dpyr

dplyr for Python. The tidyverse's verbs — filter, mutate, group_by, summarize, joins, across, tidyselect — as Python method chains, executing on polars or duckdb, with real IDE autocompletion and semantics verified against dplyr itself.

pip install dpyr        # or: uv add dpyr

from dpyr import read, col, n, desc

starwars = read("starwars.parquet")   # read() takes anything tabular:
                                      # .parquet/.csv/.arrow/.db paths, dicts,
                                      # polars/pandas frames, arrow tables,
                                      # Hugging Face datasets, numpy/torch/jax

(
    starwars
    .filter(col.height > 180, col.mass < 100)
    .mutate(bmi = col.mass / (col.height / 100) ** 2)
    .group_by(col.species)
    .summarize(
        n = n(),
        mean_bmi = col.bmi.mean(),
    )
    .arrange(desc(col.mean_bmi))
)

Evaluate that in a notebook and you see rows immediately. Typo a column name and you get the error on that line, with a did-you-mean suggestion. Wrap the same code in a pipeline and only .collect() at the end, and the whole chain runs as one fused query with predicate pushdown. That combination — schema-eager, data-lazy, display-eager — is the core design.

Two backends, one semantics

import duckdb
from dpyr import read

df  = read({"x": [1, 2, 3], "g": ["a", "a", "b"]})   # polars engine
con = duckdb.connect("warehouse.db")
tbl = read(con, "events")                            # SQL pushdown

Identical chains produce identical results on both engines — enforced by a Hypothesis fuzzer that runs random verb chains on both and compares bit-for-bit, and by differential tests against real dplyr: every spec in tests/specs/ is executed by dplyr (via oracle/run_specs.R) to produce a committed golden parquet, then replayed through dpyr on both backends. Where R and the engines genuinely disagree, the decision is documented in docs/SEMANTICS.md, not left to chance.

The dplyr you know

dplyr	dpyr
filter(df, height > 180)	df.filter(col.height > 180)
mutate(df, bmi = mass / h^2)	df.mutate(bmi = col.mass / col.h ** 2)
summarise(df, n = n(), m = mean(x, na.rm = TRUE))	df.summarize(n = n(), m = col.x.mean())
arrange(df, desc(mass))	df.arrange(desc(col.mass))
select(df, name, starts_with("h"))	df.select(col.name, starts_with("h"))
select(df, -mass)	df.select(-col.mass)
across(where(is.numeric), mean)	across(where(is_numeric), "mean")
left_join(a, b, by = "k")	a.left_join(b, on = col.k)
pivot_longer(df, x:y)	df.pivot_longer([col.x, col.y])
if_else(), case_when(), n_distinct()	if_else(), case_when(), .n_unique()
lag(), lead(), row_number(), min_rank()	lag(), lead(), row_number(), min_rank()
cumsum(), dense_rank(), percent_rank()	cum_sum(), dense_rank(), percent_rank()
slice_min(x, n), slice_max(x, n) (ties kept)	slice_min(col.x, n), slice_max(col.x, n)
separate(), unite(), relocate()	separate(), unite(), relocate()
coalesce(), replace_na()	coalesce(), replace_na()

Grouped mutate/filter are windowed per group, summarize peels one grouping level, joins use .x/.y suffixes and match NAs by default — the dplyr behaviors, deliberately.

Autocompletion that actually works

• df.c.height — frame-bound proxy: column names complete from the live schema, and the returned expression is typed (.mean() on numerics, .str_detect() on strings; calling .mean() on a string column raises immediately, at build time).
• df.filter(lambda c: c.height > 180) — lambda style for the same effect.
• dpyr stubgen data/*.parquet -o schemas.py — generates typed schema modules so completion and type-checking work statically in any IDE.

The database is a destination, not just a source

db = read("warehouse.db")                 # catalog object: db.tables, db.orders
gold = db.orders.group_by(col.region).summarize(rev = col.amount.sum())
gold.to_table("gold_revenue")             # CREATE TABLE AS <sql>, fully in-engine
gold.to_view("gold_live")                 # the lazy plan as a named view
gold.write("gold.parquet")                # in-engine COPY (extension dispatch)
mem = read({"region": ["east"], "target": [1000.0]})
gold.inner_join(mem, on = col.region)     # in-memory frames bridge into duckdb
                                          # automatically (arrow, zero-copy)

Interactive by default, lazy when you need it

df.persist()           # checkpoint: materialize now (duckdb: temp table)
df.lazy()              # this frame never executes implicitly
dpyr.options.interactive = False   # global opt-out for production pipelines

Results are cached by plan hash, so re-displaying a frame in a notebook never recomputes it.

Documentation

Full guides at maximerivest.github.io/dpyr — get started, grouped data, joins, window functions, column-wise operations, reshaping, expressions & autocompletion, and the backends guide (connecting and operating polars and duckdb).

Project documents

Doc	What it pins down
docs/DESIGN.md	API design, the materialization model, autocompletion strategy, architecture
docs/SEMANTICS.md	Every deliberate decision where R, polars and duckdb disagree
docs/TESTING.md	dplyr-as-oracle goldens, backend-agreement fuzzing, Hypothesis properties
docs/ROADMAP.md	What shipped in 1.0 and what's next

License

MIT © Maxime Rivest