[GLUTEN-5074][VL] fix: UDF load error in yarn-cluster mode #5075
Conversation
|
Thanks for opening a pull request! Could you open an issue for this pull request on Github Issues? https://github.com/apache/incubator-gluten/issues Then could you also rename commit message and pull request title in the following format? See also: |
kecookier
changed the title
| if (!canAccessSparkFiles) { | ||
| throw new IllegalArgumentException( | ||
| "On yarn-client mode, driver only accepts absolute paths, but got " + f) | ||
| val uri = Utils.resolveURI(f) |
There was a problem hiding this comment.
Looks like f can also be a relative filename or file tag here without a scheme. Would this logic be able to handle such case?
e.g.
--files /path/to/gluten/cpp/build/velox/udf/examples/libmyudf.so
--conf spark.gluten.sql.columnar.backend.velox.udfLibraryPaths=libmyudf.so
There was a problem hiding this comment.
yes，if f is libmyudf.so, Utils.resolveURI() will return file://${PWD}/libmyudf.so
There was a problem hiding this comment.
But file://${PWD}/libmyudf.so is not the expected path, right? It should be /path/to/gluten/cpp/build/velox/udf/examples/libmyudf.so
There was a problem hiding this comment.
The call to Utils.resolveURI() is used to determine whether the file is local or remote.
In your example, libmyudf.so is a local file, and we will not use uri.path directly.
The --files argument copies this file to a different destination directory. When --master=yarn is specified, the file is copied to the working directory on all nodes (both the driver and executors). In local mode, the files are added using SparkContext.addFile, and they can then be accessed using SparkFiles.get(f).
There was a problem hiding this comment.
In your example, libmyudf.so is a local file
That doesn't make sense to me. libmyudf.so here should refer to the file uploaded by --files. But here it's resolved as a relative path to the runtime directory.
The call to Utils.resolveURI() is used to determine whether the file is local or remote.
It would be better to check whether it's a relative path first. Although it's resolved as a local file and pass the if condition, the path doesn't even exist.
When --master=yarn is specified, the file is copied to the working directory on all nodes (both the driver and executors).
Based on my previous experience, for yarn client mode, the files will be copied to all executor container + AM container, so they won't be copied to the driver node. In this case, if we only use the two configurations below
--files /path/to/gluten/cpp/build/velox/udf/examples/libmyudf.so
--conf spark.gluten.sql.columnar.backend.velox.udfLibraryPaths=libmyudf.so
the driver will fail to get the actual path for libmyudf.so, and that's the reason for adding spark.gluten.sql.columnar.backend.velox.driver.udfLibraryPaths
There was a problem hiding this comment.
Thank you for your patient explanation, Rong!
I previously thought that under yarn-client mode, spark.yarn.dist.files would also copy the files to the driver. I got the AM and driver mixed up.
Let's fix this issue.
There was a problem hiding this comment.
I would suggest preserving 2 kinds of configuration for spark.gluten.sql.columnar.backend.velox.udfLibraryPaths
- relative path
- URI
And make `spark.gluten.sql.columnar.backend.velox.udfLibraryPaths=/path/to/..." invalid.
There was a problem hiding this comment.
@marin-ma I have already fix the code according to the previous discussion in the comments, could you help review it again?
backends-velox/src/main/scala/org/apache/spark/sql/expression/UDFResolver.scala
Outdated
Show resolved
Hide resolved
kecookier
force-pushed
the
vec-udf
branch
2 times, most recently
from
4c3a092 to
76bde19
Compare
| @@ -82,7 +82,8 @@ class VeloxUdfSuiteLocal extends VeloxUdfSuite { | |||
| override val master: String = "local[2]" | |||
| override protected def sparkConf: SparkConf = { | |||
| super.sparkConf | |||
| .set("spark.gluten.sql.columnar.backend.velox.udfLibraryPaths", udfLibPath) | |||
| .set("spark.files", udfLibPath) | |||
| .set("spark.gluten.sql.columnar.backend.velox.udfLibraryPaths", "libmyudf.so") | |||
There was a problem hiding this comment.
Please do not hard code "libmyudf.so" here. Use a helper function to extract the filename from udfLibPath. Note the udfLibPath can also be a comma-separated string with multiple paths.
|
Could you also update the examples in the document with either relative or URI path? |
What changes were proposed in this pull request?
driverUdfLibPathand retain onlyudfLibPath, as there is no need to differentiate between the driver and executor. The method used to access the file is determined by whether --master=yarn is specified.VeloxBackend::initUdfin both the driver and the executor.UdfResolverdoes not load the UDF repeatedly; it only retrieves the function signatures.(Fixes: #5074)
How was this patch tested?
I tested with and without the --files/--archives arguments in local, yarn-client, and yarn-cluster modes.