Appendix D — Appendix D: Production Checklist

SPCS operations, monitoring, and sizing

Keywords

snowflake, R, RStudio, Posit, VS Code, workspace notebooks, snowflakeR, RSnowflake, mlops

Use this checklist before promoting R + snowflakeR workloads from development to production.

D.1 Workspace R: org CRE onboarding (platform / IT)

Reference CRE v1 registered (cre@sfnb_multilang_r or org name) — Bootstrap — CRE vs bootstrap
Image built with create_cre.sh / CI; snow custom-image validate passed
CREATE OR REPLACE CUSTOM RUNTIME ENVIRONMENT after each image push; digest communicated to users
GRANT USAGE ON CUSTOM RUNTIME ENVIRONMENT to notebook roles
Snowsight default or runbook points analysts to CRE (bootstrap only for sandbox)
Separate GPU CRE + BASE_IMAGE_TYPE = GPU if using R/Python GPU packages
Multilang public docs: org operating model

D.2 Configuration & packaging

Pre-built R package tarballs in bootstrap configs and CRE build profile (Appendix B)
snowflaker_*.yaml pins warehouse, database, schema, role — not session defaults
CRE image built and validated for repeat workloads (CRE chapter)
reticulate>=1.25 and compatible rpy2 pinned in conda packages
Python snowflake-ml-python version meets feature requirements (sfr_check_environment())

D.3 Security & network

EAI and network rules reviewed by security — hostname allowlists vs open egress
No PATs or private keys committed to Git notebooks
Service users use key-pair JWT with least-privilege roles
Row access / masking policies applied on sensitive training data

D.4 SPCS & parallel

Docker images built with --platform linux/amd64
Compute pool cycled (suspend/resume) after image updates
Stage I/O via volume mounts on workers — no SQL GET/PUT over REST API
Worker images include all R package dependencies for fit and predict
STAGE_MOUNT env var documented in worker runbooks

D.5 Model Registry & serving

Inference dependencies on conda-forge or custom inference image
Column names and types match between training and scoring
R models tested end-to-end: sfr_log_model() → deploy → sfr_predict() / batch
CustomModel + rpy2 path validated in target SPCS image

D.6 Monitoring & operations

Inference log table populated with timestamp, prediction, optional label columns
sfr_add_monitor() configured for critical models
Alerting wired to drift/performance metrics (dashboard, email, ticketing)
SYSTEM$GET_SERVICE_LOGS and event table access for SPCS fault diagnosis
Warehouse and compute pool auto-suspend policies set

D.7 Documentation & support

Runbook for failed bootstrap (EAI toggle, admin SQL)
Runbook for CRE attach / %%R unknown (wrong runtime or image tag, Appendix G)
Runbook for exec format error / job 395030 (wrong CPU arch)
Ownership and escalation for community packages (snowflakeR, RSnowflake)
Compliance sign-off for community-project disclaimer (Introduction)

D.8 Sizing (guidance)

Workload	Starting point
Interactive notebook	XS–S warehouse; default compute pool
Feature materialization	Size warehouse to refresh SLA
Many-model SPCS	Pool with N workers × `CPU_X64_*` family; benchmark chunk size
Batch inference	Warehouse or `sfr_run_batch()` per registry docs

Benchmark on representative data before locking pool/warehouse sizes.

D.9 Disclaimer

Community packages — validate SLAs, support, and compliance with your organization before production cutover.

--- title: "Appendix D: Production Checklist" subtitle: "SPCS operations, monitoring, and sizing" --- Use this checklist before promoting R + snowflakeR workloads from development to production. --- ## Workspace R: org CRE onboarding (platform / IT) {#sec-cre-org} - [ ] **Reference CRE v1** registered (`cre@sfnb_multilang_r` or org name) — [Bootstrap — CRE vs bootstrap](../../06_workspace_bootstrap/index.qmd#sec-cre-vs-bootstrap) - [ ] Image built with `create_cre.sh` / CI; `snow custom-image validate` passed - [ ] `CREATE OR REPLACE CUSTOM RUNTIME ENVIRONMENT` after each image push; digest communicated to users - [ ] `GRANT USAGE ON CUSTOM RUNTIME ENVIRONMENT` to notebook roles - [ ] Snowsight default or runbook points analysts to CRE (bootstrap only for sandbox) - [ ] Separate **GPU** CRE + `BASE_IMAGE_TYPE = GPU` if using R/Python GPU packages - [ ] Multilang public docs: [org operating model](https://github.com/Snowflake-Labs/snowflake-notebook-multilang/blob/main/docs/org_cre_operating_model.md) --- ## Configuration & packaging {#sec-config} - [ ] Pre-built **R package tarballs** in bootstrap configs and **CRE build** profile ([Appendix B](../B_config_reference/index.qmd)) - [ ] `snowflaker_*.yaml` pins warehouse, database, schema, role — not session defaults - [ ] **CRE image** built and validated for repeat workloads ([CRE chapter](../../09_custom_runtime_and_ml_jobs/index.qmd)) - [ ] `reticulate>=1.25` and compatible `rpy2` pinned in conda packages - [ ] Python `snowflake-ml-python` version meets feature requirements (`sfr_check_environment()`) --- ## Security & network {#sec-security} - [ ] **EAI** and network rules reviewed by security — hostname allowlists vs open egress - [ ] No PATs or private keys committed to Git notebooks - [ ] Service users use **key-pair JWT** with least-privilege roles - [ ] Row access / masking policies applied on sensitive training data --- ## SPCS & parallel {#sec-spcs} - [ ] Docker images built with **`--platform linux/amd64`** - [ ] Compute pool **cycled** (suspend/resume) after image updates - [ ] Stage I/O via **volume mounts** on workers — no SQL `GET`/`PUT` over REST API - [ ] Worker images include all R package dependencies for fit **and** predict - [ ] `STAGE_MOUNT` env var documented in worker runbooks --- ## Model Registry & serving {#sec-registry} - [ ] Inference dependencies on **conda-forge** or custom inference image - [ ] Column names and types match between training and scoring - [ ] R models tested end-to-end: `sfr_log_model()` → deploy → `sfr_predict()` / batch - [ ] CustomModel + rpy2 path validated in target SPCS image --- ## Monitoring & operations {#sec-ops} - [ ] **Inference log** table populated with timestamp, prediction, optional label columns - [ ] `sfr_add_monitor()` configured for critical models - [ ] Alerting wired to drift/performance metrics (dashboard, email, ticketing) - [ ] `SYSTEM$GET_SERVICE_LOGS` and event table access for SPCS fault diagnosis - [ ] Warehouse and compute pool **auto-suspend** policies set --- ## Documentation & support {#sec-docs} - [ ] Runbook for failed bootstrap (EAI toggle, admin SQL) - [ ] Runbook for CRE attach / `%%R` unknown (wrong runtime or image tag, [Appendix G](../G_workspace_container_internals/index.qmd)) - [ ] Runbook for `exec format error` / job 395030 (wrong CPU arch) - [ ] Ownership and escalation for community packages (snowflakeR, RSnowflake) - [ ] Compliance sign-off for community-project disclaimer ([Introduction](../../00_introduction/index.qmd)) --- ## Sizing (guidance) {#sec-sizing} | Workload | Starting point | |----------|----------------| | Interactive notebook | XS–S warehouse; default compute pool | | Feature materialization | Size warehouse to refresh SLA | | Many-model SPCS | Pool with N workers × `CPU_X64_*` family; benchmark chunk size | | Batch inference | Warehouse or `sfr_run_batch()` per registry docs | Benchmark on representative data before locking pool/warehouse sizes. --- ## Disclaimer Community packages — validate SLAs, support, and compliance with your organization before production cutover.