changes to support db 13.3+ #716
Conversation
|
build |
| dim = len(cluster_centers[0]) | ||
| # inject unsupported expr (slice) that is essentially a noop | ||
| df_for_scoring = df_for_scoring.select( | ||
| F.slice(feature_col, 1, dim).alias(feature_col), output_col |
There was a problem hiding this comment.
Interesting. Any intuition as to why slice can get the hanging resolved?
There was a problem hiding this comment.
Yes. It is an unsupported expression in spark-rapids so falls back to cpu and injects columnartorow and rowtocolumnar transformations which do some batching that doesn't (but should after patching) happen otherwise.
|
|
||
| 6. Monitor progress periodically in case of a possible hang, to avoid incurring cloud costs in such cases. | ||
| 2. Monitor progress periodically in case of a possible hang, to avoid incurring cloud costs in such cases. |
There was a problem hiding this comment.
Do we keep item number 5, 6 or replace with 1, 2?
There was a problem hiding this comment.
Interesting. Didn't intend to change the numbers, will revert.
| "spark.sql.execution.arrow.pyspark.enabled": "true", | ||
| "spark.sql.files.maxPartitionBytes": "2000000000000", | ||
| "spark.databricks.delta.optimizeWrite.enabled": "false", | ||
| "spark.rapids.sql.concurrentGpuTasks": "2" |
There was a problem hiding this comment.
While we keep this here, do we need to remove it from gpu_cluster_spec.sh (the one without GPU ETL)?
There was a problem hiding this comment.
Yes. Will delete there. Good catch.
Signed-off-by: Erik Ordentlich <eordentlich@gmail.com>
Signed-off-by: Erik Ordentlich <eordentlich@gmail.com>
Signed-off-by: Erik Ordentlich <eordentlich@gmail.com>
eordentlich
force-pushed
the
eo_db_13.3_patch
branch
from
40b9e84
to
9c85bd1
Compare
|
build |
Also adds db version as option to db benchmark script and a gpu only option (i.e. no spark rapids plugin).
Plus some other misc. updates (e.g. to logistic reg. notebooks, spark-rapids versions).
Includes a temporary patch till NVIDIA/spark-rapids#10770 is fixed.