Hi folks, I’m running a test with bentoml 1.2.4 on Spark. But I got

'APIMethod' object has no attribute 'output'

from run_in_spark when I ran

bentoml.batch.run_in_spark(bento, df, spark)

. Stacktrace:

results = bentoml.batch.run_in_spark(bento, df, spark)
  File "/usr/local/lib/python3.8/site-packages/bentoml/_internal/batch/spark.py", line 169, in run_in_spark
    output_schema = api.output.spark_schema()
AttributeError: 'APIMethod' object has no attribute 'output'

I checked the APIMethod implementation and did not see an “output” or an “input” attribute. Is this expected? Anybody else ran into issues running bentos on Spark? Thanks!

Hi, the spark guide is a bit outdated right now for BentoML 1.2, this api unfortunately only work with 1.1.x using DataFrame as input and output. We are building out a more generic solution for spark batch inference at the moment

Thanks @Chaoyu! I switched to version 1.1.11 and now I don’t get the same error anymore, but now I have a new question: I want my server to take in an numpy array of shape Nx1 and return an numpy array of shape NxM, where batch size N is unknown since we rely on the spark batch inference to decide how big a batch is. If I don’t define a shape in my NumpyNdarray descriptor, I get error

Cannot perform batch inference with a numpy output without a known shape; please provide a shape

, but I don’t know how to define the shape since I don’t know the batch size… What should I do if I want to use Spark batch inference and expect the output to be an NxM numpy array?