polyglotSQL

polyglotSQL parses, tokenizes, validates, formats, analyzes and translates SQL between 30+ dialects — PostgreSQL, MySQL, BigQuery, Snowflake, DuckDB, SQLite, T-SQL, Oracle, ClickHouse, Trino, Spark and many more — directly from R. It embeds the polyglot-sql Rust crate (a Rust port of the excellent SQLGlot Python library), so everything runs natively in-process: no Python, no Java, no database connection, no network calls.

Why polyglotSQL?

Analytics teams rarely live in a single SQL dialect. Warehouses migrate (Redshift → Snowflake, anything → DuckDB), pipelines target several engines at once, and auditing queries requires understanding them structurally, not as strings. polyglotSQL gives R a real SQL compiler front-end: a proper parser, an AST, dialect-aware generators, column-level lineage and structural analysis.

How it compares to related R packages:

Package	What it does	How polyglotSQL differs
SqlRender	OHDSI templating + rule-based translation from its own SQL flavor (Java)	polyglotSQL parses any dialect into an AST and regenerates it; no Java, no template markup
SQLFormatteR	SQL formatting	formatting is one of a dozen polyglotSQL features
sqlfluffr	wrapper around the Python sqlfluff linter	polyglotSQL needs no Python; validation comes from a native parser
dbplyr	generates SQL from dplyr code per backend	polyglotSQL starts from existing SQL and translates/analyzes it

Installation

Binaries built by CI need no Rust toolchain. Installing from source requires Rust (cargo and rustc >= 1.88):

# install Rust once, via rustup:
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

Then:

# install.packages("remotes")
remotes::install_github("StrategicProjects/polyglot-sql-r")

The package builds fully offline: all Rust dependencies are vendored in the source tarball and nothing is downloaded during R CMD INSTALL.

Transpile SQL between dialects

library(polyglotSQL)

sql_transpile(
  "SELECT IFNULL(a, b) FROM t",
  from = "mysql",
  to = "postgres"
)
#> [1] "SELECT COALESCE(a, b) FROM t"

sql_transpile(
  "SELECT TOP 5 name, GETDATE() FROM users",
  from = "tsql",
  to = "duckdb"
)
#> [1] "SELECT name, CURRENT_TIMESTAMP FROM users LIMIT 5"

By default unsupported constructs raise a classed error (unsupported = "raise"); use "warn" or "ignore" to force best-effort output.

Format

cat(sql_format("select id,sum(x) total from t where y=1 group by id"))
#> SELECT
#>   id,
#>   SUM(x) AS total
#> FROM t
#> WHERE
#>   y = 1
#> GROUP BY
#>   id

Parse and inspect

ast <- sql_parse("SELECT a, b FROM t WHERE x = 1")
ast
#> <polyglot_ast> 1 statement (generic)
#> [1] select

sql_tokenize("SELECT a FROM t")
#> <polyglot_tokens> 4 tokens
#>     type   text line column start end
#> 1 Select SELECT    1      7     0   6
#> 2    Var      a    1      9     7   8
#> 3   From   FROM    1     14     9  13
#> 4    Var      t    1     16    14  15

Validate

sql_validate("SELECT FROM WHERE")
#> <polyglot_validation> invalid (generic)
#> [E003] Expected table name or subquery, got Where at 1:18

sql_validate(
  "SELECT nonexistent FROM t",
  schema = list(t = c(id = "INT", name = "TEXT"))
)
#> <polyglot_validation> invalid (generic)
#> [E201] Unknown column 'nonexistent' in table 't'

Column-level lineage

sql_lineage(
  "WITH base AS (SELECT id, amount FROM payments)
   SELECT id, amount * 2 AS doubled FROM base"
)
#> <polyglot_lineage> 2 columns (generic)
#> id ← payments
#> doubled ← payments

Supported dialects

sql_dialects()
#>  [1] "generic"     "postgresql"  "mysql"       "bigquery"    "snowflake"  
#>  [6] "duckdb"      "sqlite"      "hive"        "spark"       "trino"      
#> [11] "presto"      "redshift"    "tsql"        "oracle"      "clickhouse" 
#> [16] "databricks"  "athena"      "teradata"    "doris"       "starrocks"  
#> [21] "materialize" "risingwave"  "singlestore" "cockroachdb" "tidb"       
#> [26] "druid"       "solr"        "tableau"     "dune"        "fabric"     
#> [31] "drill"       "dremio"      "exasol"      "datafusion"

Semantic limitations

Transpilation is syntactic and best-effort. Two engines can accept the same query text and still disagree about implicit casts, collations, NULL ordering, integer division, time zones, or function edge cases. polyglotSQL (like SQLGlot) does not emulate engine semantics.

Always test transpiled SQL against the target database before relying on it in production.

Acknowledgements and licensing

polyglotSQL (MIT) statically links the polyglot-sql crate by Tobias Müller (MIT), which derives from SQLGlot by Toby Mao (MIT). Licenses of all vendored Rust crates are collected in inst/COPYRIGHTS.