columna API reference - v0.3.0
    Preparing search index...

    columna vs Arquero / DuckDB-Wasm / Polars / DuckDB — same data, same operations, checked results

    Generated 2026-09-17 · Node v24.19.0 · win32 x64 · 2 000 000 rows × 8 columns (2 int, 4 float, 2 low-cardinality text), CSV 107 MB. Each library in its own process. Cold total = sum of timed operation durations on the first run in that process (module import / DuckDB-Wasm init are outside these timers). Warm = median of 2 further runs. Peak RSS sampled every 20 ms (synchronous work can hide intermediate peaks). Times in ms.

    Library Version Cold total Warm total Peak RSS Notes
    columna 0.2.1 1728 1694 1102 MB
    arquero 8.0.3 9199 10362 3546 MB reads CSV from a string (fs.readFile + fromCSV); filter/derive/orderby/join/unique reified so the work is inside the timing; corr is JS pearson over column arrays
    duckdb-wasm 1.33.1-dev57.0 4104 3962 1847 MB single-threaded WebAssembly build (eh), blocking Node bindings; the browser deployment target
    polars 0.26.1 2203 2303 4327 MB eager API; native multi-threaded core
    polars-lazy 0.26.1 3235 3828 2532 MB lazy API: each op = scanCSV + op + collect (read is inside every timing; query optimiser + streaming)
    duckdb 1.4.4 968 1118 1706 MB native addon, multi-threaded

    Rows are operations; columns are libraries. Lower is better. polars-lazy includes CSV scan inside every op (its read cell is 0 by design).

    Operation columna arquero duckdb-wasm polars polars-lazy duckdb
    read CSV 269 1752 790 47 0 252
    filter (age>30 ∧ salary>45k) 35 169 149 40 201 296
    select 4 columns 0 0 106 0 313 19
    drop (z,y) 0 0 166 0 106 23
    rename city→town 0 0 192 0 95 29
    withColumn (xy = x+y) 6 18 212 8 113 31
    groupBy city + agg 30 230 33 41 145 7
    sort salary desc 175 1509 530 119 206 96
    sort city asc, salary desc 236 2743 640 1140 301 109
    unique (age,city) 180 291 60 53 193 15
    head 1000 0 0 3 0 2 1
    tail 1000 0 0 5 0 95 11
    sample n=1000 1 13 2 2 99 4
    inner join regions on city 47 234 36 39 192 5
    left join regions on city 62 357 35 23 155 4
    semi join regions on city 43 215 33 37 161 3
    anti join regions on city 21 123 35 30 148 4
    melt x,y (id=id,city) 15 729 230 43 124 59
    pivot segment × mean(salary) 257 704 49 106 186 8
    valueCounts(city) 9 112 28 41 142 6
    describe (age mean) 257 85 3 82 205 1
    corr(x, y) 37 61 10 399 468 3
    filter age>30 → groupBy city sum(salary) 29 144 32 53 169 5
    writeCsv (filtered) 44 1288 603 51 44 137
    Operation columna arquero duckdb-wasm polars polars-lazy duckdb
    read CSV 248 1638 801 44 0 253
    filter (age>30 ∧ salary>45k) 36 171 171 40 109 203
    select 4 columns 0 0 127 1 78 20
    drop (z,y) 0 0 167 0 86 19
    rename city→town 0 0 192 0 105 22
    withColumn (xy = x+y) 5 13 214 10 98 24
    groupBy city + agg 31 127 38 44 126 4
    sort salary desc 163 1037 556 100 211 97
    sort city asc, salary desc 219 2502 655 1128 298 107
    unique (age,city) 173 252 64 53 171 14
    head 1000 0 0 7 0 1 1
    tail 1000 0 0 6 0 84 11
    sample n=1000 1 13 4 1 94 4
    inner join regions on city 42 233 39 32 150 5
    left join regions on city 58 421 36 22 121 3
    semi join regions on city 43 207 34 39 138 3
    anti join regions on city 20 107 35 26 130 3
    melt x,y (id=id,city) 15 668 225 30 116 28
    pivot segment × mean(salary) 278 260 49 68 181 5
    valueCounts(city) 6 115 28 36 116 3
    describe (age mean) 259 85 3 80 195 1
    corr(x, y) 58 63 10 363 457 2
    filter age>30 → groupBy city sum(salary) 28 148 30 43 125 4
    writeCsv (filtered) 43 1139 614 44 44 133

    Reference (columna): filter 1240671, select 4, drop 6, rename 1, Σxy 199965448.286, groupBy 100355262.877, sort 139999.95/20000.08, sortMulti first salary 139999.95, unique 300, head Σ 499500, tail Σ 1999499500, sample rows 1000, join 1600000/127980215810.788, left 2000000, semi 1600000, anti 400000, melt 4000000/199965448.286, pivot Σ 800113.544, valueCounts Σ 2000000, describe age μ 47.525, corr(x,y) 0, filter→groupBy Σ 125361576887.11.

    • arquero: summary-match
    • duckdb-wasm: summary-match
    • polars: summary-match
    • polars-lazy: summary-match
    • duckdb: summary-match
    • Wall-clock of a single process on one machine; Polars / native DuckDB use every core end-to-end. columna engages its worker pool / native Rayon kernels on heavy ops above per-op thresholds (see README → Engines); Arquero and DuckDB-Wasm stay single-threaded.
    • polars-lazy pays the CSV scan inside every operation (that is the point of a streaming optimiser); compare its per-op numbers with read + op of the eager rows.
    • Peak RSS includes the runtime itself (a WebAssembly heap, Polars’ thread pool) and the CSV read; it is the number to budget, not the column bytes. Same-process 20 ms sampling underestimates peaks during long synchronous stretches.
    • SQL engines (DuckDB / DuckDB-Wasm) time aggregate-over-join (count/sum on a join), not full join materialization — do not compare that join cell directly to columna/Arquero/Polars full join output.
    • unique keeps the first row per (age,city) key (~300 groups from 2M rows) — a hash-dedup stress with a tiny result.
    • head / tail are first/last 1 000 rows in storage order (head id Σ = 499500 on this fixture).
    • sample checks only row count (RNG differs across libraries). corr is Pearson of x and y.
    • summary-match checks aggregates (row counts, sums, sort extremes) with tolerance — not byte-identical tables.
    • columna numbers come from collect() on the CPU engine; collectWithReport() confirms no other backend was involved.

    Settings
    Member Visibility

    On This Page
    columna vs Arquero / DuckDB-Wasm / Polars / DuckDB — same data, same operations, checked results