RSnowflake in Custom SPCS Services

This vignette covers RSnowflake in custom Snowpark Container Services (SPCS) workloads — for example RStudio Server deployed as a long-running service. It is not the Workspace Notebook path (%%R via rpy2); see vignette("workspace-rsnowflake", package = "RSnowflake") for that.

For the full deploy kit (Docker, service spec, smoke tests), see the snowflakeR package:

system.file("rstudio-spcs", package = "snowflakeR")

Or the Hitchhiker’s Guide chapter RStudio Server on SPCS.

Execution model

In a custom SPCS service container:

  1. Snowflake injects SNOWFLAKE_HOST and /snowflake/session/token.
  2. R runs as a native process (not via rpy2).
  3. dbConnect(Snowflake()) uses the internal gateway + OAuth token.

There is no Python bootstrap cell. Set warehouse, database, schema, and role via container environment variables in the service spec.

Connect

library(DBI)
library(RSnowflake)

con <- dbConnect(Snowflake())
dbGetQuery(con, "SELECT CURRENT_USER(), CURRENT_WAREHOUSE()")

RSnowflake resolves auth in this order (RSnowflake/R/auth.R):

  1. Explicit bearer token argument
  2. SPCS OAuth — SNOWFLAKE_HOST + token file
  3. SNOWFLAKE_PAT
  4. Key-pair JWT

Warehouse and context

Custom SPCS services may not export SNOWFLAKE_WAREHOUSE automatically. Set it in the service YAML:

env:
  SNOWFLAKE_WAREHOUSE: ANALYTICS_WH
  SNOWFLAKE_DATABASE: MY_DB
  SNOWFLAKE_SCHEMA: MY_SCHEMA
  SNOWFLAKE_ROLE: MY_ROLE

RSnowflake passes these to the SQL API session. Do not rely on:

dbExecute(con, "USE WAREHOUSE MY_WH")   # not supported — SQL REST API

The SQL REST API does not support USE, USE WAREHOUSE, GET, or PUT. Use env vars or fully qualified SQL instead.

Verify token injection

Run once after the service starts:

cat("SNOWFLAKE_HOST:", Sys.getenv("SNOWFLAKE_HOST"), "\n")
cat("Token file:", file.exists("/snowflake/session/token"), "\n")

If the token file is missing, check your SPCS service spec and platform version, or fall back to PAT/key-pair against the public endpoint (with EAI).

RStudio Connections pane

RSnowflake registers with the RStudio Connections pane. Probe messages tagged [RSnowflake pane] are normal when the pane checks driver availability. To show verbose pane diagnostics:

options(rsnowflake.connections_pane_verbose = TRUE)

Stage file I/O

For bulk file transfer in custom services, mount a stage as a volume:

volumeMounts:
- name: stage-vol
  mountPath: /data/stage
volumes:
- name: stage-vol
  source: "@MY_DB.MY_SCHEMA.MY_STAGE"

Read and write via the filesystem path in R. Do not use GET/PUT SQL with RSnowflake.

Optional ADBC

When adbcsnowflake is baked into the image, RSnowflake can use Arrow-native bulk paths via the same internal gateway. See WORKSPACE_ADBC.md in the RSnowflake repository for gateway and redirect-chain detail (applies to SPCS containers as well as Workspace).

Companion: snowflakeR

ML platform APIs (sfr_connect(), Feature Store, Model Registry) require Snowpark Python via reticulate. In custom SPCS services, sfr_connect() does not auto-detect the environment — use sfr_connect_spcs() from the rstudio-spcs kit or vignette("rstudio-spcs", package = "snowflakeR").