diff options
author | Herman van Hovell <hvanhovell@questtec.nl> | 2015-07-18 23:44:38 -0700 |
---|---|---|
committer | Yin Huai <yhuai@databricks.com> | 2015-07-18 23:44:38 -0700 |
commit | a9a0d0cebf8ab3c539723488e5945794ebfd6104 (patch) | |
tree | 1adb8d37a3259a3677cd3f9b686c9f404a1fcbab /data/mllib | |
parent | 04c1b49f5eee915ad1159a32bf12836a3b9f2620 (diff) | |
download | spark-a9a0d0cebf8ab3c539723488e5945794ebfd6104.tar.gz spark-a9a0d0cebf8ab3c539723488e5945794ebfd6104.tar.bz2 spark-a9a0d0cebf8ab3c539723488e5945794ebfd6104.zip |
[SPARK-8638] [SQL] Window Function Performance Improvements
## Description
Performance improvements for Spark Window functions. This PR will also serve as the basis for moving away from Hive UDAFs to Spark UDAFs. See JIRA tickets SPARK-8638 and SPARK-7712 for more information.
## Improvements
* Much better performance (10x) in running cases (e.g. BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW) and UNBOUDED FOLLOWING cases. The current implementation in spark uses a sliding window approach in these cases. This means that an aggregate is maintained for every row, so space usage is N (N being the number of rows). This also means that all these aggregates all need to be updated separately, this takes N*(N-1)/2 updates. The running case differs from the Sliding case because we are only adding data to an aggregate function (no reset is required), we only need to maintain one aggregate (like in the UNBOUNDED PRECEDING AND UNBOUNDED case), update the aggregate for each row, and get the aggregate value after each update. This is what the new implementation does. This approach only uses 1 buffer, and only requires N updates; I am currently working on data with window sizes of 500-1000 doing running sums and this saves a lot of time. The CURRENT ROW AND UNBOUNDED FOLLOWING case also uses this approach and the fact that aggregate operations are communitative, there is one twist though it will process the input buffer in reverse.
* Fewer comparisons in the sliding case. The current implementation determines frame boundaries for every input row. The new implementation makes more use of the fact that the window is sorted, maintains the boundaries, and only moves them when the current row order changes. This is a minor improvement.
* A single Window node is able to process all types of Frames for the same Partitioning/Ordering. This saves a little time/memory spent buffering and managing partitions. This will be enabled in a follow-up PR.
* A lot of the staging code is moved from the execution phase to the initialization phase. Minor performance improvement, and improves readability of the execution code.
## Benchmarking
I have done a small benchmark using [on time performance](http://www.transtats.bts.gov) data of the month april. I have used the origin as a partioning key, as a result there is quite some variation in window sizes. The code for the benchmark can be found in the JIRA ticket. These are the results per Frame type:
Frame | Master | SPARK-8638
----- | ------ | ----------
Entire Frame | 2 s | 1 s
Sliding | 18 s | 1 s
Growing | 14 s | 0.9 s
Shrinking | 13 s | 1 s
Author: Herman van Hovell <hvanhovell@questtec.nl>
Closes #7057 from hvanhovell/SPARK-8638 and squashes the following commits:
3bfdc49 [Herman van Hovell] Fixed Perfomance Regression for Shrinking Window Frames (+Rebase)
2eb3b33 [Herman van Hovell] Corrected reverse range frame processing.
2cd2d5b [Herman van Hovell] Corrected reverse range frame processing.
b0654d7 [Herman van Hovell] Tests for exotic frame specifications.
e75b76e [Herman van Hovell] More docs, added support for reverse sliding range frames, and some reorganization of code.
1fdb558 [Herman van Hovell] Changed Data In HiveDataFrameWindowSuite.
ac2f682 [Herman van Hovell] Added a few more comments.
1938312 [Herman van Hovell] Added Documentation to the createBoundOrdering methods.
bb020e6 [Herman van Hovell] Major overhaul of Window operator.
Diffstat (limited to 'data/mllib')
0 files changed, 0 insertions, 0 deletions