Skip to content

Physical AI Edge/Fleet Connector Production Readiness Plan

Date: 2026-06-23 Status: Experimental runtime path

This plan lists the gates required before the Physical AI edge/fleet Action-Outcome connector can be promoted from research/experimental runtime to a production ZeptoDB feature.

In production, the connector is the bridge between edge-local safety memory and fleet-global audit memory:

  • edge systems keep immediate safety decisions local and latency-bounded,
  • edge outbox events are transferred through bounded, retryable passes,
  • fleet systems converge later for audit, replay, and policy improvement,
  • every decision remains explainable as time-series action/outcome evidence.
GateStatusCompletion Criteria
Runtime state machineDone, devlog 202Bounded passes, duplicate/late handling, retry accounting, checkpointing, metrics.
Live SQL/HTTP replay adapter proofDone, devlog 203Two ZeptoDB HTTP nodes replay outage, dropped, duplicate, late, restart, ACK convergence, and audit JOIN checks.
Server lifecycle control planeDone, devlog 204Admin-gated configure/start/stop/delete/status and metrics.
Server-managed worker foundationDone, devlog 205Runtime owns bounded runOnce() and background worker loop with injected loader/sink hooks, status, and metrics.
Built-in SQL/HTTP adapterDone, devlog 212Server can load the configured edge outbox table and apply fleet inbox/final/ACK/telemetry rows without zepto_edge_fleet_replay or embedding-only hooks.
Persisted connector configurationDone, devlog 213Server-local versioned config persistence reloads SQL/HTTP hooks and enabled state; ACK/cursor state persists separately through checkpoint_path.
Idempotent sink contractDone, devlog 213HTTP/design docs define feed_event_id idempotency, ACK-boundary replay, duplicate final-row behavior, and ACK-ledger source-of-truth rules.
Backpressure and rate limitsDone, devlog 213Worker has explicit max rows, max bytes, retry/backoff, and per-pass failure-budget behavior.
Live restart/fault testsDone, devlogs 203, 213, and 215Standalone two-node replay covers live outage/fault cases; server tests cover HTTP restart config reload and checkpoint replay skip over live tables; devlog 215 adds server-runtime restart soak, node-replacement, and two-live-HTTP-node convergence through the built-in SQL/HTTP adapter.
Security and auditDone, devlog 213Admin endpoints are RBAC/rate-limit gated and mutating edge/fleet config/start/stop/clear outcomes emit audit-buffer events without adapter secrets.
Cross-architecture verificationDone, devlogs 212-213x86_64 and aarch64 focused/broad builds/tests pass with matching connector semantics.
Promotion documentationControlled pilot approved, devlog 222API/design/operations docs describe the supported controlled-pilot scope, limits, idempotency, monitoring, fault validation, rollback/disable procedure, and promotion criteria. README-level GA positioning remains intentionally conservative.
  1. Built-in SQL/HTTP adapter (done in devlog 212):

    • extract reusable SQL adapter pieces from zepto_edge_fleet_replay,
    • bind them to EdgeFleetConnectorRuntimeHooks,
    • keep table names and limits explicit,
    • reject unsupported schemas with clear errors.
  2. Persistence (done in devlog 213):

    • persist runtime config separately from ACK checkpoint state,
    • load config before enabling the worker,
    • document rollback and disable behavior.
  3. Operational hardening (done in devlog 213 for controlled pilots):

    • add bounded retry pacing,
    • expose last error and per-pass failure-budget status,
    • add audit events for admin lifecycle changes.
  4. Live production-style validation:

    • run two-node edge/fleet replay through the server runtime,
    • inject dropped, duplicated, late, outage, restart, and ACK-boundary cases,
    • verify edge-local safety decisions remain independent of fleet outage,
    • verify fleet audit eventually converges.
  5. Promotion review:

    • run focused cross-architecture verification,
    • update API/design/README docs,
    • mark the feature as promoted only after governance gates pass.

The explicit rollout decision is Controlled Pilot. The server path is approved only for admin-gated, opt-in pilot environments using the documented SQL/HTTP edge/fleet contract: durable config reload, checkpoint cursor reload, ACK-ledger paging for bounded SQL loads, idempotent sink behavior, retry/failure-budget controls, restart regressions, node-replacement evidence, two-live-HTTP-node convergence, and admin audit coverage are in place.

This is not a broad GA/operator release. Limited Operator Feature promotion requires completed pilot soak/fault evidence, operator dashboards/alerts, public support-boundary wording, and a new production gate that explicitly changes the status.