An MLOps course project with a factory-based FastAPI application named
recommender, served by Uvicorn. App and router instances are created inside
factories. RecommenderApplication owns the app and receives its service in the constructor;
RecommendationRouter owns the versioned routes and service reference.
The injected RecommendationService implements __aenter__ and __aexit__
for model loading and deterministic shutdown. Its idempotent close() also
runs from __del__ as a fallback. Application, router and service state is
private; app and router access use explicit read-only properties. Imports
do not start services or create shared application instances. Response contracts
are exported from internal/models.
Implementation lives under internal/recommender/, with other future apps owning
sibling packages under internal/ and separate cmd/ entry points.
Run recommender
bazel_agent bazel run //projects/mlops_2026/cmd/recommender -- --host 127.0.0.1 --port 8000
The defaults are 127.0.0.1:8000. GET /health returns {"status": "ok"}.
FastAPI serves the generated schema at /openapi.json, Swagger UI at /docs,
and ReDoc at /redoc; the documentation UIs use FastAPI’s default external
assets. GET /api/v1/recommend?track_name=Time&n=5 returns requested_track
and song recommendations containing track_name, artists, and album_name.
track_name is required; n defaults to 5 and must be between 1 and 20.
An unknown track returns an empty list. Invalid queries return HTTP 422.
The pinned upstream data.csv and dataset.csv are included in the executable
runfiles and resolved independently of the working directory. Startup checks
that both exist. Each app loads its packaged model once during startup using
the upstream load_model function, then releases it on shutdown. Health checks
do not invoke recommendation computation. Source and integrity are
owned by third_party/com_github_yandex_practicum_mlops_freetrack/.
Build the course model
bazel_agent bazel build //projects/mlops_2026:model
The sandboxed target runs the upstream train_and_save_model.py unchanged,
stages its CSV inputs in an action directory, and produces
bazel-bin/projects/mlops_2026/model.pkl. It exposes a declared build artifact
consumed by the recommender executable. The upstream script is
also buildable as
//third_party/com_github_yandex_practicum_mlops_freetrack:train_and_save_model;
running that script directly requires a working directory with both CSVs and
writes model.pkl there. Prefer the model build target for isolated execution.
The full training data joins to 6,788 unique tracks. Its dense similarity matrix
uses about 0.34 GiB; the serialized artifact is about 359 MiB. The model target and isolated model
verification are marked manual, but building or testing the recommender now
requires this artifact and can trigger training when it is not cached. Shared Python dependencies are declared and locked under tools/py/.
Verify
bazel_agent bazel test //projects/mlops_2026/test/e2e:recommender_test
bazel_agent bazel test //projects/mlops_2026/test/e2e:model_test
These checks exercise real HTTP, factory isolation, startup conflict handling,
shutdown, service context exit, repeated and deletion cleanup, course runfiles
and checksums, and a recommendation from the built
model. Repeatable JSON artifacts are saved in each test’s undeclared outputs
under bazel-testlogs/projects/mlops_2026/test/e2e/.
There is no deployment or release target.