A Ruby/Rails engineering showcase for external integrations, release automation and recovery.
MESH starts as a creator marketplace: brief → proposal → agreement → versioned delivery → acceptance. Its Publishing lab demonstrates what happens when a partner accepts a release but the response is lost: retain the operation, diagnose uncertainty and reconcile without publishing twice.
Watch the short demo · Screenshot walkthrough · Verified results
Actual application recording with English captions. The application UI is in Russian. The public presentation is static; it does not expose the backend or real partner credentials.
| Engineering problem | Implementation |
|---|---|
| Repeated or concurrent commands | Idempotency keys, PostgreSQL constraints, row locks and immutable agreements |
| Uncertain external effects | Durable operation identity, lookup-only reconciliation, lease fencing and signed callback deduplication |
| Unsafe or stale release inputs | Private digest-bound ZIP artifacts, bounded validation, real ClamAV, policy/configuration fingerprint |
| Reliable delivery | Outbox, Solid Queue, durable dispatcher and metrics that include unfinished work |
| Trusted release and recovery | Signed image provenance, deployment by digest, migration/runtime DB roles, rollback journal and encrypted restore |
Architecture: eight-domain modular Rails monolith. Publishing execution has framework-independent Domain/Application layers with explicit ports and PostgreSQL/HTTP/storage adapters. Code map and tradeoffs.
Stack: Ruby / Rails · PostgreSQL · Solid Queue · Next.js / TypeScript · Node.js partner simulator · Docker · GitHub Actions · Ansible · Terraform · Helm / Kubernetes · Prometheus / Grafana / Alertmanager.
The accepted application release is 3d849af. Later documentation commits retain that release identity.
| Verification | Result | Evidence |
|---|---|---|
| Ruby and browser | 153 Ruby examples in each of development and native Alpine; 13 browser scenarios; no failures, pending, skipped or flaky scenarios | CI |
| Exact-image security | Five runtime digests scanned; zero HIGH/CRITICAL findings, including unfixed; four signed build outputs | Release |
| Ubuntu deployment | Ansible second apply changed=0, mixed load, rejected release rollback and eight SIGKILL recovery points | Deployment |
| Kubernetes | Terraform drift repair, real NetworkPolicy controls, readiness/liveness and rollout/rollback | Cluster exercise |
| Clean demo preparation | Fresh Ubuntu checkout; actual npm run demo; all 17 stages and 4 browser checks passed | Cold preparation |
| Separate-VM recovery | Primary/queue/private artifacts restored; API/workers resumed; no repeated external publication | Recovery |
The clean-run preparation and gallery distinguish fresh checkout verification from an existing local preview. Reports retain source revisions, workload/fault scope, original artifact digests and per-file hashes. Verify the saved catalogue with npm run test:partner-support:evidence. A passing scan is a dated observation, not a permanent security guarantee.
Requirements: Node.js 22/npm, Docker with Linux containers, registry access, Microsoft Edge on Windows (or the matching Playwright Chromium runtime on Linux) and enough memory for the application plus ClamAV. Prepare before the interview; the command includes builds and can take several minutes.
npm run demoOpen Publishing, marketplace and Grafana. Synthetic operator: ops@mesh.local / MeshDemo2026! — lab credentials only.
npm run demo:publishing # real partner, lost response, reconciliation and rollback
npm run demo:partner-support # disposable peer: credential rotation, fencing and uncertainty
npm run demo:incident # worker outage, delivered alert and once-only recoverySetup and environment guide · 15-minute technical walkthrough · Detailed marketplace guide (RU)
This is a reproducible engineering lab, not a claim of operating a commercial gaming platform. Uploaded code is never executed; payments and partners are simulators. Hosted environments are ephemeral, Kubernetes is single-node, and restore uses a quiescent snapshot with a small fixture. Multi-node HA, online WAL/PITR, large-restore throughput, permanent backup retention and a long production SLO window are not demonstrated.
