Files
computer-playbook/docs/contributing/design/inventory.md
kevinveenbirkenbachandClaude Fable 5.1 113113e553 perf(applications): render the applications SPOT once per play instead of once per task
Ansible forks a fresh worker process for every task, so the in-process
cache behind lookup('applications') was cold in every task. Each task that
touched the merged applications tree (config, domains, nginx, tls, cdn and
their callers) paid a full render: about 16 s on main and 21 to 23 s on this
branch, where ten more application roles carry MCP and AI Jinja. Measured on
the workspace jobs of run 33574056480 against the main run: 331 such tasks
(93 min, 39 % of the job) versus 391 (141 min, 43 %); the same unit is 9 to
43 % of every compose and swarm deploy job on disk. The parent process
rendered once per play, late, while templating a task name, and only the
tasks after that point ran warm.

The constructor now renders once through the new applications_carrier
action and parks the payload with its cache key in the host fact
_INFINITO_APPLICATIONS_RENDERED. Every later worker inherits the fact at
fork time and get_merged_applications serves it while the key (roles dir,
applications, users, DOMAIN_PRIMARY, SYSTEM_EMAIL_DOMAIN) still matches; a
mismatch falls back to the old render. The action masks its own result: the
payload is the whole tree with every credential, and CI runs with
MASK_CREDENTIALS_IN_LOGS false.

The task sits after the networks, design and service_provider merges, after
the add_host placements and after the token store. An independent review
showed that an earlier position bakes stale values into the play: the
web-app-wordpress title rendered to 'Blog by ' because service_provider did
not exist yet. User records inside the payload are the token store's state
at render time; readers of tokens captured later keep using lookup('users')
or lookup('config', app, 'users.…'), which read the users cache fresh.

Expected savings per job: about 130 min on a fork workspace run, 16 min on
an average compose deploy and 28 min on an average swarm deploy. The
constructor's own lookups before the carrier stay cold.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
2026-09-02 23:55:32 +02:00

92 lines
10 KiB
Markdown

# Inventory 🏗️
How an Ansible inventory is assembled from the per-role `meta/services.yml` declarations, what gets resolved at build time, and what stays as an unrendered Jinja template until play runtime.
For general documentation rules such as links, writing style, RFC 2119 keywords, and Sphinx behavior, see [documentation.md](../documentation.md).
For the matrix-variant folder-per-round model that wraps this pipeline, see [variants.md](variants.md). For the on-disk shape of a role's metadata see [layout.md](role/services/layout.md).
## Pipeline overview 🛤️
The build pipeline runs entirely on the operator's host, before any Ansible play starts. Three stages produce the inventory file plus the matching `host_vars/<host>.yml`:
| Stage | Code path | Output |
|---|---|---|
| 1. Variant resolution | [applications.py (cache)](../../../utils/cache/applications.py) `_build_variants`, [inventory/payload.py](../../../cli/administration/deploy/development/inventory/payload.py) `_resolve_variant_payloads` | Per-app payload = `meta/services.yml` deep-merged with the round's `meta/variants.yml` entry. Jinja strings are preserved verbatim. |
| 2. Group materialisation | [cli.administration.inventory.devices](../../../cli/administration/inventory/devices/command.py), [filters.py](../../../cli/administration/inventory/provision/filters.py) `filter_dynamic_inventory` | Inventory file lists one group per included `application_id`, with the host as a member. |
| 3. Host-vars baking | [cli.administration.inventory.provision](../../../cli/administration/inventory/provision/cli.py) `apply_vars_overrides`, [host_vars.py](../../../cli/administration/inventory/provision/host_vars.py) | `host_vars/<host>.yml` carries the variant-resolved `applications.<app>` block plus connection vars, become password, etc. |
After stage 3 the inventory is fully variant-resolved on disk. Ansible loads it as plain YAML, no rendering yet.
## What decides which roles enter the inventory 🎯
The include set passed to stage 2 is built per round by the matrix planner: [deploy/cli.py](../../../cli/administration/deploy/development/deploy/cli.py) and [init.py](../../../cli/administration/deploy/development/init.py) call `plan_dev_inventory_matrix` in [planner.py](../../../cli/administration/deploy/development/inventory/planner.py), which resolves each round's include via `_resolve_round_include` in [legacy_resolver.py](../../../cli/administration/deploy/development/inventory/legacy_resolver.py) using the [CombinedResolver](../../../cli/meta/roles/applications/resolution/combined/resolver.py) with `follow_run_after=False`. The swarm test path computes its include separately via [derive_includes.py](../../../utils/tests/swarm/derive_includes.py), reading the active variant map from `INFINITO_APP_VARIANTS`. Both expand the selection along two transitive paths:
1. **Ansible meta `dependencies:`** from each role's `meta/main.yml`.
2. **Shared-service auto-include** from each role's `meta/services.yml.<entity>.<service>` block, when both `enabled` AND `shared` pass `is_explicit_truth` in [registry.py](../../../utils/roles/applications/services/registry.py): the literal `True` or the dynamic form `"{{ '<role>' in group_names }}"`. A literal `false` (as baked by a variant override) does NOT qualify.
`run_after:` from `meta/services.yml.<entity>.run_after` is an ordering hint only (see [base.md](role/services/base.md)); it MUST NOT pull roles into the include set.
This is the mechanism that lets variants control whether a shared service is pulled in: a variant override pins `services.<service>.{enabled,shared}` to a literal boolean; the round's include resolution reads the variant-merged services map; the provider role's group materialises in (or drops out of) the inventory accordingly.
## What lives in the inventory file vs. host_vars 📋
| Artifact | Contents |
|---|---|
| `inventory.yml` (or equivalent) | Group-per-application map plus host membership. Source of truth for `group_names` at play time. |
| `host_vars/<host>.yml` | Connection vars, the variant-baked `applications.<app>` block, vault-encrypted `ansible_become_password`, plus any `--vars` / `--vars-file` overlays (which is where deployment-wide flags like `DOMAIN_PRIMARY`, `TLS_ENABLED`, and `networks.internet.{ip4,ip6}` enter when set; otherwise group_vars provides the env-driven defaults). |
`host_vars/<host>.yml` is where variant data ends up after the deep-merge. The strings inside are still raw, including any Jinja templates from `meta/services.yml` that no variant override replaced.
## When Jinja gets rendered ⏱️
Rendering happens at **play runtime**, never during the build pipeline above. The render call is centralised in [applications.py (cache)](../../../utils/cache/applications.py) `get_merged_applications`, which:
1. Loads variant 0 as the default application payload.
2. Layers on inventory `variables['applications']` (the host_vars block) as overrides via deep merge.
3. Calls `_render_with_templar` with the play's full variable scope, including `group_names`, `inventory_hostname`, and any other facts.
4. Returns the fully rendered payload.
`get_merged_applications` is invoked transitively by `lookup('applications', ...)` and by `lookup('config', application_id, 'services.<x>.<y>')`. Tasks that read service flags through these lookups always see real Python values, never raw Jinja strings.
The render runs once per play. The constructor stage parks the payload and its cache key in the host fact `_INFINITO_APPLICATIONS_RENDERED`, and every task's worker process (Ansible forks one per task) serves that payload instead of rendering again. See [applications.md](../artefact/files/plugins/lookup/applications.md) for the carrier contract.
A role's task that reads its own service flag therefore picks up the runtime-resolved value, even when the inventory carried a Jinja string for that flag: the conditional evaluates against the play's actual `group_names`, which reflects which other application groups the host joined.
## Interaction between the three inclusion mechanisms 🧬
A role can be pulled into the inventory through any of three paths and they compose, not exclude one another:
| Path | When it fires |
|---|---|
| Operator-supplied `--apps` (or `--include`) | Always. The named roles enter the include set verbatim. |
| `run_after:` walk | When a role in the include set declares `run_after: [<other>]`, `<other>` is added to the include set. |
| Shared-service auto-include | When a role in the include set declares `services.<svc>` with literal `enabled: true and shared: true`, the role behind that service is added. |
The matrix-variant planner (see [variants.md](variants.md)) calls the expansion once per round, so each round's inventory gets the right transitive closure for that round's variant set.
## Worked example: WordPress and Discourse 🧪
The wp-discourse plugin install in [roles/web-app-wordpress/tasks/addons/wp-discourse.yml](../../../roles/web-app-wordpress/tasks/addons/wp-discourse.yml) does `container exec discourse rake api_key:create_master[...]` during the WordPress play. The Discourse container therefore has to exist by the time WordPress runs.
The wiring that makes this work is purely in the `services` declaration, not in `run_after`:
- [roles/web-app-wordpress/meta/services.yml](../../../roles/web-app-wordpress/meta/services.yml) declares `services.discourse.shared: true` plus a fallback `enabled: "{{ 'web-app-discourse' in group_names }}"`.
- [roles/web-app-wordpress/meta/variants.yml](../../../roles/web-app-wordpress/meta/variants.yml) variant 0 (canonical Single-Site) overrides `services.discourse.enabled: true`. Variant 1 (Multisite) overrides it to `false`.
- The variant override is baked as a real boolean into the inventory, so the shared-service auto-include picks it up at build time. Variant 0 therefore pulls `web-app-discourse` into the include set. Variant 1 does not.
- Discourse's role itself has no `run_after`, so deploying just `--apps web-app-discourse` does not drag WordPress along.
The fallback Jinja string in the base `meta/services.yml` is the safety net for the case where neither variant overrides the flag and the operator still adds `web-app-discourse` to the inventory groups manually: the runtime `lookup('config', application_id, 'services.discourse.enabled')` then renders the conditional against the real `group_names` and the wp-discourse plugin install enables itself accordingly.
## Reference files 📌
| File | Purpose |
|---|---|
| [applications.py (cache)](../../../utils/cache/applications.py) | `get_variants`, `get_application_defaults`, `get_merged_applications`. Stage 1 variant loader and the runtime renderer. |
| [service_registry.py](../../../utils/roles/applications/services/registry.py) | `build_service_registry_from_roles_dir`, `resolve_service_dependency_roles_from_config`. Strict-identity auto-include. |
| [in_group_deps.py](../../../utils/roles/applications/in_group_deps.py) | `applications_if_group_and_all_deps`. Recursive walk over Ansible meta `dependencies:` plus shared-service deps. |
| [inventory/](../../../cli/administration/deploy/development/inventory/__init__.py) | Package root re-exporting the public API. Per-submodule split: [`payload.py`](../../../cli/administration/deploy/development/inventory/payload.py) (`_resolve_variant_payloads`, `_bake_overrides`) and [`builder.py`](../../../cli/administration/deploy/development/inventory/builder.py) (`build_dev_inventory`). Stage 1 variant resolution and stage 3 host-vars baking. |
| [cli.administration.inventory.devices](../../../cli/administration/inventory/devices/command.py) | Stage 2 group materialisation: emits one group per invokable `application_id` for the given host. |
| [filters.py](../../../cli/administration/inventory/provision/filters.py) | `filter_dynamic_inventory`. Restricts the materialised groups to the `--include` set. |
| [host_vars.py](../../../cli/administration/inventory/provision/host_vars.py) | Connection vars, vars-file/JSON overlays, become-password handling. |
| [config.py (lookup)](../../../plugins/lookup/config.py) | `lookup('config', application_id, 'services.<x>.<y>')`. Resolves nested config paths and renders Jinja via `_render_with_templar`. |
| [applications.py (lookup)](../../../plugins/lookup/applications.py) | `lookup('applications'[, application_id])`. Returns the rendered application payload. |