AVD pipeline overview

Developer Guide

If you want to use the system to produce configs, start with Quick Start.

The Arista Validated Design (AVD) pipeline transforms Infrahub's network data model into PyAVD-compatible input data, then renders Arista EOS configurations and human-readable documentation from it.

PyAVD version

Version-sensitive

The integration targets pyavd >= 6.3.0, < 6.4.0 (pinned in pyproject.toml).

The following sections are version-sensitive — review them when upgrading PyAVD:

• Hostvars Reference — the PyAVD input schema.
• Role Mapping — AVD device type names (for example, l3leaf, super-spine).
• Transforms — the PyAVD functions the transforms call (validate_inputs, get_avd_facts, get_device_structured_config, get_device_config, get_fabric_documentation).

The two-phase pipeline

Phase 1 — Hostvars

Generator: generate-avd-device-hostvar Target: each DcimDevice in the avd_devices group (one task per device).

For each device the generator:

1. Extracts device attributes — name, role, BGP ASN, node ID, loopback, management IP.
2. Determines the uplink role based on the device's role: spine → super_spine interfaces, leaf and border_leaf → spine interfaces, l2leaf → leaf interfaces, super_spine → no uplinks.
3. Extracts connected endpoints (servers) from interfaces with role = "server", including tagged/untagged VLANs.
4. For leaves, extracts the MLAG peer information, then the virtual router MAC.
5. For leaves and spines, queries EVPN tenants, VRFs, SVIs, and L2 VLANs associated with the fabric (skipped entirely for l2leaf).
6. For Border Leafs, evaluates NetworkLink objects with role=dci in the fabric and emits valid links as profile-free PyAVD l3_edge.p2p_links entries. DCI addressing resolves from NetworkFabric.fabric_ip_pools role dci first, falls back to the legacy NetworkFabric.dci_pool, and then uses deterministic Fabric Supernet fallback when the required DCI prefix-pool role is missing.
7. Builds a complete PyAVD hostvars dict (see Hostvars Reference).
8. Serialises to JSON, computes a SHA256 checksum, and compares against the previous content. If changed (or absent), writes a new AvdHostvarFile as a child of the device's AvdArtifact node.

Phase 2 — structured config

Generator: generate-avd-device-structured-config Target: each NetworkFabric in the fabrics group (one task per fabric).

For the fabric the generator:

1. Walks the fabric hierarchy (fabric → pods → devices, fabric → pods → racks → devices) to collect every device.
2. Verifies each device has a hostvar artifact; fails fast if any is missing (meaning Phase 1 didn't complete for that device).
3. Fetches the hostvars JSON for every device.
4. Calls pyavd.validate_inputs() across all hostvars.
5. Calls pyavd.get_avd_facts() once for the fabric to derive shared facts (routed-uplink allocations, VLAN assignments, etc.).
6. For each device, calls pyavd.get_device_structured_config(hostvars, facts) and gets a dict of structured AVD config.
7. Serialises to JSON, computes a SHA256 checksum, and compares against the previous content. If changed (or absent), writes a new AvdStructuredConfigFile as a child of the device's AvdArtifact.

Transforms — on demand

When an operator opens an AVD artifact in the Infrahub UI, the matching transform runs:

• avd_eos_config — reads structured_config_file, calls pyavd.get_device_config(), returns text/plain.
• avd_device_doc — reads structured_config_file, calls the PyAVD device documentation function, returns text/markdown.
• avd_fabric_doc — reads hostvars and structured configs for all devices in the fabric, calls pyavd.get_fabric_documentation(), returns text/markdown.

See Transforms for the full transform-by-transform reference.

Components at a glance

Generator / Transform	Target	File
generate-avd-device-hostvar	per device	generators/generate_avd_device_hostvar.py
generate-avd-device-structured-config	per fabric	generators/generate_avd_device_structured_config.py
avd_eos_config	per device	transforms/avd_eos_config.py
avd_device_doc	per device	transforms/avd_device_doc.py
avd_fabric_doc	per fabric	transforms/avd_fabric_doc.py

Related pages

• Hostvars Reference — the exact PyAVD input structure built per role.
• AvdArtifact & File Storage — schema, relationships, and how Phase 1 and Phase 2 share data.
• Role Mapping — Infrahub roles → AVD device types.
• Transforms — detailed transform-by-transform breakdown.
• Extending the Pipeline — worked examples for adding roles, transform outputs, or hostvar fields.
• Debugging the Pipeline — object-store inspection, forced regeneration, single-generator re-runs.