single-writer replication isolates old leaders.
Disk persistence only after validation (all-or-nothing) · persist-before-broadcast · monotonic serial.
Forced catch-all in forward + CoreDNS loop plugin prevents infinite resolution loops.
SBOM provided — dependency modules and licenses transparent. Minimal external deps.
Stateless LB
Group stateless endpoints (MinIO gateways, Trino coordinators) under dns-rr, publishing only live instances as multiple A records. Dead nodes are auto-excluded.
Internal Resolver
internal.corp. such as in-house domains use a dedicated upstream / policy, while the rest forward to a public resolver. Per-domain cache TTL is applied separately.
High Availability
Configuration edited on the leader replicates to followers. Even on leader failure, follower CoreDNS keeps responding, and a manual promote can safely take over.
PADION integration
Provides consistent names and health-aware routing for internal services of PADION lakehouse (Trino · MinIO · metastore, etc.). Owns "name → live place" at the infrastructure layer.
| DNS engine | stock CoreDNS (file + reload). Decoupled from code → version independence. In single-container mode dnsd starts and supervises CoreDNS as a child. |
|---|---|
| Zone model | Multi-zone (origin). Zone = dynamic services (health) + static records. k8s custom zones map 1:1. |
| Record types | A · AAAA · CNAME · TXT · MX · NS. Multiple values per (name, type, value) identifier. |
| Health-check | TCP / HTTP / HTTPS (health path) · interval · timeout · rise / fall · debounce. all_dead: fail_open / fail_closed. |
| Forward | Per-source domain upstream / policy / health_check / max_fails / per-rule cache. Mandatory catch-all '.' with loop detection. |
| SOA / Cache | Global defaults + per-zone override (field inheritance). Auto-monotonic serial. Global / per-rule cache (TTL · success · denial). |
| Replication | Single writer · multi reader · safe sync · manual promote · peer status · sync history. |
| Authentication | Web UI login session (PBKDF2, forced default-password change) + Bearer token for REST/Replication (constant-time comparison, configured via file). Login/logout logged with IP and ID. |
| TLS | HTTPS for admin/UI (certificate + key). Enabled when both cert and key are configured. |
| Audit | Append-only · periodic rotation (audit-YYYY-MM.log) · retention months · time-range (from/to) query. |
| Observability | Standard metrics · CoreDNS apply verification · SBOM provided. |
| Architecture | Go single binary · No-DB (config file as source of truth) · minimal external deps · atomic updates. |
| Deployment | Host + Docker (single-container dnsd+CoreDNS) today · k8s ConfigMap sink planned. Single config for operational paths. |
| PADION integration | data lakehouse Internal service discovery · datakeeper · gatekeeper |
For auth, authz, ETL, storage, and analytics to work, services must first find each other by name and connect to live places. dnsd owns that discovery layer.
PoC, internal network deployment, and multi-node redundancy.