R bindings to libmdbx, an embedded transactional key-value store. The library is vendored and compiled into the package, so there is no server to run and nothing to install beside it.
A database is a single file. Keys and values are bytes, held in sorted key order and optionally split across named databases within that one file. Every read and write happens inside a transaction: commits are ACID and fully durable by default, and a reader sees a consistent snapshot without blocking the writer. Several processes can share a database — many readers, one writer.
Three limits shape how it is used. An environment runs one transaction at a time, because libmdbx binds a transaction to the thread that began it and R is single-threaded. A handle does not survive fork(), so each process opens its own. And storing bytes is where the package stops: serializing R objects is left to you, as are cursors, duplicate keys and batched calls, which are not implemented yet.
Installation
mdbx is not on CRAN yet. Install the development version from GitHub with:
# install.packages("pak")
pak::pak("pedrobtz/mdbx")Usage
Data lives in an environment (a file on disk, .mdbx by convention), and every read or write happens inside a transaction: mdbx_with_read() and mdbx_with_write() open one, run your code, and commit if it returns or abort if it throws. Keys and values are bytes: a string goes in as its UTF-8 bytes, so "k" and charToRaw("k") are the same key. Reads decode back to text by default, which MDBX cannot vouch for: pass as = "raw" for the rest; a wrong guess errors, never corrupts.
library(mdbx)
path <- tempfile(fileext = ".mdbx")
env <- mdbx_env_open(path)
mdbx_with_write(env, function(txn) {
mdbx_put(txn, "answer", "42")
mdbx_put(txn, "config", serialize(list(retries = 3L), NULL))
})
mdbx_with_read(env, function(txn) {
mdbx_get(txn, "answer")
})
#> [1] "42"
# Not text, so read the bytes and decode them yourself.
mdbx_with_read(env, function(txn) {
unserialize(mdbx_get(txn, "config", as = "raw"))
})
#> $retries
#> [1] 3
# A key that is not there reads as NULL.
mdbx_with_read(env, function(txn) mdbx_get(txn, "missing"))
#> NULL
mdbx_env_close(env)Listing what is there
mdbx_keys() and mdbx_items() walk the database in key order, in a single crossing into C rather than one per record.
env <- mdbx_env_open(tempfile(fileext = ".mdbx"))
mdbx_with_write(env, function(txn) {
mdbx_put(txn, "banana", "2")
mdbx_put(txn, "apple", "1")
})
mdbx_with_read(env, function(txn) mdbx_keys(txn))
#> [1] "apple" "banana"
items <- mdbx_with_read(env, function(txn) mdbx_items(txn))
stats::setNames(items$values, items$keys)
#> apple banana
#> "1" "2"
mdbx_env_close(env)Keys come back in key order, not insertion order. Both accept as = "raw", needed for keys or values that are not text, and both take limit to bound the read. A scan with no limit refuses to run past mdbx_scan_max (a million records) rather than quietly materializing the lot — pass limit = n, or limit = Inf to mean it.
Managing a transaction by hand
mdbx_txn_begin(), mdbx_txn_commit() and mdbx_txn_abort() drive a transaction directly, for when the decision to keep the writes is only made at the end. The worked example shows the pattern; the with_* helpers are those calls plus on.exit().
Testing
testthat covers the API, the fork() and cross-process contracts, and panic recovery. Generated operation sequences are replayed against a reference state model: valid calls must agree with it, forged handles must be refused changing nothing, and injected faults must stay recoverable. CI adds R CMD check on five OS/version legs, ASan/UBSan, valgrind, LTO, gctorture, rchk and shuffled test order; tools/interop-check.sh round-trips against an independently built libmdbx.