Transpile simple choice-type regular expressions into lists of choices to use with string replace multi #7967
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Navin Kumar <navink@nvidia.com>
…-transpile-replace-multi
…-transpile-replace-multi
…e are backrefs Signed-off-by: Navin Kumar <navink@nvidia.com>
Signed-off-by: Navin Kumar <navink@nvidia.com>
|
build |
andygrove
reviewed
Mar 30, 2023
andygrove
reviewed
Mar 30, 2023
sql-plugin/src/main/scala/org/apache/spark/sql/rapids/stringFunctions.scala
Show resolved
Hide resolved
| def test_regexp_replace_fallback(): | ||
| gen = mk_str_gen('[abcdef]{0,2}') | ||
|
|
||
| conf = { 'spark.rapids.sql.regexp.enabled': 'false' } |
There was a problem hiding this comment.
nit, here and other places, could use typed constants
Suggested change
| conf = { 'spark.rapids.sql.regexp.enabled': 'false' } | |
| conf = { 'spark.rapids.sql.regexp.enabled': False } |
| @@ -593,6 +593,11 @@ object GpuOverrides extends Logging { | |||
| lit.value == null | |||
| } | |||
|
|
|||
| def isSupportedStringReplacePattern(strLit: String): Boolean = { | |||
| // check for regex special characters, except for \u0000 which we can support | |||
| !regexList.filterNot(_ == "\u0000").exists(pattern => strLit.contains(pattern)) | |||
There was a problem hiding this comment.
I think it's more readable in a conjunctive form, but it is not part of your PR, so very optional:
Suggested change
| !regexList.filterNot(_ == "\u0000").exists(pattern => strLit.contains(pattern)) | |
| !regexList.exists(pattern => pattern != "\u0000" && strLit.contains(pattern)) |
Signed-off-by: Navin Kumar <navink@nvidia.com>
Signed-off-by: Navin Kumar <navink@nvidia.com>
…-transpile-replace-multi
|
build |
Signed-off-by: Navin Kumar <navink@nvidia.com>
|
build |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fixes #7907.
This uses the updated stringReplace(Multi) API from rapidsai/cudf#12858 and rapidsai/cudf#12979 to optimize scenarios that involve simple choices in regular expressions. For example, the regular expressions
aa|bbcan be transpiled to a list["aa", "bb"]which can be converted to a ColumnVector and passed to the new stringReplace(Multi) API without using any regex. This results in an improved speedup (especially for large strings).Some performance numbers:
This test created a Parquet file with 4096 rows of a single string column for each of string lengths. It then used a regular expression that was a simple choice (e.g.
aaaaa|bbbbb), and then called regexp_replace on the dataframe using SQL, and then writes the result back to Parquet. This Parquet method was used due to the fact that keeping the entire dataset in memory results in OOM (running out of heap space on the JVM) on the CPU as the string length is increased. Posted is the effective speedup in a table. You can see this in the graph here:One interesting result is that there is no obvious inflection point in terms of string length. In general, it looks like the multi-replace optimization will either perform about the same, and as the string size increases in the dataframe, the speedup difference between the GPU regexp_replace and the optimized version that transpiles the regex choice into a list of strings which can then be passed to the stringReplace(Multi) API and executed in parallel on the GPU.
Source Code of Benchmark