- Add secmarks (id-keyed) and vars (key-keyed) resources: migration 0009, model,
store CRUD, REST handlers.
- Compiler rendering for the global-compiled tail: blrules, conntrack, secmarks
render on enforcing devices; vars render on every device. This closes the
rendering gap left by batch 1 (blrules/conntrack were stored but not rendered).
Storage + CRUD (migration 0007, model, id-keyed store, REST handlers) + compiler
rendering, each owned by a device. proxy_arp/proxy_ndp share the ProxyEntry
shape via table-parameterized store helpers.
Storage + CRUD (migration 0006, model, id-keyed store, REST handlers) plus
compiler rendering: each section is owned by a device and projected into that
device's rendered config (hosts/providers/routes/routing_rules).
The compiler now projects the global NAT intents through each device's bindings
into the rendered config:
- snat/masquerade: renders on a device that binds the egress zone (and, when the
source is a zone, that zone too), resolving the egress interface — this
auto-scopes masquerade to edge devices. A literal-CIDR source needs only the
egress binding.
- netmap: renders on the device its anchor (device:zone|interface) names,
resolving a zone anchor to its bound interface.
- 1:1 nat: renders on the device it is bound to.
Adds RenderedSNAT/RenderedNetmap/RenderedNAT to the rendered config, fetches the
tiers in Compile, and unit-tests binding-scoping (edge vs interior/other device).
Devices carry a reachable_prefixes set (migration 0004) reported by their agent
from the kernel FIB, via a new agent-authed POST /devices/{name}/routes endpoint.
This is scoping/observability data, so it does not bump the config generation and
is exposed on the device resource.
It deliberately does NOT limit which rules a device enforces: the compiler still
emits every applicable rule on every enforcing device. The interface-agnostic,
address-matched rule form is correct under ECMP precisely because it does not
depend on which device is on the path, and most routers hold a default route
anyway — so reachability could not meaningfully narrow the rule set. The reported
FIB is kept for fleet visibility and future zone-origin validation.
Expand asn address groups to concrete prefixes centrally (one iplocate key,
consistent fleet-wide) and refresh them on a per-group TTL (default 24h). A
background Refresher scans for due groups, unions each group's ASNs to a deduped
prefix set, and writes them to a new resolved/resolved_at column (migration
0002). Fail-safe: a lookup error or empty expansion keeps the last-good set,
never emptying it. The compiler folds resolved prefixes into the rendered set
members; membership churn bumps the generation but never rewrites rules. The
iplocate client is endpoint-configurable and response-tolerant, documented as
needing endpoint/key confirmation. Unit tests cover TTL parsing, due-checks, and
union/dedup/error propagation with a fake expander.
Project the fleet-global model through a device's bindings into a rendered,
interface-agnostic config: rules compile to saddr/daddr forward matches with no
iif/oif so they are correct under FRR/ECMP. Firewalls always enforce; routers
enforce only when their fabric opts into defense-in-depth. Referenced address
groups are emitted as named sets carrying their source (static CIDRs, dns FQDNs,
or asn numbers) so membership churns out-of-band without a rule reload. Wire
GET /devices/{name}/config to compile and serve YAML, generation-stamped. Add
portgroups/policies/settings store methods and portgroup CRUD. Pure Render is
unit-tested for enforcement gating, ASN set emission, and resolver precedence.
benvin