v0.3.0

What's Changed

• Map the file for a parallel CSV read by @randyzwitch in #144
• Take signed decimals on the one-pass Float64 parser by @randyzwitch in #145
• Pipeline CSV decoding and reduce parsing and assembly overhead by @randyzwitch in #153
• Avoid reparsing wide decimal floats in CSV and casts by @randyzwitch in #154
• Accelerate numeric reductions, ordered joins, and decimal parsing by @randyzwitch in #155
• CSV: translate Polars atoi_simd integer parsing by @randyzwitch in #159
• Replace public CSV reader with the Polars-derived pipeline by @randyzwitch in #161
• Avoid owned strings in numeric casts and measure the path by @randyzwitch in #162
• Append single CSV chunks directly during assembly by @randyzwitch in #163
• Stop rechunking fused CSV expressions on every batch by @randyzwitch in #164
• Move benchmark research to wiki and remove third_party directory by @randyzwitch in #166
• Filter large chunked frames without input rechunking by @randyzwitch in #167
• Keep CSV source attribution in central notice by @randyzwitch in #168
• Keep moderately repeated grouping keys on serial route by @randyzwitch in #169
• Encode chunked string group keys without rechunking by @randyzwitch in #170
• Gather serial group keys from source chunks by @randyzwitch in #171
• Assemble parallel string gathers as chunks by @randyzwitch in #172
• Gather medium multicolumn frames by column by @randyzwitch in #173
• Gather partitioned group inputs in one pass by @randyzwitch in #174
• Share left columns for identity-order join output by @randyzwitch in #175
• Rechunk parallel gather inputs by column by @randyzwitch in #176
• Count grouped bare columns without row batches by @randyzwitch in #177
• Accumulate Float64 reductions eight values at a time by @randyzwitch in #178
• Keep CSV integer source history out of code comments by @randyzwitch in #179
• Size CSV decode workers to mapped input by @randyzwitch in #180
• Scan chunked Float64 sums without materializing chunk views by @randyzwitch in #181
• Sort medium low-cardinality first keys by independent buckets by @randyzwitch in #182
• Pack fused expression validity by byte by @randyzwitch in #184
• Use constant-size hash storage for group cardinality sampling by @randyzwitch in #186
• Reuse hash worker histograms for group scatter by @randyzwitch in #187
• Encode single string group keys in parallel by @randyzwitch in #189
• Select simple Float64 filter rows directly by @randyzwitch in #183
• Use direct row lookup for dense ascending Int64 right joins by @randyzwitch in #190
• Evaluate large fused expressions on source chunk windows by @randyzwitch in #185
• Keep parallel expression outputs chunked by @randyzwitch in #188
• Split large sorted chunk gathers across workers by @randyzwitch in #191
• Speed up low-cardinality Int64 grouping with direct key lookup by @randyzwitch in #192
• Reduce filter row-index concatenation allocations by @randyzwitch in #193
• Use standard sort for parallel index runs by @randyzwitch in #194
• Inline hot CSV field splitting path by @randyzwitch in #195
• Filter aligned chunks directly with bounded workers by @randyzwitch in #196
• Keep radix sort keys local to low-cardinality buckets by @randyzwitch in #197
• Avoid partition gather for skewed small-domain groups by @randyzwitch in #198
• Parse short CSV integers with scalar digits by @randyzwitch in #199
• Preallocate row evaluation outputs by @randyzwitch in #201
• Skip redundant bounds checks in short CSV integers by @randyzwitch in #203
• Scan plain CSV records with one delimiter vector by @randyzwitch in #204
• perf: write fused comparison results directly to bitmap by @randyzwitch in #205
• Aggregate numeric sum and count directly by hash bucket by @randyzwitch in #200
• perf: preserve absent validity in all-valid column gathers by @randyzwitch in #206
• Scale CSV chunk count with input size by @randyzwitch in #207
• Evaluate expressions in eight-row SIMD batches by @randyzwitch in #208
• Use larger bounded batches for with_columns by @randyzwitch in #209
• Decode CSV columns in bounded row groups by @randyzwitch in #210
• Skip validity work for all-valid parallel gathers by @randyzwitch in #211
• Fuse low-cardinality grouped sum and count across workers by @randyzwitch in #212
• Pipeline lazy CSV filters through decode workers by @randyzwitch in #213
• Fuse bounded grouped Float64 aggregates across worker ranges by @randyzwitch in #214
• Defer sort gathers and push row-local filters below sort by @randyzwitch in #215
• Benchmark general joins against Polars at large row counts by @randyzwitch in #216
• Use direct membership for bounded Int64 semi and anti joins by @randyzwitch in #230
• Parallelize membership probes and select join rows by chunk by @randyzwitch in #231
• Parallelize right joins and avoid unused row index by @randyzwitch in #233
• Index high-cardinality joins by original right rows by @randyzwitch in #234
• Pack join hash entries and specialize Int64 probes by @randyzwitch in #235
• Consolidate related tests into fewer CI modules by @randyzwitch in #243
• Index regular Int64 join domains by compact stride by @randyzwitch in #244
• Skip row-chain traversal for unique bounded join keys by @randyzwitch in #245
• Build large bounded join indexes across key ranges by @randyzwitch in #246
• Gather ordered repeated join rows within source chunks by @randyzwitch in #247
• Compare contiguous string join keys by bytes by @randyzwitch in #248
• Use interval membership for consecutive Int64 join keys by @randyzwitch in #249
• Map ordered strided Int64 join keys directly to rows by @randyzwitch in #250
• Use bounded strided index for smaller right joins by @randyzwitch in #251
• Map repeated ordered join keys directly to row runs by @randyzwitch in #252
• Use open addressing for high-cardinality join indexes by @randyzwitch in #253
• Fuse bounded membership with aligned chunk filtering by @randyzwitch in #254
• Avoid source validity checks in null-extended gathers by @randyzwitch in #259
• Skip redundant validity checks in string join probes by @randyzwitch in #260
• Share null-extension bits across gather columns by @randyzwitch in #261
• Mix string partition hashes once by @randyzwitch in #262
• Probe nonnullable two-key Int64 joins directly by @randyzwitch in #263
• Keep native DuckDB join benchmark in oracle suite by @randyzwitch in #264
• Parallelize full outer join row construction by @randyzwitch in #265
• Read Parquet through a minimal Arrow build behind a swappable backend by @randyzwitch in #275
• Benchmark key layouts and sizes that miss the fast paths by @randyzwitch in #281
• Probe the hash index for semi and anti joins; sweep two row cutoffs by @randyzwitch in #282
• Add list and struct column types with split, explode, unnest and Arrow I/O by @randyzwitch in #283
• CI: build libdfparquet on Linux and require the Parquet tests to run by @randyzwitch in #284
• Add implode, as_struct, and struct columns as group, unique and join keys by @randyzwitch in #285
• Fix and verify the Parquet reader build on Apple Silicon by @randyzwitch in #286
• Measure join cutoffs and bound direct-address tables by bytes by @randyzwitch in #288
• Use progression joins for matching temporal grids by @randyzwitch in #289
• Calibrate bucket sort limits across rank widths and machines by @randyzwitch in #290
• Add concise tuple schemas to read_csv by @randyzwitch in #293
• Stream Parquet row groups into chunked columns by @randyzwitch in #291
• Calibrate worker limits on Threadripper and Apple M1 by @randyzwitch in #292
• Parallelize Arrow import with measured serial fallbacks by @randyzwitch in #295
• Execute lazy plans as bounded ordered batch pipelines by @randyzwitch in #296
• Establish upstream join benchmarks and profile structural gaps by @randyzwitch in #297
• Reuse prepared join state across streaming batches by @randyzwitch in #298
• Choose the smaller join build input at measured cardinality cutoffs by @randyzwitch in #299
• Filter impossible join probes using exact integer key bounds by @randyzwitch in #300
• Count eligible inner joins without materializing matches by @randyzwitch in #301
• Borrow string-view byte blocks instead of retaining each read by @randyzwitch in #302
• Specialize scalar membership and string count probes by @randyzwitch in #303
• Use shared block hashing for long string keys by @randyzwitch in #305
• Use the shared Float64 power-of-ten table in strict parsing by @randyzwitch in #313
• Store the numeric DType inside DataType by @randyzwitch in #314
• Keep parsed DataTypes in expression nodes by @randyzwitch in #315
• Remove the GCD-stride mode from bounded Int64 join indexes by @randyzwitch in #316
• Drop equal-run repeat detection from progression joins by @randyzwitch in #317
• Replace inner-join count fusion with projection pushdown through joins by @randyzwitch in #318
• Aggregate worker row ranges with shared mergeable state by @randyzwitch in #319
• Validate join type once into an integer code by @randyzwitch in #320
• Keep selector kinds as integer codes by @randyzwitch in #321
• Benchmark against external suites with checked answers by @randyzwitch in #322
• Retire benchmarks the external suites supersede by @randyzwitch in #323
• Add arg_min, arg_max, mode, skew, kurtosis, corr and cov by @randyzwitch in #324
• Add describe, sample and value_counts (#220) by @randyzwitch in #325
• Merge streaming group-by states in batches, not per batch (#326) by @randyzwitch in #338
• Add decimal128 column type by @randyzwitch in #339
• Split streaming group-by state by key hash with a persistent key index (#326) by @randyzwitch in #340
• Return Float64 from decimal mean, as Polars does (#341) by @randyzwitch in #343
• Print decimal columns instead of aborting by @randyzwitch in #344
• Add interpolate, cut, and qcut expressions by @randyzwitch in #345
• Round decimal products, quotients and casts half to even (#342) by @randyzwitch in #346
• Keep decimal values through joins and integer casts by @randyzwitch in #347
• Turn on H2O q9 (corr) and a decimal PDS-H variant by @randyzwitch in #348
• Write an HTML benchmark report beside the Markdown one by @randyzwitch in #349
• Write comparisons and Boolean logic as packed bitmaps (#327) by @randyzwitch in #350
• Gather rows in bulk runs instead of one row at a time (#328) by @randyzwitch in #351
• Add time-zone-aware datetimes (#222) by @randyzwitch in #352
• Add binary columns (#120) by @randyzwitch in #353
• Rank numeric columns with one parallel sort per partition (#330) by @randyzwitch in #354
• Group by string keys without gathering or re-hashing them (#334) by @randyzwitch in #355
• Lazy joins on differently named keys (#329) by @randyzwitch in #356
• Write the PDS-H suite with the lazy API (#329) by @randyzwitch in #357
• Sort rows by packing their keys and row index into one integer (#331) by @randyzwitch in #358
• Select the first rows of a sort cut short by a slice (#332) by @randyzwitch in #359
• Scan ungrouped integer reductions and answer len() from the row count (#333) by @randyzwitch in #360
• Run test modules grouped into about ten programs by @randyzwitch in #361
• Count distinct values by hash partition (#336) by @randyzwitch in #362
• Fix the warnings CI reports by @randyzwitch in #363
• Capitalize a benchmark docstring summary by @randyzwitch in #364
• Select quantile positions instead of sorting, and finish groups in parallel (#337) by @randyzwitch in #365
• Concatenate join rows on every worker and stop copying them (#335) by @randyzwitch in #366
• Stream lazy group-bys over frames only when keys repeat; keep a group's first distinct number inline (#326) by @randyzwitch in #367
• Parse string casts straight into the target type, in parallel ranges (#149) by @randyzwitch in #369
• Add categorical columns: dictionary-encoded strings (#106) by @randyzwitch in #370
• Run categorical expressions once per dictionary value; count codes for n_unique (#106) by @randyzwitch in #371
• Allocate fully overwritten buffers without filling them first (#388) by @randyzwitch in #391
• Compare strings with literals and run string is_in over raw bytes on every worker (#373) by @randyzwitch in #392
• Filter fixed-width and Boolean columns straight from the mask's words (#376) by @randyzwitch in #393
• Gather string rows as views into the source's bytes (#375) by @randyzwitch in #394
• Fix the q18 and q1 slowdowns: share and deduplicate view buffers, build views directly, copy kept runs (#395) by @randyzwitch in #396
• Search strings with a SIMD byte filter, slice them as views, and add str.like (#374) by @randyzwitch in #397
• Stop dividing 128-bit integers on every decimal row (#385) by @randyzwitch in #398
• Run narrow-integer arithmetic in SIMD blocks and reduce computed columns directly (#384) by @randyzwitch in #399
• Number group keys through typed open-addressing tables instead of Dict (#380) by @randyzwitch in #400
• Read aggregation inputs through pointers and skip validity when there are no nulls (#382) by @randyzwitch in #401
• Let a single lazy join with a small left input build on it (#377) by @randyzwitch in #402
• Probe the Int64 hash join through pointers and move hashes into the index (#378) by @randyzwitch in #403
• Number over() partitions on every worker and write partitioned ids back in parallel (#387) by @randyzwitch in #404
• Run in-memory lazy group-bys eagerly when faster, and merge string keys in parallel (#381) by @randyzwitch in #405
• Write NaN as NaN in CSV so it reads back as a float (#368) by @randyzwitch in #406
• Make DATAFRAME_THREADS a limit on threads running jobs (#390) by @randyzwitch in #407
• Group several few-valued keys by worker ranges, not hash partitions by @randyzwitch in #408
• Build the deepest small-left join of a lazy chain on its small side (#377) by @randyzwitch in #409
• Read the decimal precision limit once per column, not per row by @randyzwitch in #410
• Run a lazy join eagerly when its left input comes from other joins and its right is large (#377) by @randyzwitch in #411
• Fill Boolean nulls and keep is_in's nulls on bitmaps, not per row by @randyzwitch in #412
• Index the decimal sum's pointers with unsafe_offset= by @randyzwitch in #413
• Write PDS-H's LIKE '%a%b%' filters with str.like by @randyzwitch in #414
• Add and multiply decimals without nulls through pointers, dividing by constant powers of ten by @randyzwitch in #415
• Keep every row without copying, and rank many small partitions with one sort by @randyzwitch in #416
• Number buckets of one numeric key by hash alone, without comparing rows by @randyzwitch in #417
• Stream large inputs through joins in larger batches by @randyzwitch in #418
• Keep a chunked series' first rechunk, and merge view chunks by block copy by @randyzwitch in #419
• Number small Int64 key domains through pointers by @randyzwitch in #420
• Hash the smaller side of semi and anti joins, deciding lazily from real sizes by @randyzwitch in #421
• Return all-valid comparisons without a validity bitmap by @randyzwitch in #422
• Compare string keys eight bytes at a time in hash buckets by @randyzwitch in #423
• Aggregate a moderate-cardinality Int64 key in place per worker range (#383) by @randyzwitch in #424
• Keep decimals at their declared width: Arrow decimal32 and decimal64, and Parquet's smallest decimals by @randyzwitch in #425
• Filter rank(ordinal).over(keys) <= k by keeping each partition's top k rows by @randyzwitch in #426
• Read grouped values at their source rows instead of gathering them per bucket by @randyzwitch in #427
• Order lazy inner-join chains by how much each input narrows its table by @randyzwitch in #428
• Evaluate AND filters one part at a time on the rows earlier parts kept by @randyzwitch in #429
• Keep Parquet dictionary codes beside string columns and group by them by @randyzwitch in #430
• Compare decimals unmasked and combine byte-aligned masks whole bytes at a time by @randyzwitch in #431
• Filter aligned chunks one job each for any row-local predicate by @randyzwitch in #433
• Fix repeated column references in fused chunk windows by @randyzwitch in #443
• Bound grouped top-k memory and replace quadratic insertion by @randyzwitch in #444
• Reuse measured join inputs when join order stays unchanged by @randyzwitch in #445
• Preserve holdout validation and improve benchmark provenance and coverage by @randyzwitch in #446
• Keep supported aggregates indexed and sort general keys by row prefixes by @randyzwitch in #447
• Reuse owned key codes and dense lookup for composite grouping by @randyzwitch in #448
• Batch bounded cardinality sample hashing across key columns by @randyzwitch in #449
• Reuse cardinality preference across group-by strategy selection by @randyzwitch in #450
• Assign string window ranks from packed partition ordering by @randyzwitch in #451
• Hash Int64 join probes directly without a row-sized hash buffer by @randyzwitch in #452
• Feed grouped native sums and means directly by source interval by @randyzwitch in #454
• Cache bitmap reads and separate nullable reduction loops by @randyzwitch in #456
• Cache composite string buffers for exact bucket comparisons by @randyzwitch in #457
• Avoid periodic-key aliasing in hash-aggregation sampling by @randyzwitch in #458
• Sum grouped numeric columns of every width at their own type by @randyzwitch in #459
• Probe hash joins on integer keys of every width without a hash buffer by @randyzwitch in #460
• Keep Parquet dictionary codes through filters, takes, sorts and joins by @randyzwitch in #461
• Let grouped top-k workers scan for their groups instead of routing rows by @randyzwitch in #462
• Skip the bucket gather when no aggregate needs gathered columns by @randyzwitch in #463
• Make PDS-H a development suite and hold out TPC-DS instead by @randyzwitch in #467
• Support decimals in conditionals without otherwise, fill_null and coalesce by @randyzwitch in #468
• Start lazy inner-join chains from the input that saves the most hashing by @randyzwitch in #469
• Translate 13 more TPC-DS queries for this library and Polars by @randyzwitch in #470
• Add an interactive Mermaid explorer for benchmark query plans by @randyzwitch in #471
• Split the Linux CI test step into smaller programs to fit its memory by @randyzwitch in #475
• Fix the mean and distinct count of decimal columns outside group-bys by @randyzwitch in #472
• Cast numeric columns in one typed loop when every value fits by @randyzwitch in #474
• Translate 7 more TPC-DS queries for this library and Polars by @randyzwitch in #473
• Translate 5 more TPC-DS queries for this library and Polars by @randyzwitch in #476
• Count distinct values by group through one set of (group, value) pairs by @randyzwitch in #477
• Add a cross join to LazyFrame by @randyzwitch in #479
• Encode chunked Int64 keys in place and compact column windows in bulk by @randyzwitch in #478
• Select grouped quantiles within one buffer laid out in group order by @randyzwitch in #480
• Translate 7 more TPC-DS queries for this library and Polars by @randyzwitch in #481
• Run the Linux CI tests as two driver programs, one at a time by @randyzwitch in #482
• Collect lazy execution reports in benchmark traces and list imbalanced joins by @randyzwitch in #483
• Report observed execution counters per plan node, and take every eager join direct by @randyzwitch in #484
• Collect a many-group streamed group-by and reduce it once, eagerly by @randyzwitch in #492
• Select conditional branches with SIMD and reduce masked branches in place by @randyzwitch in #493
• Probe hash join slots through a byte of tag per slot, sixteen at a time by @randyzwitch in #494
• Evaluate expressions over dictionary-coded String columns on their codes by @randyzwitch in #495
• Do not stream a frame scan whose only consumer runs eagerly by @randyzwitch in #496
• Drop the micro-benchmark and package smoke steps from pull request CI by @randyzwitch in #497
• Compare composite group keys through typed views; push key filters below group_by by @randyzwitch in #498
• Bound a streamed top-k by its k-th first key, like DuckDB's Top-N filter by @randyzwitch in #499
• Add geospatial storage and GeoJSON, GeoArrow, and GeoParquet interchange by @randyzwitch in #500
• Preserve Arrow field metadata and unknown extensions by @randyzwitch in #501
• Add WKB geometry type, CRS and bounding-box accessors by @randyzwitch in #502
• Pass join key sets sideways; single-pass integer is_in; small-left semi joins first by @randyzwitch in #503
• Complete GeoParquet interoperability with GeoPandas and DuckDB by @randyzwitch in #504
• Do not trade stacked filters over a join by @randyzwitch in #505
• Add eager and lazy as-of joins by @randyzwitch in #506
• Key sets from join-chain inputs, per key, through renames, and into a grouping left side by @randyzwitch in #508
• Answer filters on ascending columns by binary search by @randyzwitch in #509
• Add calendar-aware time windows and rolling std/var by @randyzwitch in #507
• Time each node in the execution report; skip restoring join order nobody reads by @randyzwitch in #510
• Read explicit categorical CSV fields directly into dictionary codes by @randyzwitch in #516
• Add native shapefile reader with typed DBF attributes by @randyzwitch in #515
• Whole-row unique as a parallel group_by; encode short string keys in registers by @randyzwitch in #511
• Balance streamed batches: four a worker, claimed as workers free up by @randyzwitch in #512
• Keep the busy and wall times from execution reports in results.json by @randyzwitch in #513
• Run each query's parallel work on one crew of threads by @randyzwitch in #514
• Build hash join indexes from hashes and keys in bucket order by @randyzwitch in #517
• Count distinct values per group by segments, not one bucket-wide set by @randyzwitch in #518
• Evaluate a filter's column-versus-constant comparisons in one pass by @randyzwitch in #519
• Move example builds off ordinary PRs and record CI costs by @randyzwitch in #520
• Reuse correctness binaries and skip benchmark work in answer checks by @randyzwitch in #521
• Copy only the columns read after a streamed filter by @randyzwitch in #523
• Decide a streamed group-by's collection from the rows held, not one batch by @randyzwitch in #530
• Execute a subplan used in two places once by @randyzwitch in #531
• Sum a column plus a constant from the column's own sum and extremes by @randyzwitch in #532
• Fuse comparisons on columns with nulls by @randyzwitch in #533
• Record the CPU frequency governor in benchmark provenance by @randyzwitch in #534
• Put hash tables and partition buffers on transparent huge pages by @randyzwitch in #536
• Share the filter selection across its jobs; direct-lookup codes for narrow integer keys by @randyzwitch in #537
• Design: thread-owned pipelines over morsels (executor redesign) by @randyzwitch in #539
• Stage 1 follow-ups: filter path selection, morsel size and alignment by @randyzwitch in #548
• Add provider-owned automatic execution dispatch by @randyzwitch in #551
• Fix logical-plan assertions for automatic engine diagnostics by @randyzwitch in #552
• Group rows in the pipeline's sink: per-worker tables or one late gather by @randyzwitch in #553
• Print the plan alone from explain(); the ENGINE line with engine="auto" by @randyzwitch in #554
• Run joins as probe steps of the thread-owned pipeline by @randyzwitch in #555
• Divide the eager group-by at the partition floor, not the scan floor by @randyzwitch in #556
• Keep the k best rows in the pipeline's sink by @randyzwitch in #557
• Load Parquet decimals at the smallest width their precision allows by @randyzwitch in #558
• Cast decimals to Float64 through the pointer under a mask too by @randyzwitch in #559
• Widen decimal32 operands to 64 bits before the decimal kernels by @randyzwitch in #560
• Share accelerator row planning and semantic capability checks by @randyzwitch in #561
• Fold a worker's top-k candidates by the sort keys alone by @randyzwitch in #562
• Encode chunked categorical keys by their codes in place by @randyzwitch in #563
• Add optional native Apple Metal accelerator runtime by @randyzwitch in #565
• Add LazyFrame.concat (UNION ALL) as a plan node by @randyzwitch in #567
• Fuse Metal projections and shrink intermediate storage by @randyzwitch in #566
• Measure Apple Metal DataFrame execution with retained M1 results by @randyzwitch in #568
• feat(metal): support all native integer widths by @randyzwitch in #569
• Keep a filter above a with_columns that holds a window by @randyzwitch in #572
• Running sums of decimal columns, exact at their scale by @randyzwitch in #574
• Add native Metal mixed numeric expressions and casts by @randyzwitch in #570
• feat(metal): preserve native types in extrema reductions by @randyzwitch in #571
• Translate 34 more TPC-DS queries (#564 phases 1 and 2) by @randyzwitch in #575
• Group build-side string keys by their codes in the pipeline by @randyzwitch in #580
• feat(metal): retain temporal and decimal column metadata by @randyzwitch in #573
• feat(metal): execute masked conditionals and integer operators by @randyzwitch in #576
• feat(metal): sort resident rows by multiple stable keys by @randyzwitch in #577
• feat(metal): aggregate resident rows by native typed keys by @randyzwitch in #578
• feat(metal): reduce rows with picks, logic, and arg extrema by @randyzwitch in #579
• Execute native Metal joins and joined lazy queries by @randyzwitch in #581
• Group String keys exactly in the pipeline's insert tables by @randyzwitch in #585
• Check join key progressions over raw values by @randyzwitch in #586
• Drop all-valid bitmaps when importing Arrow arrays by @randyzwitch in #587
• Keep right join keys for every join kind when coalesce=False by @randyzwitch in #588
• Update to Mojo 1.2.0.dev2026100905 by @randyzwitch in #591
• Join name lists losslessly so any column name round-trips by @randyzwitch in #589
• Extend Metal residency, typed algorithms, dictionaries, and analytical placement by @randyzwitch in #590
• Run cumulative windows in one pass and number partitions directly by @randyzwitch in #592
• Add forced-Metal correctness coverage for unchanged external queries by @randyzwitch in #599
• Preserve non-coalesced keys in accelerator joins by @randyzwitch in #593
• Extend Metal dictionary string operations and exact integer membership by @randyzwitch in #594
• Keep distinct rows, slices and UNION ALL resident on Metal by @randyzwitch in #595
• Preserve opaque Float64 payloads through Metal row plans and joins by @randyzwitch in #596
• Execute provably bounded Int64 and UInt64 sums on Metal by @randyzwitch in #597
• Compose Metal aggregates and add distinct counts and temporal fields by @randyzwitch in #598
• Rebuild test suite around maintained dataframe contracts by @randyzwitch in #624
• Fix macOS native Arrow consumer allocator isolation by @randyzwitch in #634
• Sink grouped pipelines into a radix-partitioned aggregate table by @randyzwitch in #641
• Compile fatal contract tests once and preserve isolated crash checks by @randyzwitch in #638
• Keep probe columns under a selection through pipeline joins by @randyzwitch in #648
• Keep decimal sums and means inline in the aggregate table by @randyzwitch in #650
• Share subplans read whole by several joins or aggregations by @randyzwitch in #653
• Extend Metal numeric execution, windows, and zero-column sources by @randyzwitch in #656
• Preserve full-domain Gregorian Date fields on CPU and Metal by @randyzwitch in #657
• Add Apple Metal duration totals across the full tick domain by @randyzwitch in #658
• Number window partitions with the aggregate table's hashing by @randyzwitch in #659
• Match multi-column join keys a column at a time by @randyzwitch in #667
• Add native median and quantiles; preserve UInt64 and decimal samples by @randyzwitch in #660
• Add native scalar-only selections with one-row results by @randyzwitch in #661
• Add checked fixed temporal offsets to Apple Metal by @randyzwitch in #662
• Create a query's worker threads on first parallel use by @randyzwitch in #663
• Upload chunked Apple Metal inputs without host rechunking by @randyzwitch in #664
• Add calendar month and year truncation to Apple Metal by @randyzwitch in #665
• Add signed calendar offsets to Apple Metal by @randyzwitch in #666
• Add Metal exp/log and correct CPU exceptional logarithms and exponentials by @randyzwitch in #668
• Add linear and nearest interpolation to resident Metal windows by @randyzwitch in #669
• Add Metal coordinate interpolation and fix CPU chunked coordinate windows by @randyzwitch in #670
• Read execution reports that lead with a calibration tag by @randyzwitch in #680
• Finish aggregate tables without copying keys or sorting groups by @randyzwitch in #671
• Add floating power on Metal and correct CPU IEEE power semantics by @randyzwitch in #672
• Parse a temporal interval once per column by @randyzwitch in #673
• Support decimal duration arithmetic on CPU and Metal by @randyzwitch in #675
• Support UTC duration rolling and wider calendar day periods on Metal by @randyzwitch in #677
• Evaluate window partitions contiguously, many at once by @randyzwitch in #679
• Keep a pipeline's selection through projections of bare columns by @randyzwitch in #683
• Clear every compiler warning in the library and its tests by @randyzwitch in #685
• Make Arrow string views the one in-memory layout for text by @randyzwitch in #686
• Migrate Metal execution into Mojo and retire native runtime by @randyzwitch in #684
• Keep plan schemas while optimizer passes read them by @randyzwitch in #687
• Run windows in one pass for decimals and share partition layouts by @randyzwitch in #692
• Run full joins through the direct hash route by @randyzwitch in #695
• Cast integers to decimals in one typed pass; multiply small decimals in 64 bits by @randyzwitch in #694
• Push filters through renames, union inputs and shared subplans by @randyzwitch in #688
• Filter the other input of an equi-join by a filter on its keys by @randyzwitch in #689
• Gather string views without per-row appends, prefetching ahead by @randyzwitch in #693
• Spread one-chunk pipeline sources over the workers by @randyzwitch in #696
• Extend Mojo GPU Lists, ordered execution and QCUT by @randyzwitch in #699
• Narrow pipeline morsels by each join build's key range by @randyzwitch in #690
• Clear the compiler warnings in the Mojo GPU modules by @randyzwitch in #703
• Give decimal arithmetic DuckDB's result types by @randyzwitch in #702
• Run a unique over every column as the pipeline's aggregate table by @randyzwitch in #691
• Let freed crew threads join a job list already running by @randyzwitch in #704
• Ask only moved filters' neighbours again; derive empty schemas from types by @randyzwitch in #697
• Join unique integer keys over a small span through slots (perfect hash) by @randyzwitch in #698
• Run a union's inputs at once by @randyzwitch in #705
• Remove an unused reference in the Mojo GPU source projection by @randyzwitch in #707
• Build on a small left inside the stream when its keys index directly by @randyzwitch in #700
• Run a join's computed inputs, and a stream's breaker inputs, at once by @randyzwitch in #708
• Filter aggregates built on by the probe side's keys (semi-join reduction) by @randyzwitch in #709
• Validate Metal on a local Apple silicon Mac and retire the CI job by @randyzwitch in #710
• Feed a filtered morsel to the aggregate table in place by @randyzwitch in #711
• Evaluate when(c).then(x) without an otherwise by validity, not choice by @randyzwitch in #712
• Send few-valued keys that form many groups to the aggregate table by @randyzwitch in #713
• Sort a string key whose values repeat by the values' ranks by @randyzwitch in #714
• Keep each group's top k rows in the aggregate table by @randyzwitch in #715
• Run the aggregate table's insert loops without a selection check when none is given by @randyzwitch in #716
• Reduce aggregates built on by several join keys to the probe side's key tuples by @randyzwitch in #717
• Pack SIMD comparison lanes into bits with one move-mask by @randyzwitch in #718
• Join a nested inner join chain's tables from the one supplying the keys by @randyzwitch in #719
• Reject absent integer probe keys with a bitmap of the build keys by @randyzwitch in #721
• Build on the smaller join side under distinct counts, judging their batches by key-value pairs by @randyzwitch in #720
• Revert join-tree flattening (#719) by @randyzwitch in #723
• Pack Boolean validity lists into bits eight lanes at a time by @randyzwitch in #724
• Spread the aggregate table's keep flags a mask byte at a time by @randyzwitch in #727
• Add Mojo GPU source List EXPLODE and align checked Decimal execution with CPU by @randyzwitch in #726
• Move file-format IO into a dataframe/io subpackage by @randyzwitch in #728
• Order join graphs by estimated cost, after DuckDB's join order optimizer by @randyzwitch in #725
• Fix UTF-8 temporal format literals and escaped percent by @randyzwitch in #730
• Join a crew posting before taking a budget slot, so retire waits for every slot by @randyzwitch in #729
• Match Apple GPU moment updates to CPU rounding and NaN payloads by @randyzwitch in #731
• Preserve checked integer producer errors before GPU consumers by @randyzwitch in #732
• Declare the GPU explode's output row count without an unused initial value by @randyzwitch in #734
• Build a stream's join on its left side when that side is the smaller by @randyzwitch in #733
• Match CPU order for serial grouped floating reductions on Apple Silicon by @randyzwitch in #735
• Preserve eager CUT/QCUT errors and widened integer SUM fusion by @randyzwitch in #737
• Pack multi-column row keys into one 64-bit key per numbering pass by @randyzwitch in #736
• Fix main: split the GPU capability requirement fused by the #735/#737 merge by @randyzwitch in #744
• Encode subset uniques and key counts one hash bucket at a time by @randyzwitch in #738
• Retire benchmark programs that predate the external suites by @randyzwitch in #743
• Fix merged backend capability policy constructor by @randyzwitch in #739
• Run IANA timezone row mutations with Mojo kernels on Apple Silicon by @randyzwitch in #741
• Tidy .gitignore and remove CHANGELOG.md by @randyzwitch in #750
• Narrow filter rows in place with selection trees, and read null tests from bitmaps by @randyzwitch in #745
• Move benchmark scripts into benchmarks/ and drop obsolete scripts by @randyzwitch in #746
• Remove retired dev notes and one-off benchmark results from docs by @randyzwitch in #748
• Build eager joins on the side of fewer rows times columns by @randyzwitch in #747
• Support direct grouped primitive Lists on Apple Silicon by @randyzwitch in #742
• Remove the unused legacy dfq_read_parquet entry point by @randyzwitch in #749
• Support IANA row producers before resident consumers on Apple Silicon by @randyzwitch in #751
• Preserve categorical code identity and dictionary owners on Apple GPU by @randyzwitch in #740
• Evaluate null-otherwise conditionals by validity when nothing can raise by @randyzwitch in #752
• Remove dead code found by two read-only scans by @randyzwitch in #753
• Keep primitive grouped Lists resident through downstream GPU consumers by @randyzwitch in #754
• Support partitioned cut and qcut on Apple GPU by @randyzwitch in #756
• Make TPC-DS a development suite; ClickBench stays held out by @randyzwitch in #757
• Derive Apple query coverage reports from verified receipts by @randyzwitch in #758
• Preserve typed grouped keys through resident GPU explode by @randyzwitch in #755
• Keep DuckDB's distinct-count statistics for loaded tables and order joins by them by @randyzwitch in #759
• Estimate joins the order cannot reorder instead of running them by @randyzwitch in #760
• Compose strict cast row producers with Apple GPU consumers by @randyzwitch in #761
• Cut the fixed cost of making, slicing and framing columns by @randyzwitch in #762
• Preserve eager checked Decimal SUM feed order on Apple GPU by @randyzwitch in #763
• Compose checked integer SUM with Apple GPU consumers by @randyzwitch in #764
• Diagnose complete Apple GPU query comparison differences by @randyzwitch in #765
• Maintain Apple GPU scalar floating SUM and MEAN contracts by @randyzwitch in #766
• Maintain Apple GPU grouped floating SUM and MEAN contracts by @randyzwitch in #767
• Admit checked scalar SUM with infallible Float64 mean siblings by @randyzwitch in #768
• Export complete typed benchmark answers for independent diagnostics by @randyzwitch in #769
• Force staged scalar aggregate materializations through Metal by @randyzwitch in #770
• Keep suite policy entries as markers in the test modules they describe by @randyzwitch in #784
• Exercise populated Apple query results against independent oracles by @randyzwitch in #771
• Fold casts of constants into typed literals while optimizing by @randyzwitch in #772
• Keep checked integer rank operands on the GPU through consumers by @randyzwitch in #773
• Restore a smaller-build join's left-major order on several workers by @randyzwitch in #774
• Observe checked Decimal GROUP operands before consumers by @randyzwitch in #775
• Compose scalar Float64 SUM over checked Decimal operands by @randyzwitch in #776
• Preserve coded String keys through checked Decimal GROUP by @randyzwitch in #777
• Observe checked Decimal operands before grouped Float64 SUM and MEAN by @randyzwitch in #778
• Observe bounded Decimal row producers containing Date YEAR by @randyzwitch in #779
• Prove same-scale Decimal selections before grouped consumers by @randyzwitch in #780
• Allow proven safe arithmetic after observed Decimal group sums by @randyzwitch in #781
• Size aggregate partitions by workers and finish one chunk per worker by @randyzwitch in #782
• Keep whether a key buffer is a progression with the buffer by @randyzwitch in #783
• Gather an unordered left build's rows with a gather that takes any order by @randyzwitch in #785
• Run every Metal contract on the Mac and fix the failures it finds by @randyzwitch in #786
• Build joins on progression-keyed relations instead of hashing the other side by @randyzwitch in #787
• Add lazy tail, reverse, shift and with_row_index by @randyzwitch in #788
• Add lazy group_by_dynamic and rolling by @randyzwitch in #812
• Count a join graph's computed relations at once by @randyzwitch in #814
• Put NaN first in descending sorts, as Polars and DuckDB do by @randyzwitch in #815
• Decide Metal data faults on the CPU instead of proving fault order per query shape by @randyzwitch in #819
• Add lazy unpivot, fill_null, drop_nulls, null_count, n_unique and duplicate flags by @randyzwitch in #813
• Leave sort ties unordered unless maintain_order=True by @randyzwitch in #817
• Run whole-input selections ahead of a join stream like other breakers by @randyzwitch in #818
• Run a stream's computed source through the morsel pipelines like a table by @randyzwitch in #821
• Add string functions that need no regex: titlecase, strip_prefix/suffix, find, split_exact and more by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/822
• Add lazy rename and cast, group_by on expressions, and group_by len and reductions by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/823
• Add lazy pivot with its output columns named up front by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/824
• Pass read_csv's options through scan_csv to every CSV read path by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/825
• Add lazy describe, clear, gather, top_k/bottom_k and diagonal and horizontal concat by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/826
• Sum, average and cum_sum Bool columns as Polars does by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/827
• Give a union's inputs the same columns after projection pruning by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/834
• Add log(base), log10, log1p, sign, cbrt, truncate, trig, arctan2, product and cum_prod by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/828
• Probe progression and perfect-hash joins through pointers by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/829
• Keep a null-otherwise branch's values whole in aggregate-table inputs by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/830
• Add sum/min/max/mean/any/all_horizontal by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/831
• Move the horizontal ops off the math functions' op numbers by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/848
• Keep repeated subplans whole so they are still shared by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/847
• Estimate filters between joins as join graph edges, as DuckDB does by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/838
• Add replace and replace_strict for recoding values by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/832
• Copy validity a run at a time when compacting a selection by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/833
• Add a linear-time regex engine in the RE2 / Rust regex family (#219) by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/835
• Merge chunks of numeric columns on every worker by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/836
• Use the regex engine in the .str() namespace, as Polars does (#219) by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/841
• Bound a join's large side by the small side's extreme for inequality conditions by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/837
• Reduce aggregates by probe keys through renames and filtering joins by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/839
• Take a probe key's span from its zone maps instead of scanning it by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/842
• Compute repeated subexpressions of aggregations and projections once by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/840
• Restore lazy.mojo's import of _int_bounds (main does not compile) by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/852
• Group small integer keys by perfect hash, as DuckDB does by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/843
• Leave a failing perfect hash aggregation to the batch-ordered routes by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/858
• Derive join and group_by schemas from their inputs instead of running them by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/845
• Measure complete CPU and Metal queries on Apple Silicon by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/849
• Render standalone Apple query coverage and performance reports by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/850
• Count a sampled filter's matches from its mask, on copied runs by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/846
• Accelerate small dictionary-coded String sums on Metal by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/851
• Prune captured columns before Metal dictionary staging by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/853
• Reuse retained dictionary codes during Metal preparation by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/854
• Upload only source columns into Metal row matrices by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/855
• Sum bounded pairs of String dimensions directly on Metal by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/856
• Share immutable device storage for resident projections by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/857
• Prune source payload before dictionary lookup lowering by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/859
• Keep wide payload out of GPU row-selection matrices by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/860
• Bring the TPC-DS coverage in docs/benchmarks.md up to 89 of 99 by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/879
• Compute a dense selection's with_columns without compacting the morsel by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/863
• Skip top-k morsels by their key's zone maps, as DuckDB's dynamic filter does by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/867
• Add bounded GPU equality hash partition construction by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/864
• Sum Int64 column statistics in Int128 lanes so they cannot wrap by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/880
• Skip the DFA passes for anchored capture searches, as RE2 does by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/882
• Keep rows together and reject mismatched infinities in the answer check by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/884
• Use equality hash partitions for general Metal grouped reducers by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/865
• Add separate Metal grouping development performance gates by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/866
• Establish an attributable Mojo Metal performance baseline by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/883
• Choose the in-memory lazy group-by's route by the routes' own tests (#608) by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/887
• Attribute Metal row execution and final export phases by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/885
• Finish aggregate table groups a vector of rows at a time, as DuckDB's scan does by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/886
• Add native Metal storage, reusable resources and bounded fused sums by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/888
• Archive verified M1 development-suite Metal baseline by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/895
• Archive paired Metal reduction, grouping and join performance evidence by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/889
• Document Metal replacement coverage and retirement gates by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/890
• Prune repeated subplans to the union of their readers' needs by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/891
• Find repeated subplans by hash and shallow comparison, as Polars does by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/892
• Keep node schemas across the optimizer passes and renumbering by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/893
• Reuse Metal device service across resident DAG queries by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/894
• Update mojo/max to newest nightly by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/906
• Update benchmark suite to Polars 2.0 by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/907
• Update to DuckDB 2.0 by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/910
• Use stable radix passes and tiled merges for Metal sorting by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/896
• Audit Metal answers with explicit numerical and selection policies by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/903
• Fix Metal detection crash under --target-accelerator=unknown by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/912
• Fix DuckDB benchmark reference workers by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/911
• Remove redundant borrowed-value transfers by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/909
• Preserve native typed columns through gathers and Float32 join inputs by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/897
• Gather native join payloads through shared device row maps by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/898
• Select bounded top-k runs before gathering sorted payloads by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/900
• Add native single-key Mojo GPU hash joins by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/901
• Parallelize grouped integer reductions with bounded partial states by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/902
• Build large Metal concat dictionaries from observed GPU pairs by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/904
• Delete one-off probe programs from tests/tooling by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/917
• Check native hash join memory at exact sizes instead of falling back by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/915
• Run benchmark-relevant kernel and suite checks from the Mac script by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/916
• Keep benchmark evidence out of the tree by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/918
• Remove ineffective ownership transfers by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/919
• Split docs into user docs and design notes by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/921
• Keep only runnable programs under tests/ by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/920
• Remove unused assignment initializers by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/922
• Restore the probe programs the Mac script runs by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/927
• Run contract areas as parallel CI jobs by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/923
• Specify the accelerator contract and check the core/backend boundary by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/925
• Add an on-demand CPU benchmark checkpoint report by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/928
• Shrink the accelerator provider to what the core calls and move planner types into core by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/930
• Add an on-demand Apple CPU/Metal checkpoint report by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/929
• Add temporal constructors, reporting fields, and calendar helpers by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/932
• Move the Metal backend into its own dataframe_metal package by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/933
• Remove temporal row boxing and index business-calendar holidays by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/934
• Update to Mojo 1.2 by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/936
• Fix Metal contracts: radix sort ignored categorical ranks, selectors hit the coded sum, coded sums merged labels, dates wrapped by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/942
• Admit float reassociation in the exact CPU/Metal receipt and record the match kind by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/938
• Square floats by multiplication instead of calling the platform pow by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/940
• min, max, cum_min, cum_max and rolling extrema skip NaN, as in Polars by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/937
• Pin the CPU Date overflow boundary behind the Metal calendar contracts by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/941
• Group categorical keys by label: equal labels are equal keys on every route by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/944
• Add Unicode regex properties, capture structs, and regex selectors by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/945
• Release v0.3.0 by @randyzwitch in https://github.com/randyzwitch/dataframe_mojo/pull/946

Full Changelog: v0.2.0...v0.3.0