Open alamb opened 2 weeks ago
I think we should aim for a first "milestone" of showing improvements for some clickbench queries
Will we completely change StringArray to StringViewArray in Datafusion? While I try to utilize StringViewArray in #10976 , I found there is schema mismatched issue UTF8 vs UTF8View. To avoid converting StringViewArray to StringArray, we might need to change the schema to UTF8View overall from logical plan to physical plan. If we need to keep both String and StringView, then we need to think about how to deal with the conversion between these two types.
A more concrete example is
statement ok
create table t(a int, b varchar, c int) as values (1, 'a', 3), (2, 'c', 1), (1, 'c', 2), (1, 'a', 4);
We have the string column b
as StringArray and DataType::Utf8 now. Should we convert it to StringViewArray and DataType::Utf8View?
If not, if we somehow want to utilize StringViewArray, how do we minize the cost of conversion between String and StringView?
It seems Polars completely refactor their String to StringView 🤔
Will we completely change StringArray to StringViewArray in Datafusion?
I think since they are two separate types in Arrow we couldn't fully switch to StringView the way polars could as it controls the whole stack. Users could still feed DataFusion StringViewArray from custom TableProviders and would expect StringView at the output.
However what I think we could do is internally to DataFusion (e.g. within the plan, before the final output) is use StringView in the batches that flow through intermediate nodes in the plan.
I found there is schema mismatched issue UTF8 vs UTF8View. To avoid converting StringViewArray to StringArray, we might need to change the schema to UTF8View
Indeed, As you point out, I don't think we can transparently switch to using StringView -- instead we would have to start encoding information in the plans about the new types.
I wonder if we could have a new logical optimzier pass that tried to annotate operations that support it to use StringView in their schema rather than String. Then the ExecutionPlans would know if they were supposed to generate StringView as output or the more traditional StringArray 🤔
Here is an idea of one place to start: https://github.com/apache/datafusion/issues/9403#issuecomment-2178347730
I think @XiangpengHao is looking into another place to use StringView which is https://github.com/apache/datafusion/issues/10921 -- where we have a similar idea to use StringView in some sub portion of the plan. Here is more info about the optimizer pass idea: https://github.com/apache/datafusion/issues/10921#issuecomment-2178364966
I think this project is going pretty well :bowtie:
We are at the point of starting to implement some basic functions using StringView.
Is your feature request related to a problem or challenge?
StringView
/BinaryView
were added to the Arrow format that make it more suitable for certain types of operations on strings. Specifically when filtering with string data, creating the outputStringArray
requires copying the strings to a new, packed binary buffer.GenericByteViewArrary
was designed to solve this limitation and the arrow-rs implementation, tracked by https://github.com/apache/arrow-rs/issues/5374, is now complete enough to start adding support into DataFusionI think we can improve performance in certain cases by using StringView (this is also described in more details in the Pola.rs blog post)
StringViewArray
/BinaryViewArray
rather thanStringArray
/BinaryArray
saves a copy (and @ariesdevil is quite close to having it integrated into the parquet reader https://github.com/apache/arrow-rs/issues/5530)substr(url, 4) = 'http'
) as the intermediate result ofsubstr
can be called without copying string valuesDescribe the solution you'd like
I would like to support StringView / BinaryView support in DataFusion.
While my primary usecase is for reading data from parquet, I think teaching DataFusion to use
StringView
(at least as intermediate values when evaluating expressions may help significantlyDevelopment branch:
string-view
Since this feature requires upstream arrow-rs support https://github.com/apache/arrow-rs/issues/5374 that is not yet released we plan to do development on a
string-view
feature branch:https://github.com/apache/datafusion/tree/string-view
Task List
Here are some high level tasks (I can help flesh these out for anyone who is interested in helping)
Utf8View
->Utf8
/BinaryView
->Binary
for compatibilityDescribe alternatives you've considered
No response
Additional context
Polars implemented it recently in rust so that can serve as a motivation Blog Post https://pola.rs/posts/polars-string-type/ https://twitter.com/RitchieVink/status/1749466861069115790
Facebook/Velox's take: https://engineering.fb.com/2024/02/20/developer-tools/velox-apache-arrow-15-composable-data-management/
Related PRs: https://github.com/pola-rs/polars/pull/13748 https://github.com/pola-rs/polars/pull/13839 https://github.com/pola-rs/polars/pull/13489