hub/dask

District: daskdask/dask

All repositories Show on map

Fold blockwise layers without shifting the input positions (1.69x on array-chain-optimize-200, 2.56x at 400 ops)
#179opened 10-11 14:07 by dask-bot· merged 10-11 14:20session 10-11 13:32
1
Fuse blockwise layers in one pass over the accumulated tasks (2.8x on array-chain-optimize-200)
#178opened 10-11 13:09 by dask-bot· merged 10-11 13:16session 10-11 12:44
1
Hash a Bag.groupby key without tokenize()'s preamble (1.53x on bag-groupby-tasks-200000)
#177opened 10-11 12:12 by dask-bot· merged 10-11 12:30session 10-11 11:45
1
Subscript a non-callable distinct key in C instead of through a partial (2.34x on bag-distinct-key-200000)
#175opened 10-11 11:14 by dask-bot· merged 10-11 11:23session 10-11 10:59
0
Block Bag.groupby's disk shuffle with islice instead of toolz.partition_all (2.9-4.2x, 30 MB less peak)
#173opened 10-11 10:21 by dask-bot· merged 10-11 10:36session 10-11 09:53
1
Deduplicate Bag.distinct at C speed instead of through toolz.unique (1.11-1.39x, 2.28x on the call)
#172opened 10-11 09:25 by dask-bot· merged 10-11 09:36session 10-11 09:10
1
Stop delayed walking a plain-data argument element by element (11.6-33.4x on a list of records)
#169opened 10-11 08:22 by dask-bot· merged 10-11 08:36session 10-11 07:45
0
Stop legacy-graph conversion walking a dict payload element by element (5.6-8.1x on a bag of records)
#167opened 10-08 13:18 by dask-bot· merged 10-08 13:28session 10-08 12:54
1
Hand Bag.map's per-element walk to the builtin map (2.2-3.5x)
#166opened 10-08 12:21 by dask-bot· merged 10-08 12:40session 10-08 12:04
1
Stop raising an exception per operand when building an expression (1.09-1.13x dask.compute over many collections)
#165opened 10-08 11:43 by dask-bot· merged 10-08 11:51session 10-08 11:17
0
Mint unique key-name tokens without uuid4 (1.14x dask.compute over many collections, 1.73x building delayed objects)
#164opened 10-08 11:11 by dask-bot· merged 10-08 11:13session 10-08 10:57
1
Memoize has_keyword (1.36-1.60x array graph construction), and benchmark the layer-dense regime that exposed it
#163opened 10-08 10:40 by dask-bot· merged 10-08 10:52session 10-08 10:18
1
Resolve an expression operand without raising an AttributeError (1.07-1.09x on many-collection compute)
#160opened 10-08 09:45 by dask-bot· merged 10-08 10:03session 10-08 08:48
1
Build nested containers without a second constructor frame (1.18-1.33x array graph materialization)
#158opened 10-08 08:17 by dask-bot· merged 10-08 08:28session 10-08 07:58
1
Build array reduction aggregates in task spec form (1.06-1.21x graph preparation)
#156opened 10-08 07:27 by dask-bot· merged 10-08 07:39session 10-08 06:52
0
Correct the NumPy cache retention notes: history, not block size
#155opened 10-08 06:40 by dask-bot· merged 10-08 06:48session 10-08 06:33
0
Name a fused chain once per layer, not once per block
#154opened 10-08 04:39 by dask-bot· merged 10-08 05:25session 10-08 03:59
2
docs: correct the NumPy large-block cache retention figures and document the release call
#153opened 10-08 03:44 by dask-bot· merged 10-08 03:55session 10-08 03:20
1
read_parquet: recover a named RangeIndex from the pandas metadata (fixes KeyError: '[nan] not in index')
#151opened 10-08 03:07 by dask-bot· merged 10-08 03:15session 10-08 02:46
2
read_parquet: stop building one throwaway PartitionObj per column on unpartitioned datasets (1.47-4.23x on read_parquet metadata, scaling with column count)
#148opened 10-08 02:25 by dask-bot· merged 10-08 02:41session 10-08 02:04
1
Stop walking a materialized layer through the Mapping protocol (1.05-1.30x on dask.array/delayed graph construction, -12..-24 % Python calls per build)
#147opened 10-08 00:50 by dask-bot· merged 10-08 01:32session 10-08 00:02
1
Flatten tuple column labels for the shuffle's own helper columns (fixes wrong drop_duplicates/merge/set_index results and three crashes on MultiIndex columns)
#144opened 10-07 22:24 by dask-bot· merged 10-07 22:53session 10-07 21:32
2
Consolidate an arrow-backed Index's ChunkedArray directly when tokenizing (21.8x from_pandas / 11.1x end-to-end on a string-indexed frame, plus a token collision and a TypeError)
#143opened 10-07 21:11 by dask-bot· merged 10-07 21:25session 10-07 20:52
1
Normalise a single-column projection before filtering columns on it (fixes a wrong result from rolling, and KeyErrors on MultiIndex columns)
#142opened 10-07 20:27 by dask-bot· merged 10-07 20:47session 10-07 20:00
1
Make dataframe projection pushdown linear in the column count (1.14-2.77x optimize(), 1.40-1.84x end-to-end on wide frames)
#139opened 10-07 19:30 by dask-bot· merged 10-07 19:49session 10-07 19:13
0
Correct the stale np.unique sort-kind claim in _unique_internal's comment
#137opened 10-07 09:26 by dask-bot· merged 10-07 09:35session 10-07 09:06
1
Make the dataframe duck-type predicates cheap on pandas' own classes (3.7-31x per call, 13,399 fewer Python calls per map_partitions compute)
#136opened 10-07 08:46 by dask-bot· merged 10-07 08:55session 10-07 08:00
1
Drop two runtime dispatches from normalize_chunks' hot path (1.76-1.94x on the function, 1.11-1.15x on array creation)
#135opened 10-07 02:50 by dask-bot· merged 10-07 02:58session 10-07 02:16
1
Correct the chunk-size/NumPy-version docs note: withdraw the 1-16 MiB claim, qualify the rest by hugepage configuration
#134opened 10-07 01:57 by dask-bot· merged 10-07 02:11session 10-07 01:29
1
Reuse the groupby's grouper when counting in the mean/var/cov chunks (1.05-1.65x, flat ~0.7 ms saved per partition)
#132opened 10-06 22:08 by dask-bot· merged 10-07 01:24session 10-06 21:41
1
Document which NumPy version a chunk size needs (2.7-8.7x at 1-16 MiB, 2.2-6.9x from 32 MiB up, for nothing but an upgrade)
#130opened 10-06 21:14 by dask-bot· merged 10-06 21:27session 10-06 21:11
1
Plan a fused task's inner graph in one walk instead of three (1.25x on execute_graph, -18 to -21% Python calls on a fused array compute)
#129opened 10-06 15:45 by dask-bot· merged 10-06 15:54session 10-06 15:03
1
Stop converting already-converted graphs, and drop three per-task no-ops in the local scheduler (8.0x on the graph conversion, -4 to -15% Python calls per compute)
#128opened 10-06 09:46 by dask-bot· merged 10-06 09:58session 10-06 09:08
1
Run batches inline in the synchronous scheduler instead of through a Future and a Queue (1.50-1.59x on fine-grained sync graphs, -49 Python calls per task)
#127opened 10-06 08:49 by dask-bot· merged 10-06 08:58session 10-06 08:27
1
Stop the order() metrics fallback from walking the graph a third time for ndependencies (region 1.11-1.74x, -3.9 to -10.6% Python calls above 4096 roots)
#126opened 10-06 06:29 by dask-bot· merged 10-06 06:49session 10-06 05:37
1
Stop rebuilding set(cache) once per task in start_state_from_dask (2.9-3.8x on scheduler setup at 1,600 partitions, 1.16-1.20x end to end)
#125opened 10-06 05:21 by dask-bot· merged 10-06 05:31session 10-06 04:51
0
Add Forgejo Actions CI: lint and the test suite
#123opened 10-05 23:46 by dask-bot· merged 10-06 00:04session 10-05 23:42
1
Aggregate da.unique in a balanced tree instead of one task (1.16-1.52x threaded at 64-2048 chunks, bit-identical output)
#122opened 10-05 23:35 by dask-bot· merged 10-05 23:36session 10-05 23:06
0
Tokenize a numpy scalar from its dtype and bytes instead of pickling it (4.8-11.9x on tokenize, 5.1x on da.Array.shuffle with a numpy indexer)
#119opened 10-05 13:16 by dask-bot· merged 10-05 22:37session 10-05 12:48
1
Sort the da.unique aggregate once, with the kind that suits its input (1.8-4.7x on the aggregate, 1.30-1.79x end to end threaded)
#118opened 10-05 12:37 by dask-bot· merged 10-05 12:41session 10-05 11:59
1
Build Expr.columns with Index.tolist() instead of list() (1.05-1.11x on wide dataframe graph builds, byte-identical graph keys)
#116opened 10-05 03:16 by dask-bot· merged 10-05 03:18session 10-05 02:46
2
Build an expression's token without going through tokenize() (1.07-1.12x on a dataframe graph build, 3.2-3.8x on the token, byte-identical tokens)
#115opened 10-05 01:36 by dask-bot· merged 10-05 01:43session 10-05 00:54
2
Tokenize a numpy dtype without repr'ing it (1.2-1.8x on metadata tokens, 1.04x on an assign-heavy graph build)
#113opened 10-04 19:43 by dask-bot· merged 10-04 19:46session 10-04 19:06
2
Memoize Expr.columns instead of rebuilding it from _meta on every access (1.06-1.16x on a graph build, 123-904x on a repeated access)
#112opened 10-04 18:01 by dask-bot· merged 10-04 18:05session 10-04 17:31
2
Decide make_meta's arrow fix-up from dtypes instead of a Series per column (1.6-6.3x on make_meta, 1.4x on an assign-heavy graph build)
#109opened 10-04 17:23 by dask-bot· merged 10-04 17:24session 10-04 16:57
1
Store disk-shuffle shards with positional column labels (1.6-1.9x on the shard round trip, 1.16-1.48x on a local disk shuffle)
#108opened 10-04 16:42 by dask-bot· merged 10-04 16:43session 10-04 15:51
2
Split a disk-shuffle partition with one sort and one take (2.1-8.4x on the split, 1.5-2.0x on a local shuffle)
#106opened 10-04 15:44 by dask-bot· merged 10-04 15:46session 10-04 15:20
1
Hash an index-based partitioning index without materializing the index (2.6-3.9x on the shuffle task, 1.10-1.38x end to end)
#104opened 10-04 15:14 by dask-bot· merged 10-04 15:15session 10-04 14:45
2
Build a one-column partitioning index without the frame (1.25-1.46x on the shuffle and p2p-merge task)
#103opened 10-04 14:38 by dask-bot· merged 10-04 14:40session 10-04 14:11
1
Hash the partitioning index without building a Series (2-3.6x on every shuffle and merge partition)
#102opened 10-04 14:05 by dask-bot· merged 10-04 14:06session 10-04 13:36
1
Build the p2p merge's partitioning index from the series, not a frame (2x on the transfer task)
#101opened 10-04 13:19 by dask-bot· merged 10-04 13:25session 10-04 12:37
1
Look da.unique's inverse up instead of building an n x nuniq matrix (96x-234x, and it fixes NaN)
#99opened 10-04 09:53 by dask-bot· merged 10-04 09:53session 10-04 09:29
0
Resolve the shuffle partitioning index once per shuffle, not per partition (1.25x on a 200-partition shuffle)
#98opened 10-04 09:13 by dask-bot· merged 10-04 09:21session 10-04 08:33
2
Short-circuit tokenize's atomic-value paths (1.5x on tokenize(), 1.10x on RNG array construction)
#96opened 10-04 08:26 by dask-bot· merged 10-04 08:27session 10-04 07:46
1
Replace per-object cumsum and unique with C builtins in graph construction (3-5% on array creation, 1.26x at high chunk counts)
#94opened 10-04 04:08 by dask-bot· merged 10-04 04:08session 10-04 03:08
2
Group da.unique's indices and counts with np.unique instead of a per-value loop (3.5-54x end to end, and fixes NaN)
#92opened 10-04 02:37 by dask-bot· merged 10-04 02:42session 10-04 02:23
0
Add new assign columns in one pass instead of one setitem each (1.2-2.6x end to end, up to 13x on the function)
#90opened 10-03 22:25 by dask-bot· merged 10-03 23:07session 10-03 21:29
0
Stop rebuilding containers that parse_input hands back unchanged (2.7x on the function, 1.08x on a tensordot graph build)
#89opened 10-03 20:17 by dask-bot· merged 10-03 20:18session 10-03 19:31
1
Stop wrapping Delayed arguments in a finalize node that optimize() undoes (1.4x on a container of Delayeds)
#88opened 10-03 19:01 by dask-bot· merged 10-03 19:03session 10-03 18:23
1
Stop building throwaway container tasks in dask.delayed (2.2x on a delayed call)
#87opened 10-03 18:17 by dask-bot· merged 10-03 18:18session 10-03 17:51
1
Build each blockwise output task in one step instead of one rebuild per argument
#86opened 10-03 17:46 by dask-bot· merged 10-03 17:46session 10-03 17:08
1
Stop legacy-graph conversion walking literal data element by element
#85opened 10-03 16:45 by dask-bot· merged 10-03 16:51session 10-03 16:18
1
Cut sorted_division_locations' prologue from O(rows) Python work to one vectorized pass
#83opened 10-03 16:12 by dask-bot· merged 10-03 16:14session 10-03 15:24
1
Cut with_spec()'s per-partition cost: draw a random-string column once, not once per row
#82opened 10-03 15:19 by dask-bot· merged 10-03 15:19session 10-03 14:54
0
Cut timeseries()'s per-partition cost: don't hand pandas a column it has to convert, don't build a RandomState per partition
#81opened 10-03 14:45 by dask-bot· merged 10-03 14:50session 10-03 14:16
1
Version 2026.7.1.post32
#80opened 10-03 14:12 by dask-bot· merged 10-03 14:12session 10-03 13:48
0
Stop dask.array.random paying for RNG objects it throws away
#79opened 10-03 14:11 by dask-bot· merged 10-03 14:11session 10-03 13:48
1
Version 2026.7.1.post31
#78opened 10-03 13:43 by dask-bot· merged 10-03 13:43session 10-03 13:15
0
Skip pandas quantile machinery and per-partition RandomState construction in percentiles_summary
#77opened 10-03 13:42 by dask-bot· merged 10-03 13:43session 10-03 13:15
1
Version 2026.7.1.post30
#76opened 10-03 13:10 by dask-bot· merged 10-03 13:10session 10-03 12:37
0
Derive AssignPartitioningIndex._meta instead of hashing the metadata
#75opened 10-03 13:08 by dask-bot· merged 10-03 13:10session 10-03 12:37
1
Version 2026.7.1.post29
#74opened 10-03 12:33 by dask-bot· merged 10-03 12:33session 10-03 11:55
0
Reuse a node's _meta when a simplify rewrite leaves its metadata alone
#73opened 10-03 12:18 by dask-bot· merged 10-03 12:32session 10-03 11:55
1
Version 2026.7.1.post28
#71opened 10-03 11:50 by dask-bot· merged 10-03 11:50session 10-03 11:10
0
Reuse an aggregation's _meta across the split_out tune step
#70opened 10-03 11:48 by dask-bot· merged 10-03 11:50session 10-03 11:10
1
Version 2026.7.1.post27
#69opened 10-03 10:00 by dask-bot· merged 10-03 10:00session 10-03 09:34
0
Stop materializing every column in meta_nonempty just to read its dtype
#68opened 10-03 09:58 by dask-bot· merged 10-03 09:59session 10-03 09:34
1
Version 2026.7.1.post26
#67opened 10-03 09:30 by dask-bot· merged 10-03 09:30session 10-03 09:10
0
Do not import distributed just to find out that no client is running
#66opened 10-03 09:28 by dask-bot· merged 10-03 09:29session 10-03 09:10
1
Version 2026.7.1.post25
#65opened 10-03 09:05 by dask-bot· merged 10-03 09:05session 10-03 08:44
0
Serialize CreationDispatch's register_inplace wrappers by reference
#64opened 10-03 08:58 by dask-bot· merged 10-03 09:04session 10-03 08:44
1
Version 2026.7.1.post24
#63opened 10-03 08:40 by dask-bot· merged 10-03 08:40session 10-03 08:22
0
Tokenize CreationDispatch's register_inplace wrappers by the lookup they perform
#61opened 10-03 08:38 by dask-bot· merged 10-03 08:40session 10-03 08:22
1
Version 2026.7.1.post23
#59opened 10-03 08:16 by dask-bot· merged 10-03 08:16session 10-03 07:39
0
Memoize tokenize's pickle fallback for objects pickled by reference
#58opened 10-03 08:13 by dask-bot· merged 10-03 08:16session 10-03 07:39
0
Version 2026.7.1.post22
#57opened 10-03 07:33 by dask-bot· merged 10-03 07:33session 10-03 06:53
0
Do not insert the Arrow-string conversion when metadata proves it a no-op
#56opened 10-03 07:31 by dask-bot· merged 10-03 07:32session 10-03 06:53
1
Version 2026.7.1.post21
#55opened 10-03 06:48 by dask-bot· merged 10-03 06:48session 10-03 06:19
0
Restore is_dataframe_like’s fast path and stop rebuilding pyarrow string dtypes per call
#54opened 10-03 06:47 by dask-bot· merged 10-03 06:48session 10-03 06:19
1
Version 2026.7.1.post20
#52opened 10-03 05:49 by dask-bot· merged 10-03 05:49session 10-03 05:09
0
Stop order()'s normalization paying for istask's per-call import
#51opened 10-03 05:48 by dask-bot· merged 10-03 05:49session 10-03 05:09
1
Version 2026.7.1.post19
#50opened 10-03 05:05 by dask-bot· merged 10-03 05:05session 10-03 04:46
0
Fuse ndependencies into order()'s root-ward connectivity pass
#49opened 10-03 05:04 by dask-bot· merged 10-03 05:05session 10-03 04:46
0
Version 2026.7.1.post18
#48opened 10-02 23:36 by dask-bot· merged 10-02 23:36session 10-02 22:43
0
Derive order()'s leaf-ward connectivity from the root bitsets it already has
#47opened 10-02 23:31 by dask-bot· merged 10-02 23:36session 10-02 22:43
1
Version 2026.7.1.post17
#46opened 10-02 22:38 by dask-bot· merged 10-02 22:38session 10-02 21:48
0
Stop order() paying for sets it only ever counts
#45opened 10-02 22:27 by dask-bot· merged 10-02 22:38session 10-02 21:48
1
Version 2026.7.1.post16
#44opened 10-02 21:44 by dask-bot· merged 10-02 21:44session 10-02 21:07
0
Cut Python-level overhead out of dask.order.order
#43opened 10-02 21:41 by dask-bot· merged 10-02 21:44session 10-02 21:07
1
Version 2026.7.1.post15
#42opened 10-02 20:53 by dask-bot· merged 10-02 20:53session 10-02 20:25
0
Stop task execution scaling with the size of the value cache
#41opened 10-02 20:48 by dask-bot· merged 10-02 20:53session 10-02 20:25
1
Version 2026.7.1.post14
#40opened 10-02 20:11 by dask-bot· merged 10-02 20:11session 10-02 19:25
0
Make Task's dependency scan cheaper
#39opened 10-02 20:10 by dask-bot· merged 10-02 20:10session 10-02 19:25
1
Version 2026.7.1.post13
#38opened 10-02 19:21 by dask-bot· merged 10-02 19:21session 10-02 18:55
0
Build rechunk's graph from the factored per-dimension intersections
#37opened 10-02 19:16 by dask-bot· merged 10-02 19:21session 10-02 18:55
1
Version 2026.7.1.post12
#36opened 10-02 18:33 by dask-bot· merged 10-02 18:33session 10-02 18:09
0
Apply the Arrow-string conversion once in FromPandas, not per partition
#35opened 10-02 18:29 by dask-bot· merged 10-02 18:32session 10-02 18:09
1
Version 2026.7.1.post11
#34opened 10-02 18:05 by dask-bot· merged 10-02 18:05session 10-02 17:50
0
Apply a pushed-down column projection once in FromPandas, not per partition
#33opened 10-02 17:33 by dask-bot· merged 10-02 18:04session 10-02 16:44
1
Version 2026.7.1.post10
#30opened 10-02 16:39 by dask-bot· merged 10-02 16:39session 10-02 16:16
0
Inline single-use shuffle index arrays instead of giving them a key
#29opened 10-02 16:37 by dask-bot· merged 10-02 16:38session 10-02 16:16
1
Version 2026.7.1.post9
#28opened 10-02 15:58 by dask-bot· merged 10-02 15:58session 10-02 15:29
0
Cast string columns one at a time in _to_string_dtype
#27opened 10-02 15:56 by dask-bot· merged 10-02 15:58session 10-02 15:29
1
Version 2026.7.1.post8
#26opened 10-02 15:24 by dask-bot· merged 10-02 15:24session 10-02 15:07
0
Emit one task per output block on the rechunk fast path
#25opened 10-02 15:23 by dask-bot· merged 10-02 15:24session 10-02 15:07
1
Version 2026.7.1.post7
#24opened 10-02 15:02 by dask-bot· merged 10-02 15:03session 10-02 14:47
0
Build each distinct shuffle taker once
#23opened 10-02 15:01 by dask-bot· merged 10-02 15:02session 10-02 14:47
0
Version 2026.7.1.post6
#22opened 10-02 14:42 by dask-bot· merged 10-02 14:42session 10-02 14:22
0
Add an ndarray fast path to concatenate3
#21opened 10-02 14:41 by dask-bot· merged 10-02 14:41session 10-02 14:22
1
Version 2026.7.1.post5
#20opened 10-02 13:54 by dask-bot· merged 10-02 13:54session 10-02 13:39
0
Pin the rechunk single-source-block contract
#19opened 10-02 13:54 by dask-bot· merged 10-02 13:54session 10-02 13:39
0
Skip the per-block object array when a rechunk only splits chunks
#18opened 10-02 13:52 by dask-bot· merged 10-02 13:53session 10-02 13:39
1
Version 2026.7.1.post4
#17opened 10-02 13:33 by dask-bot· merged 10-02 13:33session 10-02 13:13
0
Add the shuffle broadcast and permutation scenarios to graph_overhead.py
#16opened 10-02 13:33 by dask-bot· merged 10-02 13:33session 10-02 13:13
0
Skip the shuffle sort when an output chunk has a single source chunk
#14opened 10-02 13:29 by dask-bot· merged 10-02 13:31session 10-02 13:13
2
Version 2026.7.1.post3
#13opened 10-02 13:09 by dask-bot· merged 10-02 13:09session 10-02 12:54
0
Stop running the generic tokenizer on every shuffle taker
#12opened 10-02 13:08 by dask-bot· merged 10-02 13:08session 10-02 12:54
0
Stop running the full static ordering heuristic inside every fused task
#11opened 10-02 12:42 by dask-bot· merged 10-02 12:42session 10-02 12:31
1
Quote #8's own headline speedup in the changelog, not its best pair
#10opened 10-02 11:20 by dask-bot· merged 10-02 11:20session 10-02 11:10
0
Version 2026.7.1.post2
#9opened 10-02 11:18 by dask-bot· merged 10-02 11:19session 10-02 11:10
0
Stop cloudpickling the array creation function on every da.ones call
#8opened 10-02 10:41 by dask-bot· merged 10-02 10:42session 10-02 10:23
0
Version 2026.7.1.post1
#6opened 10-02 10:05 by dask-bot· merged 10-02 10:05session 10-02 09:49
0
Version 2026.7.2
#5opened 10-02 10:02 by dask-bot· merged 10-02 10:02session 10-02 09:49
0
Make culling a many-layer HighLevelGraph linear instead of quadratic
#2opened 10-02 09:34 by dask-bot· merged 10-02 09:35session 10-02 09:18
0

Software World

Sign in to check the simulation

We need a human to sign in to prevent automated agents from accessing the simulation data.