coulomb/state-hub
4
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix: bound repository collection identity projection

Browse source 
Assistant: codex
Assistant-Model: gpt-6-astra
Assistant-Session: 01a06ed7-828d-7ca0-a8d4-0c3e5a0c4102
This commit is contained in:
tegwick 2026-09-05 02:03:52 +02:00
parent 4cef58cd98
commit 3578dc2f32
10 changed files with 3331 additions and 9 deletions
Download patch file Download diff file Expand all files Collapse all files

47
docs/evidence/repository-collection-projection-20260905.md Normal file
View file

@ -0,0 +1,47 @@
# Repository collection query bound — 2026-09-05
STATE-WP-0087 review of the pending patch found a second loading path to bound:
`ManagedRepo.domain` could eagerly fetch domain topics/goals, and resolving a
filter with a full `Domain` object could load its entire repository graph.
The collection now joins only the domain row needed for `domain_slug`, suppresses
other relationships, and resolves domain filters with a scalar ID query.
Canonical/alias records are fetched together in one registry query.
## Production-shaped measurement
Command: `uv run pytest -q -s tests/test_repository_collection_projection.py`
Real local PostgreSQL (`custodian_test`), FastAPI through HTTPX ASGITransport,
with canonical slug records, an alias, and a pre-backfill repository. Measurements
cover HTTP handling and serialization, excluding fixture setup. These are single
local samples, not deployed-service latency or a percentile benchmark. The
regression asserts SQL counts rather than timing thresholds.
| Repository count | Domain filter | SQL SELECTs | Request seconds |
| --- | --- | --- | --- |
| 4 | No | 2 | 0.0377 |
| 131 | No | 2 | 0.0627 |
| 501 | No | 2 | 0.1751 |
| 4 | Yes | 3 | 0.0480 |
| 131 | Yes | 3 | 0.0642 |
| 501 | Yes | 3 | 0.1686 |
Exactly one SELECT reads `repository_slugs` in each nonempty case. The 501-row
case exceeds SQLAlchemy's usual select-in batch size and guards against
accidentally restoring eager relationship batching. An empty collection skips
the registry query. Tests also compare collection/detail behavior for absent
registry entries, missing canonicals, and mismatched canonicals; classification
and secondary-domain filters retain their existing behavior.
This bounds database round trips, not response size: the existing unpaginated
list still requires memory and serialization proportional to the fleet size.
The optional SBOM owner read remains unchanged. No deployment was performed.
## Regression validation
`uv run pytest -q tests/test_repository_collection_projection.py
tests/test_routers_core.py tests/test_repository_alias_routing.py
tests/test_repository_rename_api.py tests/test_repository_rename_persistence.py
tests/test_sbom_nexus_compat.py`
Result: **102 passed in 129.14s**. `git diff --check` passed.