Set spark.executor.cores for integration tests.
#9177
Conversation
Fixes NVIDIA#9135. (By workaround.) This change sets `spark.executor.cores` to `10`, if it is unset. This allows integration tests to work around the failure seen in `parquet_test.py:test_small_file_memory`, where the `COALESCING` Parquet reader's thread pool accidentally uses 128 threads with 8MB memory each, thus consuming the entire heap. Note that this is a bit of a workaround. A more robust solution would be to scale the Parquet reader's buffers based on the amount of available memory, and the number of threads. Signed-off-by: MithunR <mythrocks@gmail.com>
|
Build |
|
Build |
|
|
||
| # Set per-executor cores, if unspecified. | ||
| # This prevents per-thread allocations (like Parquet read buffers) from overwhelming the heap. | ||
| export PYSP_TEST_spark_executor_cores=${PYSP_TEST_spark_executor_cores:-'10'} |
There was a problem hiding this comment.
why 10? We have a few other places where we try to configure things for local mode already why are the number of executor cores out of sync with the LOCAL_PARALLEL or NUM_LOCAL_EXECS?
On a side note are databricks tests being run in local mode and configured badly? In a regular databricks cluster will we also run into this type of a problem? If so this workaround feels very much like it is going in the wrong direction, like we need to really fix the underlying problem ASAP instead of trying to work around it.
There was a problem hiding this comment.
I did see that. I figured we might consider lower parallelism for local-mode than cluster-mode, and that a more appropriate number might be suggested in the review. I have verified that this works with 4.
This was an attempt to get a clean build on CDH, as quickly as possible. But I'm supportive of closing this in favour of a proper fix. |
Fixes #9135. (By workaround.)
This change sets
spark.executor.coresto10, if it is unset. This allows integration tests to work around the failure seen inparquet_test.py:test_small_file_memory, where theCOALESCINGParquet reader's thread pool accidentally uses 128 threads with 8MB memory each, thus consuming the entire heap.Note that this is a bit of a workaround. A more robust solution would be to scale the Parquet reader's buffers based on the amount of available memory, and the number of threads.