vectra

• Expressions support sin(), cos(), tan(), asin(), acos(), atan(), atan2(), sinh(), cosh(), tanh(), asinh(), acosh(), atanh(), expm1(), log1p() and ^ (#22). A math call with the wrong number of arguments is now an error instead of dropping the extra ones, and sign(), pmin()/pmax() follow base R on NaN.
• Every verb that takes expressions captures them through rlang, so !!, !!!, {{ }}, :=, .data and .env work as in dplyr (#24). group_by() and count() accept !!sym(k), .data[[k]], across(all_of(k)), pick() and computed groups (group_by(odd = x %% 2)), and gain .add.
• spatial_join() gains keep_geom. keep_geom = FALSE drops the left geometry, and with coords = the points are never encoded (#23).

Performance and memory

• vectra_mem() is now a bound on what a query allocates (#19, #28). All the buffering steps of a query (sorts, joins, grouped aggregates, fuzzy joins, k-mer counts) reserve the bytes they actually allocate from one shared pool of vectra_mem() bytes, and spill when it is exhausted; each is guaranteed a quarter of an equal share. Previously each step received the whole budget and counted only part of what it allocated, so a join feeding a grouped aggregate could peak at several times the budget. Temporary spill files are written uncompressed, which is faster at the same peak memory. On 2e7 rows with a 256 MB budget, a join followed by a grouped summary went from 50.1 s and 427 MB above idle to 7.7 s and 168 MB.
• Grouped summarise() aggregates in hash tables instead of always sorting its input (#16). When the groups outgrow their share of the budget, the rest of the input is hash-partitioned to disk and each partition aggregated in turn; results stay sorted by key, and median()/n_distinct() keep their spilling path. On 2e7 rows and 1,000 groups this took 0.9 s instead of 1.9 s, with memory falling from 1.16 GB to under 100 MB.
• arrange() with enough memory emits its result in 131,072-row batches from one permutation instead of as a single batch holding a second copy of the data (#20). arrange() followed by write_vtr() on 2e7 rows went from 54.8 s to 3.5 s.
• Column pruning passes through mutate(), select() and rename(), so a scan reads only the columns a query uses (#17). Filters prune row groups by zone map, sorted-column search and index even when written after a mutate(), select(), arrange() or another filter, and in every input of bind_rows() (#18); the plan no longer depends on the order the verbs are written in.
• spatial_join()'s native path is a lazy streaming node: the result flows into the next verb without a spill and without buffering R data frames (#23). Tagging 1e7 points and counting went from 72.5 s to 6.9 s.
• compress = "fast", the write_vtr() default, runs its LZ stage at the fastest match-search level; streaming a filtered result to .vtr is about 2.8x faster for files about 3.6% larger (#26). compress = "small" also tries that setting, so it is still never larger than "fast".

Bug fixes

• bind_rows() of stores whose column types differ keeps the widened type after column pruning instead of narrowing a double to an integer.
• Filter pushdown no longer stops after the 16th input of bind_rows().
• arrange() documentation now states that NA sorts last in both directions, which is what it has always done.
• explain() no longer prints an offload grade of "streaming scan" for plans that contain a sort or a join; the grade is shown only for replay caches and partitions that carry one (#27).

Installation

• configure now tests the candidate OpenMP flags by building, loading and calling a small parallel routine, and keeps the first that works (#15). On macOS it tries -Xclang -fopenmp -lomp first, so a binary that uses OpenMP records a real dependency on the libomp.dylib that CRAN's R ships instead of looking its symbols up at load time; if nothing works the package builds single-threaded and says so. The package now compiles without warnings when OpenMP is unavailable.

Documentation

• The engine, large-data, spatial, formats and indexing vignettes, and the README, describe the current engine: the tdc container format, hash-first aggregation, the shared memory pool, spilling joins, streaming windows and mapped hash indexes (#27).