Infrastructure
The infrastructure domain is the platform substrate that runs, isolates, and
routes Zephr across installations. It is intentionally not product semantics:
memory meaning lives in memory, ranking in retrieval, metering in
economics, telemetry in observability. What infrastructure guarantees is
that when a pod or a routed event touches a tenant, it is tenant-scoped,
authorized, and recoverable.
Four capabilities own this surface:
| Capability | Owns | This guide |
|---|---|---|
compute-plane | Ephemeral isolated pods (profile → microVM/sandbox), lifecycle, egress/budget/miner controls, pod surfaces | compute-and-managed-environments |
managed-environments | Desired-state Environment + immutable sealed EnvironmentRevision sequence; switch/rollback/sync | compute-and-managed-environments |
mcp-fabric | Hosted user-MCP-server catalog + observed deployment rows | mcp-fabric-and-event-bus |
event-bus | The single routing-bus publish surface + multi-channel dispatcher | mcp-fabric-and-event-bus |
Status honesty
This domain is evolving — parts ship, parts are partial, and some are
designed-not-built. Each companion page carries an inline status caveated per
sub-capability rather than a single page-level claim. The ground rule across all
four capabilities is status-closed: a capability is shipped, evolving
(partial), or planned/dark/unsupported — never asserted from a spec
alone.
Tenancy is the constant
Every table in this domain is tenant-scoped on installation_id and gated by
tenant_installation_matches() (RLS ENABLE + FORCE) keyed off the session
GUC app.installation_id — see the platform/tenancy spec. A connection
without the GUC sees zero rows; a cross-tenant reference is rejected by a
composite foreign key before any row is written. This is why compute_pod,
mcp_fabric_catalog/mcp_fabric_deployment, managed_environment/..._revision,
and the routing tables (routing_rules, channel_deliveries, channels) all
share the same isolation model rather than each inventing its own.
Failure model
Infrastructure here is treated as hostile to its own operator: pods run as a hostile guest, bus publishes are fail-soft so a transient routing error never rolls back an already-persisted webhook, and revocation is a durable sweep across pods, projected clients, and in-memory server processes — not a local delete. The two companion pages spell out the recovery path for each surface.
Companion pages
- Compute and managed environments — pod lifecycle, budget, egress, secrets, residency, isolation, and the desired-state environment model.
- MCP fabric and event bus — the hosted-MCP catalog and the routing event bus with its publish envelope, idempotency, and channel adapters.