Repository navigation
Conversation
|
Some tests may be not passed, I will fix them tomorrow. And maybe I also need to add a test for |
| /// # } | ||
| /// ``` | ||
| fn union(&self, dataframe: Arc<dyn DataFrame>) -> Result<Arc<dyn DataFrame>>; | ||
| fn union( |
There was a problem hiding this comment.
I think a new method e.g. union_distinct would make better sense - having booleans in APIs hurts readability.
alternatively - what about df.union(df2)?.distinct()? and not exposing a new method?
| array_str += &*array_value_to_string(column, 1)?; | ||
| vec_array.push(column.clone()); | ||
| } | ||
| if vec_str.contains(&array_str) { |
There was a problem hiding this comment.
As the code is using a Vec with all rows + contains this scales badly, namely O(n^2) instead of O(n). Besides that, it probably has extremely high overhead in terms of string formatting and memory usage to keep all batches + rows converted to strings in memory. Formatting to string is also not very robust - we don't guarantee that two different values won't be formatted as the same string value.
I think a cleaner / efficient way to go now would be to reuse the current implementations we have to drive the execution and only change the query planning.
#998 (comment)
There was a problem hiding this comment.
For some inspiration, here is the current SELECT DISTINCT implementation:
https://github.com/apache/arrow-datafusion/blob/master/datafusion/src/sql/planner.rs#L777
There was a problem hiding this comment.
I agree with @Dandandan -- specifically I think you could make a plan that implemented SELECT x from foo UNION select x from bar by effectively creating the same plan as
SELECT distinct (select x from foo UNION ALL select x from bar)
You can see the plan that gets made by running EXPLAIN:
explain select distinct x from ( select 1 as x UNION ALL select 1 as x);
+---------------+------------------------------------------------------------------------------+
| plan_type | plan |
+---------------+------------------------------------------------------------------------------+
| logical_plan | Aggregate: groupBy=[[#x]], aggr=[[]] |
| | Union |
| | Projection: Int64(1) AS x |
| | EmptyRelation |
| | Projection: Int64(1) AS x |
| | EmptyRelation |
| physical_plan | HashAggregateExec: mode=FinalPartitioned, gby=[x@0 as x], aggr=[] |
| | CoalesceBatchesExec: target_batch_size=4096 |
| | RepartitionExec: partitioning=Hash([Column { name: "x", index: 0 }], 16) |
| | HashAggregateExec: mode=Partial, gby=[x@0 as x], aggr=[] |
| | UnionExec |
| | RepartitionExec: partitioning=RoundRobinBatch(16) |
| | ProjectionExec: expr=[1 as x] |
| | EmptyExec: produce_one_row=true |
| | RepartitionExec: partitioning=RoundRobinBatch(16) |
| | ProjectionExec: expr=[1 as x] |
| | EmptyExec: produce_one_row=true |
+---------------+------------------------------------------------------------------------------+```
(so use a `UnionExec` followed by `HashAggregateExec`)
| let stream = execute_stream(plan).await?; | ||
| common::collect(stream).await | ||
| let stream = execute_stream(plan.clone()).await?; | ||
| let any_plan = plan.as_any().downcast_ref::<UnionExec>(); |
There was a problem hiding this comment.
The code to execute the UnionExec (if changed) should be changed there.
I suggest to implement it using the plan we have. If a more efficient implementation could be implemented, I think the best way would be to put that in a new node - i.e. UnionDistinctExec and
| /// Execution metrics | ||
| metrics: ExecutionPlanMetricsSet, | ||
| /// Union ALL or Union | ||
| is_all: bool, |
There was a problem hiding this comment.
Union all and union distinct are quite different, so if we need to add / change implementations, I think it makes sense to add a new node instead, like UnionDistinctExec
|
@xudong963 thanks for opening this PR! I have some comments on the current approach / direction, let me know what you think. |
@Dandandan Dan Dan, I agree with your idea #998 (comment). After I finished the current approach, I also feel it badly. Thanks for your comments and guides, I will try the PR with #998 (comment). 💪🏻 |
No problem! Thanks for trying and looking forward to the next iteration 👍 |
|
I ended my disgusting 24-hour on-call. My one-week holiday is coming, I'll be absorbed in this PR! |
|
Close the PR, the new is #1068 |
* Attempt at caching Jstrings as GlobalRefs in a HashMap to reduce reallocations. I need to confirm 1) there's actually a performance benefit to this, and 2) these GlobalRefs are being released when I want them to be. * Minor refactor and added more docs. * Undo import reordering to reduce diff. * Docs. * Avoid get() by just cloning the Arc to globalref on insert. * Store jstring cache in ExecutionContext.
* pyo3 update required changes to deprecated interfaces * Substrait feature clippy updates * PyTuple was called twice * add -D warnings option --------- Co-authored-by: Tim Saucer <timsaucer@gmail.com>
Which issue does this PR close?
Closes #998
Rationale for this change
What changes are included in this PR?
Are there any user-facing changes?