A complete R client for the Databento market data HTTP API. Databento publishes clients for Python, C++ and Rust but not for R; this fills that gap, and does it without requiring Python or the binary DBN format.
Everything the HTTP API offers is covered: dataset and schema discovery, free cost and record-count previews, streaming range downloads, symbology resolution, asynchronous batch jobs, and the reference data endpoints for corporate actions, adjustment factors and the security master. Argument names, defaults and the bytes on the wire mirror the official Python client, so a script translates line by line. Results come back as tibbles, or go straight to parquet.
Installation
# install.packages("pak")
pak::pak("sstoeckl/databentoR")Usage
library(databentoR)
db_list_datasets()
db_list_schemas("GLBX.MDP3")
# how far back does each schema actually go?
db_get_dataset_range("OPRA.PILLAR")
args <- list(dataset = "GLBX.MDP3", start = "2024-01-02", end = "2024-01-09",
symbols = "ES.FUT", schema = "ohlcv-1d", stype_in = "parent")
# previews are free; the download is billed per gigabyte
do.call(db_get_cost, args) # US dollars
do.call(db_get_record_count, args) # rows you would get
es <- do.call(db_get_range, args) # a tibblevignette("databentoR") walks through the rest: symbology, batch jobs, reference data, and exact nanosecond timestamps.
Why the types are never guessed
Column types are assigned from the field name rather than inferred. That is not fussiness: a trades slice whose action column happens to be all "T" infers as boolean, which would silently turn every trade action into TRUE. db_field_types() reports what databentoR will assign, and the live test suite checks that table against db_list_fields().
Timestamps come back as POSIXct, which stores seconds as a double and so resolves to about a quarter of a microsecond on modern dates. Pass ts_type = "integer64" when you need the exact nanosecond count. Fields that are 64-bit on the wire are never downcast either, so order_id and an undefined statistics quantity keep their value.
Row order is the server’s own, and the server does not promise a stable order among records sharing a timestamp: two downloads of one slice can return the same rows in a different sequence. Sort if you need reproducibility.
Equivalence with the Python client
The official client is pinned as a git submodule, and agreement with it is a test rather than a claim.
- Wire equivalence, offline and free, on every push. A stub transport records the exact request the pinned Python client would send for all 28 endpoints; the suite asserts that databentoR builds the same method, endpoint, and parameter names, values and order. It also checks the R value domains against the compiled DBN enums, so a schema added upstream fails the tests instead of a user’s request.
-
Data equivalence, weekly and billed in cents. Six tiny slices are frozen through the Python client and compared column by column: prices to a relative tolerance of
1e-12, timestamps and counts exactly. - Upstream watch, weekly. When a new Python release appears, the wire fixture is re-captured against it and the diff is filed as an issue naming the calls that changed.
The one deliberate divergence is timeseries.get_range, where databentoR asks for the CSV encoding instead of DBN. vignette("equivalence") documents that and the five remaining differences.
Downloaded market data is never committed. Databento’s licence does not permit redistribution, so the reference fixtures stay untracked.