This repository contains mdbplyr, an R package that
provides a disciplined, lazy dplyr-style interface for
MongoDB aggregation pipelines.
mdbplyr sits between raw mongolite usage
and broader compatibility layers. Compared with writing aggregation JSON
by hand, it lets you express supported queries with tidy verbs and
inspect the generated pipeline before execution. Compared with
approaches that try to hide MongoDB behind table-like semantics, it
stays explicit about scope, keeps translation conservative, and fails
clearly when a verb or expression is outside the supported subset.
The practical advantages are:
collect(),cursor(),show_query() for inspectable pipeline output,Suppose there is a running MongoDB instance on localhost
with default port and no authentication. The code below loads
dplyr::starwars into a collection named
starwars.
library(dplyr)
library(mongolite)
library(mdbplyr)
starwars_collection <- mongolite::mongo(
collection = "starwars",
db = "mdbplyr"
)
starwars_collection$drop()
starwars_collection$insert(dplyr::starwars)
starwars_tbl <- tbl_mongo(
starwars_collection,
schema = names(dplyr::starwars)
)Once the collection is loaded, starwars_tbl is the lazy
table used in the examples below.
mdbplyr uses a schema to know which fields are available
in a collection. This matters especially for:
`message.timestamp` or
`message.measurements.Fx`;filter(), mutate(), and
summarise();select(),
flatten_fields(), and unwind_array().The most reliable approach is to pass schema = ...
explicitly when creating the lazy table:
starwars_tbl <- tbl_mongo(
starwars_collection,
schema = c("name", "species", "height", "mass", "homeworld")
)When writing the schema by hand is inconvenient,
infer_schema() can populate it from the first document in
the collection:
This is convenient for exploratory work, but it has an important limitation: it only sees one document. If the collection is heterogeneous, fields that do not appear in the first document may still need to be added manually.
You can inspect the currently known fields with:
Inspect the known schema and the generated pipeline without executing the query.
Each subsection below shows one of the supported
dplyr-like verbs on the starwars
collection.
select()Selecting dotted paths preserves nested MongoDB structure by default. It does not flatten nested fields unless you explicitly ask for that:
mutate()arrange()summarise()flatten_fields()Use flatten_fields() when you explicitly want nested
object leaves to become flat tibble columns. By default the output names
are the schema dot paths.
You can also target a specific nested root and optionally rename the flattened output columns:
unwind_array()Use unwind_array() when a document field contains an
array and you want one output row per array element.
If array elements are nested objects, unwind_array() and
flatten_fields() can be chained:
The features in this section were added in mdbplyr
v0.4.0. Each heading is marked (New from v0.4.0) so it
is easy to tell them apart from the longer-standing subset above.
mdbplyr now records the MongoDB server version when it
can probe it, and some features are gated on a minimum version (for
example median() needs 7.0+ and window functions need
5.0+). When the version cannot be determined it is left unknown and the
feature is allowed through, so the server still has the final say.
group_by() keys (New from v0.4.0)Grouping is no longer limited to bare field names. A
named computed key is translated into the MongoDB
$group._id.
summarise() accumulators (New from v0.4.0)summarise() gained sd(),
var(), first(), last(), and
n_distinct(), alongside the original n(),
sum(), mean(), min(), and
max().
starwars_tbl |>
group_by(species) |>
summarise(
n = n(),
sd_height = sd(height),
var_mass = var(mass),
first_name = first(name),
distinct_worlds = n_distinct(homeworld)
) |>
collect()median() and quantile(probs = ...) are also
supported, but compile to $percentile and therefore require
MongoDB 7.0+:
select() (New from v0.4.0)When the schema is known, select() accepts name-based
tidyselect helpers (starts_with(),
ends_with(), contains(),
matches(), everything(),
all_of(), any_of()), ranges, and negation.
(where() is intentionally not supported, because column
types are unknown without reading data.)
starwars_tbl |>
select(name, starts_with("home"), ends_with("_color")) |>
collect()
# Drop columns with negation, or select from a character vector:
starwars_tbl |>
select(-ends_with("_color")) |>
collect()
cols <- c("name", "height", "mass")
starwars_tbl |>
select(all_of(cols)) |>
collect()Pure bare-name selections still work exactly as before, including
nested root paths such as `message.measurements`.
coalesce() (New from v0.4.0)coalesce() compiles to $ifNull, returning
the first non-missing value.
across() (New from v0.4.0)across() works in both mutate() and
summarise(). It expands a column selection and a function
spec into one assignment per (column, function) pair. .fns
accepts a bare function name, a ~ lambda (using
.x), or a named list of those, and .names glue
({.col}, {.fn}) controls the output names.
# One aggregate over several columns
starwars_tbl |>
group_by(species) |>
summarise(across(c(height, mass), mean, na.rm = TRUE)) |>
collect()
# A named list of aggregates -> {.col}_{.fn} output names
starwars_tbl |>
group_by(species) |>
summarise(across(c(height, mass), list(avg = mean, max = max))) |>
collect()Window functions in mutate() / transmute()
compile to MongoDB $setWindowFields and therefore require
MongoDB 5.0+. They partition by the current
group_by() keys.
Ranking functions take the column to order by (optionally wrapped in
desc()):
starwars_tbl |>
group_by(species) |>
mutate(height_rank = min_rank(desc(height))) |>
select(name, species, height, height_rank) |>
collect()Cumulative and offset windows take their order from a preceding
arrange() (and the stage reorders the output by that
key):
starwars_tbl |>
arrange(height) |>
mutate(cumulative_mass = cumsum(mass)) |>
select(name, height, mass, cumulative_mass) |>
collect()starwars_tbl |>
arrange(height) |>
mutate(
prev_height = lag(height),
next_height = lead(height, default = 0)
) |>
select(name, height, prev_height, next_height) |>
collect()row_number() is available as an alias for the existing
1:n() row-numbering.
inner_join(), left_join(),
semi_join(), and anti_join() compile to
MongoDB $lookup against another collection in the same
database. The right-hand side must be a plain tbl_mongo (a
collection reference with a known schema and no lazy operations).
films_tbl <- tbl_mongo(
mongolite::mongo(collection = "films", db = "mdbplyr"),
schema = c("film_id", "title", "director")
)A mutating join flattens the match by default (one row per matched pair, with colliding columns suffixed):
left_join() keeps unmatched left rows. Filtering joins
keep the left columns and only test whether a match exists:
Pass unnest = FALSE to keep the matched documents as a
native nested array column instead of flattening them:
The examples above stay within the currently supported subset:
.data$... for
explicit field references and .env$... for local
values,mutate() and
transmute(), plus the special 1:n()
row-numbering case,across() in
select()/mutate()/summarise()
(v0.4.0),$setWindowFields (v0.4.0),inner_join() / left_join() /
semi_join() / anti_join() via
$lookup against a plain right-hand tbl_mongo
(v0.4.0),flatten_fields(),unwind_array(),If a query falls outside that subset, mdbplyr is
designed to fail explicitly rather than guess or silently change
execution semantics.