2010YOUY01 opened a new pull request, #22921:
URL: https://github.com/apache/datafusion/pull/22921
## Which issue does this PR close?
<!--
We generally require a GitHub issue to be filed for all bug fixes and
enhancements and this helps us generate change logs for our releases. You can
link an issue to this PR using the GitHub syntax. For example `Closes #123`
indicates that this PR will close issue #123.
-->
- Closes #.
## Rationale for this change
<!--
Why are you proposing this change? If this is already explained clearly in
the issue then this section is not needed.
Explaining clearly why changes are proposed helps reviewers understand your
changes and offer better suggestions for fixes.
-->
Attempt to simplify the `approx_distinct` implementation, the existing
complexity is due to there is no generic API to calculate hash for a
`array.elem(i)`, so we have to implement specialization for many different
types like primitive/string/stringview, and bloated the code size.
This PR used a existing `create_hashes` for batched hashing that is
applicable to all array types, and it reduced 261 lines of code in
`approx_distinct.rs`
### Performance
<details>
<summary> Cargo bench result </summary>
```sh
cargo bench -p datafusion-functions-aggregate \
--bench approx_distinct \
-- --baseline main
```
```sh
nuplot not found, using plotters backend
Benchmarking approx_distinct i64 80% distinct: Collecting 100 samples in
estimated 5.0135 s (884k ite
approx_distinct i64 80% distinct
time: [5.6406 µs 5.6477 µs 5.6550 µs]
change: [−0.9680% −0.7111% −0.4639%] (p = 0.00 <
0.05)
Change within noise threshold.
Found 4 outliers among 100 measurements (4.00%)
2 (2.00%) low mild
2 (2.00%) high mild
Benchmarking approx_distinct utf8 short 80% distinct: Collecting 100 samples
in estimated 5.0360 s (3
approx_distinct utf8 short 80% distinct
time: [12.970 µs 12.977 µs 12.985 µs]
change: [+15.898% +16.116% +16.339%] (p = 0.00 <
0.05)
Performance has regressed.
Found 7 outliers among 100 measurements (7.00%)
2 (2.00%) low severe
1 (1.00%) low mild
4 (4.00%) high severe
Benchmarking approx_distinct utf8view short 80% distinct: Collecting 100
samples in estimated 5.0171
approx_distinct utf8view short 80% distinct
time: [8.7402 µs 8.7455 µs 8.7511 µs]
change: [+22.516% +22.703% +22.893%] (p = 0.00 <
0.05)
Performance has regressed.
Found 9 outliers among 100 measurements (9.00%)
1 (1.00%) low severe
5 (5.00%) high mild
3 (3.00%) high severe
Benchmarking approx_distinct utf8 long 80% distinct: Collecting 100 samples
in estimated 5.0120 s (26
approx_distinct utf8 long 80% distinct
time: [19.060 µs 19.085 µs 19.108 µs]
change: [+9.9923% +10.224% +10.429%] (p = 0.00 <
0.05)
Performance has regressed.
Found 10 outliers among 100 measurements (10.00%)
1 (1.00%) low severe
4 (4.00%) low mild
4 (4.00%) high mild
1 (1.00%) high severe
Benchmarking approx_distinct utf8view long 80% distinct: Collecting 100
samples in estimated 5.0800 s
approx_distinct utf8view long 80% distinct
time: [21.281 µs 21.306 µs 21.335 µs]
change: [+1.8930% +2.0965% +2.3087%] (p = 0.00 <
0.05)
Performance has regressed.
Benchmarking approx_distinct i64 99% distinct: Collecting 100 samples in
estimated 5.0037 s (884k ite
approx_distinct i64 99% distinct
time: [5.6507 µs 5.6645 µs 5.6805 µs]
change: [+0.1104% +0.3620% +0.6143%] (p = 0.00 <
0.05)
Change within noise threshold.
Found 4 outliers among 100 measurements (4.00%)
2 (2.00%) low mild
1 (1.00%) high mild
1 (1.00%) high severe
Benchmarking approx_distinct utf8 short 99% distinct: Collecting 100 samples
in estimated 5.0298 s (3
approx_distinct utf8 short 99% distinct
time: [12.956 µs 12.968 µs 12.979 µs]
change: [+15.776% +16.044% +16.296%] (p = 0.00 <
0.05)
Performance has regressed.
Found 6 outliers among 100 measurements (6.00%)
4 (4.00%) low mild
2 (2.00%) high severe
Benchmarking approx_distinct utf8view short 99% distinct: Collecting 100
samples in estimated 5.0295
approx_distinct utf8view short 99% distinct
time: [8.8005 µs 8.8054 µs 8.8106 µs]
change: [+22.620% +22.909% +23.202%] (p = 0.00 <
0.05)
Performance has regressed.
Found 7 outliers among 100 measurements (7.00%)
2 (2.00%) low severe
2 (2.00%) low mild
3 (3.00%) high severe
Benchmarking approx_distinct utf8 long 99% distinct: Collecting 100 samples
in estimated 5.0467 s (26
approx_distinct utf8 long 99% distinct
time: [19.134 µs 19.197 µs 19.309 µs]
change: [+9.0293% +9.4204% +9.8109%] (p = 0.00 <
0.05)
Performance has regressed.
Found 4 outliers among 100 measurements (4.00%)
2 (2.00%) low severe
1 (1.00%) low mild
1 (1.00%) high severe
Benchmarking approx_distinct utf8view long 99% distinct: Collecting 100
samples in estimated 5.0953 s
approx_distinct utf8view long 99% distinct
time: [21.295 µs 21.332 µs 21.384 µs]
change: [+1.9350% +2.2141% +2.4823%] (p = 0.00 <
0.05)
Performance has regressed.
Found 1 outliers among 100 measurements (1.00%)
1 (1.00%) high mild
Benchmarking approx_distinct u8 bitmap: Collecting 100 samples in estimated
5.0022 s (4.2M iterations
approx_distinct u8 bitmap
time: [1.1961 µs 1.1976 µs 1.1994 µs]
change: [−3.8254% −3.1554% −2.5920%] (p = 0.00 <
0.05)
Performance has improved.
Found 5 outliers among 100 measurements (5.00%)
2 (2.00%) low mild
3 (3.00%) high mild
Benchmarking approx_distinct i8 bitmap: Collecting 100 samples in estimated
5.0058 s (4.2M iterations
approx_distinct i8 bitmap
time: [1.2043 µs 1.2058 µs 1.2075 µs]
change: [−0.6865% −0.3102% −0.0076%] (p = 0.07 >
0.05)
No change in performance detected.
Found 9 outliers among 100 measurements (9.00%)
1 (1.00%) low severe
3 (3.00%) low mild
4 (4.00%) high mild
1 (1.00%) high severe
Benchmarking approx_distinct u16 bitmap: Collecting 100 samples in estimated
5.0052 s (1.1M iteration
approx_distinct u16 bitmap
time: [4.3272 µs 4.3392 µs 4.3521 µs]
change: [−1.5999% −1.1667% −0.7366%] (p = 0.00 <
0.05)
Change within noise threshold.
Found 6 outliers among 100 measurements (6.00%)
5 (5.00%) low mild
1 (1.00%) high mild
Benchmarking approx_distinct i16 bitmap: Collecting 100 samples in estimated
5.0171 s (1.1M iteration
approx_distinct i16 bitmap
time: [4.4383 µs 4.4431 µs 4.4479 µs]
change: [+2.4499% +2.8213% +3.1689%] (p = 0.00 <
0.05)
Performance has regressed.
Found 5 outliers among 100 measurements (5.00%)
4 (4.00%) low mild
1 (1.00%) high mild
Benchmarking approx_distinct_grouped/Int64 50000 groups: Collecting 10
samples in estimated 5.4914 s
approx_distinct_grouped/Int64 50000 groups
time: [9.8851 ms 9.9005 ms 9.9293 ms]
change: [−3.5210% −2.9384% −2.4270%] (p = 0.00 <
0.05)
Performance has improved.
Found 1 outliers among 10 measurements (10.00%)
1 (10.00%) high mild
Benchmarking approx_distinct_grouped/Utf8 50000 groups: Collecting 10
samples in estimated 5.0165 s (
approx_distinct_grouped/Utf8 50000 groups
time: [10.116 ms 10.148 ms 10.186 ms]
change: [−4.9237% −4.6468% −4.3582%] (p = 0.00 <
0.05)
Performance has improved.
Benchmarking approx_distinct_grouped/Utf8View 50000 groups: Collecting 10
samples in estimated 5.4867
approx_distinct_grouped/Utf8View 50000 groups
time: [9.9450 ms 9.9498 ms 9.9556 ms]
change: [−3.4418% −3.0161% −2.5685%] (p = 0.00 <
0.05)
Performance has improved.
Found 1 outliers among 10 measurements (10.00%)
1 (10.00%) high severe
```
</details>
It shows some get 5% faster due to batched hashing, some utf cases get
slower (the worst one 22% slower)
I think it's still a good idea to ignore the regression and simplify the
code due to:
- Amdahl's law (20% faster on function X, but function X only takes 1% of
query time, then the complexity to win the performance might not be worthy):
specifically the microbench only measured `update_batch()` function, this piece
of code is highly vectorizable, and it can very unlikely to be significant on
any real query.
I tried to construct a query that is very heavy on `update_batch`, still
can't observe end-to-end difference
```sql
> select approx_distinct(v1)
from (
select arrow_cast(v1, 'Utf8View')
from generate_series(100000000)
as t1(v1)) as t_string(v1);
+------------------------------+
| approx_distinct(t_string.v1) |
+------------------------------+
| 99201889 |
+------------------------------+
1 row(s) fetched.
Elapsed 0.139 seconds.
-- Runtime almost the same on PR v.s. main
```
- For the slowest microbench, I think the root cause is that LLVM can
optimize the manually simplified code more easily.
The existing implementation has the following fast path:
https://github.com/apache/datafusion/blob/b8998c762bb864a9f3607a518384b03dcf40eb61/datafusion/functions-aggregate/src/approx_distinct.rs#L254-L261
The same optimization also exists in the common, simpler API `create_hashes`:
https://github.com/apache/datafusion/blob/b8998c762bb864a9f3607a518384b03dcf40eb61/datafusion/common/src/hash_utils.rs#L352
The existing implementation is still faster likely because the code is
manually unrolled, while `create_hashes` is more branchy.
However, this kind of optimization can be applied endlessly and would
introduce complexity everywhere, so I do not think it is worth preserving here.
## What changes are included in this PR?
<!--
There is no need to duplicate the description in the issue here but it is
sometimes worth providing a summary of the individual changes in this PR.
-->
1. Extend `create_hashes` with a hash state that is optimized for
statistical quality
2. Simplify `approx_distinct` with create_hashes
## Are these changes tested?
<!--
We typically require tests for all PRs in order to:
1. Prevent the code from being accidentally broken by subsequent changes
3. Serve as another way to document the expected behavior of the code
If tests are not included in your PR, please explain why (for example, are
they covered by existing tests)?
-->
## Are there any user-facing changes?
<!--
If there are user-facing changes then we may require documentation to be
updated before approving the PR.
-->
<!--
If there are any breaking changes to public APIs, please add the `api
change` label.
-->
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
