Skip to contents

Changelog

Source: NEWS.md

RTrino 0.1.0

First release: a DBI backend for Trino over the cluster’s HTTP REST API.

Connections

  • Trino(), dbConnect(), dbDisconnect() and dbIsValid() manage a connection’s life cycle. dbConnect() probes /v1/info, so a wrong address or a rejected credential fails at connection time. Its arguments are validated before the coordinator is contacted, with plain messages.
  • trino_auth_basic(), trino_auth_jwt() and trino_auth_oauth2() supply credentials, and trino_ssl() sets the TLS options, including ca_bundle to trust an internal certificate authority without turning verification off. dbConnect() refuses to send credentials to an http:// host unless allow_http_auth = TRUE.
  • timeout bounds each HTTP request (60 seconds by default) and query_max_run_time bounds a query on the cluster.
  • 429, 502, 503 and 504 responses are retried after a short pause, as Trino’s protocol asks.

Queries and results

  • dbSendQuery(), dbFetch(), dbGetQuery(), dbExecute(), dbSendStatement(), dbHasCompleted() and dbClearResult() run queries, with automatic nextUri pagination, chunked fetching through dbFetch(n = ), and server-side cancellation when a running result is cleared.
  • dbExecute() returns the number of rows Trino says a statement changed, and 0 for DDL. dbSendStatement() runs the statement to completion, so that dbGetRowsAffected() answers straight away.
  • dbColumnInfo() and dbFetch(n = 0) describe the columns before the first row is fetched.
  • dbIsValid() on a result reports whether it has been cleared, not whether the query has finished, as DBI requires: dbColumnInfo() and dbGetRowCount() keep working until dbClearResult(). dbHasCompleted() reports that every row has been consumed.
  • Query failures are raised as classed conditions, trino_query_error and trino_canceled, carrying the server’s error_name, error_code, error_type and query_id, so a caller can react to a specific Trino error without matching on the message text.
  • Queries do not pause between the empty pages Trino returns while it queues and plans. A capped pause still guards against a server that returns empty pages instantly and forever.
  • Passing params is an error, and dbBind() says that parameters are not supported. Put values into a statement with DBI::sqlInterpolate() or glue::glue_sql(): dbQuoteLiteral() writes typed Trino literals (DATE '...', TIMESTAMP '...', TRUE, doubles that read back exactly).

Types

  • BIGINT is returned as exact bit64::integer64 by default (bigint = "numeric" or "character" for the alternatives), and DOUBLE bit for bit. Trino sends both as JSON numbers, which are converted without going through text. The lowest BIGINT, which bit64 keeps for NA, is NA with a warning.
  • Timestamps keep their microseconds, and TIMESTAMP WITH TIME ZONE is resolved per value whether the server sends a zone name or an offset.
  • A column of a type RTrino does not know warns once per result; INTERVAL columns are read as text, and a bare NULL as a logical.
  • dbDataType() and trino_type_to_r() map between R and Trino types.

Tables, writes and transactions

  • dbListTables(), dbExistsTable(), dbListFields() and dbGetInfo() read the catalog. Table names are strings of one to three parts, DBI::Id() or quoted identifiers, and are matched ignoring case, as Trino does. dbExistsTable() counts a row in information_schema.tables, so its cost does not grow with the size of the schema, and reports FALSE rather than an error for a name in a catalog that does not exist.
  • dbWriteTable(), dbAppendTable() and dbCreateTable() write data frames. Rows go in as INSERT INTO ... VALUES statements of chunk_size rows each (1000 by default), every value a literal, since Trino’s protocol has no parameter binding. dbWriteTable(overwrite = TRUE) renames the existing table rather than dropping it, and restores it if the write fails. dbRemoveTable() and dbRenameTable() drop and rename tables, and temporary = TRUE is an error, because Trino has no temporary tables.
  • dbBegin(), dbCommit(), dbRollback() and dbWithTransaction() run statements inside a Trino transaction. Whether a ROLLBACK undoes a write depends on the connector behind the catalog; dbBegin() warns for the catalogs that hold no real storage (memory, system, tpch, tpcds, jmx and blackhole).

dplyr

  • dplyr pipelines are translated to Trino SQL through a dbplyr dialect, which requires dbplyr >= 2.6.0. The methods are registered when dbplyr is loaded rather than through NAMESPACE, because a NAMESPACE directive for sql_dialect(), which older versions lack, would make such a dbplyr fail to load.
  • Translations are written so that an expression computes in Trino what it computes in R: / divides as a double, %% takes the sign of the divisor, as.integer() and as.integer64() truncate, grepl(), sub() and gsub() honour fixed and ignore.case and accept R’s \\1 back-references, and paste(), paste0() and str_c() accept columns of any type. days(), weeks(), months(), years(), hours(), minutes() and seconds() build intervals, for date + days(1). A Date or POSIXct from R becomes a typed literal in a pipeline.
  • filter_out(), from dplyr 1.2.0, uses Trino’s native IS DISTINCT FROM rather than the CASE WHEN comparison dbplyr falls back to. dplyr 1.2.0’s when_any()/when_all() and recode_values()/replace_values()/ replace_when() are not translated, because dbplyr 2.6.0 has no translation for them on any backend.
  • copy_to() and compute() write through dbWriteTable(), and need temporary = FALSE and a table name. compute() does not run ANALYZE, which not every connector supports.
  • Column discovery uses dbplyr’s default (WHERE 0 = 1) rather than LIMIT 0. Both plan without reading data on Trino, and the default avoids depending on dbplyr internals.
  • simulate_trino() returns a connection that carries the dialect without a session, for inspecting translated SQL with no cluster to hand.
  • The vignette lists the few places where a translation still differs from R: round() on halves, the approximate median() and quantile(), DECIMAL beyond 15 digits, and paste() of a DOUBLE.

Testing

  • The test suite runs offline against a fake coordinator that replays responses recorded from a real Trino (dev/record-fixtures.R), so everything about how values are encoded is tested against payloads a cluster really sends.
  • tests/testthat/test-live.R runs against a real Trino when RTRINO_TEST_URL is set, and CI runs it against trinodb/trino.