GH-26818: [C++][Python] Preserve order when writing dataset multi-threaded #44470
Conversation
|
|
github-actions
bot
added
Component: C++
Component: Python
awaiting review
EnricoMi
force-pushed
the
preserve-order-2
branch
2 times, most recently
from
9ca8c76 to
28cb588
Compare
|
This pull request seems to functionally overlap with this one. Some changes are almost exactly the same. Ordering of data is kept in threaded execution with use of batch index. Can you check whether it fixes your use case also? |
EnricoMi
force-pushed
the
preserve-order-2
branch
from
28cb588 to
0377ef9
Compare
| @@ -103,8 +103,8 @@ class ARROW_ACERO_EXPORT SourceNodeOptions : public ExecNodeOptions { | |||
| std::shared_ptr<Schema> output_schema; | |||
| /// \brief an asynchronous stream of batches ending with std::nullopt | |||
| std::function<Future<std::optional<ExecBatch>>()> generator; | |||
|
|
|||
| Ordering ordering = Ordering::Unordered(); | |||
There was a problem hiding this comment.
The constructor has a default value for ordering and initializes ordering with the value given to the constructor. No point for another default value here, I think.
github-actions
bot
added
awaiting committer review
EnricoMi
force-pushed
the
preserve-order-2
branch
from
2776b34 to
6024066
Compare
| acero::ConsumingSinkNodeOptions{ | ||
| std::move(consumer), | ||
| {}, | ||
| /*sequence_output=*/write_options.preserve_order})); |
There was a problem hiding this comment.
Right, looks like this requires some refactoring, since TeeNode is not used for writing datasets I'd leave this to a separate PR.
|
Since you are fixing dataset write ordering I think this check never fires. It should be moved to InsertBatch. |
EnricoMi
force-pushed
the
preserve-order-2
branch
from
6024066 to
697c378
Compare
|
@gitmodimo I think that refactoring should be done in a separate PR keeping this PR focused on fixing the issue. |
Rationale for this change
The order of rows in a dataset might be important for users and should be preserved when writing to a filesystem. With multi-threaded write, the order is currently not guaranteed,
What changes are included in this PR?
Preserving the dataset order of rows requires the
SourceNodeto useImplicitOrdering(this gives exec batches an index), and theConsumingSinkNodeto sequence exec batches (preserve order of batches by their index).User-facing changes:
preserve_ordertoFileSystemDatasetWriteOptionsDev-facing changes:
orderingtoSourceNodeOptionsimplicit_orderingtoScanNodeOptionsDefault behaviour is current behaviour.
Are these changes tested?
Unit tests have been added,
Are there any user-facing changes?
Users can set
FileSystemDatasetWriteOptions.preserve_order = true(C++) /arrow.dataset.write_dataset(..., preserve_order=True)(Python).