h2oplyr

Tidy-style data manipulation with H2O as the only data backend. The only non-base runtime dependency is h2o. knitr and rmarkdown are optional vignette-building dependencies. No dplyr, tidytable or data.table is used.

# Install h2o first, then this package from its source archive.
install.packages("h2oplyr_0.3.0.tar.gz", repos = NULL, type = "source")
h2o::h2o.init(nthreads = 2, max_mem_size = "2G")
library(h2oplyr)
p <- h_penguins()
h_summarize(p, n = n(), mass = mean(body_mass_g, na.rm = TRUE), .by = species)
vignette("h2oplyr-penguins")

The matching H2O server and Java runtime are external system requirements. This package never starts or shuts down a cluster automatically.

See the vignette for the supported expression language, examples and limits. Help pages have independently executable examples for all 76 exports. Additional asserted examples and regression tests are installed under system.file("examples", package = "h2oplyr").

Grouped windows (0.3.0)

p |> h_filter(!is.na(body_mass_g)) |>
  h_mutate(input_row = h_row_number()) |>
  h_arrange(species, body_mass_g, input_row) |>
  h_mutate(previous_mass = h_lag(body_mass_g),
           rolling_mass = h_roll_mean(body_mass_g, k = 3),
           mass_rank = h_min_rank(desc(body_mass_g)),
           position = h_row_number(), .by = species)

Windows respect the current order within each group and return the input row order. h_group_by() is also supported. Grouped mutation accepts only direct calls to the five supported helpers (including h_lead()); other expressions must be calculated before grouping or after explicit ungrouping. Missing group keys form groups. See NEWS.md for the release history and help pages for NA, tie and partial-window rules. Large rolling widths require more remote shifts.

Migration from the standalone prototype

• h2o_tbl() requires an existing H2OFrame; upload local files with H2O first.
• h_collect() returns the H2OFrame; it does not collect rows locally.
• h_nest_by() returns remote keys and a list of remote handles.
• Unsupported grouped operations and arbitrary R functions fail explicitly.
• Grouping and join character keys are encoded as factors in H2O. Joins have unspecified order. Bind rows requires matching column-name sets.
• Pivots reject duplicate cells; aggregate first. Empty summaries are rejected.
• Selection helpers inspect names/types. They never inspect a local data sample.

Tests

Start a matching H2O cluster explicitly on a local port, then set H2OPLYR_TEST_PORT when running R CMD check. Without that variable, static tests run and integration tests state that they were skipped. Live integration tests verify that H2O namespace function bindings remain unchanged. Rebuilding vignettes against a live server can be requested with H2OPLYR_VIGNETTE_PORT.

Copyright 2026 Carlos Ortega. Code: MIT. Included Palmer penguin data: CC0; see inst/extdata/PROVENANCE.md.