07 / DEPRESSION RISK

Artifacts, process counts and resource budgets

Keep adapter and model operations separate. Resource arithmetic is reproducible; latency, concurrency capacity and cost require measurement.

DSAIL startup

1

Provision

TrueFoundry deploys four replicas

2

Resolve artifacts

Registry FQNs provide model/explainer directories

3

Start workers

Command specifies seven Uvicorn processes

4

Load bundle

Each worker loads model and explainer pickles

5

Align schema

feature_list.yml and model.feature_names_

6

Serve

Port 8000; JWT at gateway; direct model inference

No Gunicorn pre-fork copy-on-write sharing is documented for DSAIL. Model_fn is shown in the request sequence and in startup; exact reuse/loading frequency needs source profiling.

Two-tier resource planner

Adapter replicas DSAIL replicas DSAIL worker processes /replica

Adapter GiB and DSAIL MB remain separate. Model workers do not multiply pod resource reservations. EFS200 GB is the reported artifact volume; its per-pod/shared scope is not established.

Documented

Deployment configuration

Setting	Adapter	DSAIL
Replicas	6	4
CPU request /limit	0.5 /1	5 /6
Memory request /limit	1 /2 GiB	5000 /6000 MB
Temporary storage	Not supplied	1000 /2000 MB
Artifact cache	Adapter mapping from S3	200 GB EFS
Runtime	Python 3.12	Python 3.9
Worker command	Not supplied	--workers 7
Concurrency env	Not supplied	UVICORN_WEB_CONCURRENCY=10
Auth	No app inbound auth	JWT gateway with client_id
Health probes	Disabled in Helm snapshot	/health; does not check external dependencies

Concurrency is unresolved: command says 7 workers/replica; environment says 10; summary sometimes calls 10 workers; table says total 70 while its text calculates 4×7×10=280. Four replicas with seven worker processes means 28 processes. The environment value does not establish a further concurrency multiplier or measured throughput.

Documented

Artifact and version map

Item	Current /planned
Model	model.pickle; CatBoost 1.2.2; model 2023-12-05; deploymentv 12
Explainer	explainer.pickle; SHAP0.43.0
Feature list	feature_list.yml;1,618 features
Registry	TrueFoundry ihs-saas-nonprod-depr-repo
Identity	MODEL_FQN /EXPLAINER_FQN; register_model.py
Resolved directories	depr_model /depr_explainer
Planned S3 model	rcs-ehr-cai-s 3-{env}/depression_risk/v 1/model.pickle
Planned S3 explainer	depression_risk/v 1/explainer.pickle
Overrides	DEPR_MODEL_S3_BUCKET /DEPR_MODEL_S3_KEY /DEPR_EXPLAINER_S3_KEY

S3 hosting and layered monorepo reorganisation are planned/in progress, not a confirmed production cutover. Preserve model, explainer, schema and factor mapping compatibility.

Documented

Runtime environment and pinned stack

Variable /dependency	Purpose /reported value
MODEL_FQN /EXPLAINER_FQN	Registry artifact versions
depr_model /depr_explainer	Local artifact directories
DATA_CAPTURE /ENVIRONMENT	Capture table /prod ornonprod
WORKSPACE_FQN /INTEGRATION_FQN	TrueFoundry workspace /auth integration
ENDPOINT /MODEL_ENDPOINT	Endpoint parameters; not evidence of a model API call in predict_fn
UVICORN_WEB_CONCURRENCY /LOGGING_LEVEL	10 /INFO defaults as reported
API_IHS_CACHE	Adapter cache flag
API_DEPRESSION_POC_SHAP_THRESHOLD	0.04 ADL cutoff
API_DEPRESSION_DIAGNOSIS_SHAP_THRESHOLD	0.0 diagnosis cutoff
Pinned DSAIL stack	Python 3.9; CatBoost 1.2.2; scikit-learn 1.2.0; SHAP0.43.0; numba 0.58.1; numpy 1.24.4

Documented

Observability and performance

Adapter: loguru and Datadog APM with 0.1% sampling. Track both cache avoided calls and downstream latency.
Model: stdout/stderr, TrueFoundry/Kubernetes instrumentation and warehouse capture. Exact live alerts/tracing are not documented.
README states 50–200 ms/resident and SHAP<100 ms, but benchmark methodology is absent. Treat as reported estimates, not SLOs.
~2000+RPS is explicitly estimated; do not quote it as measured capacity. Batch size, worker saturation and payload 1,618 features/resident matter.
Proposed: end-to-end p 50/p 95, downstream errors, capture failure backlog, cache freshness and clinical flag distribution.

Proposed design

Migration and resilience plan

Obtain model card, training lineage and baseline request/response fixtures.
Re-export/retrain pinned model/explainer for a supported Python runtime; compare predictions, SHAP, factor selections and final flags.
If moving artifacts to S3, validate startup identity/access, schema alignment and per-worker memory.
Release version-aware cache invalidation with model rollout; define rollback bundle and cache behaviour.
Add bounded retry/token reuse, inbound adapter auth and enabled probes.
Clinical/regulatory/privacy review and measured staging capacity are gates; owning hosting alone does not transfer training/governance ownership.

Incident playbook

Documented

Tests and branch boundaries

Listed tests: test_model_load.py, test_preprocess.py, test_app.py, test_shap_values.py. README states integration CI on prod pull requests; results are not attached.
Manual checks: /health, authenticated /inferance and /dc-table.
Original DSAIL main is described as development. The rcs-ehr-cai main mirror has imported production code undergoing reorganisation. These are distinct main branches.
The structural tree mentions training workflows/notebooks, while explicit text says training code is absent. Do not infer a reproducible training pipeline from a directory label.
Proposed end-to-end tests should assert adapter evidence selection, seven-category deduplication, cache policy and application-error parsing.