PyFlink矢量化UDF抛出NullPointerException

Question

我有一个ML模型，它接受两个numpy.ndarray-users和items-，并返回一个numpy.ndarraypredictions。在普通的Python代码中，我会做：

model = load_model()
df = load_data() # the DataFrame includes 4 columns, namely, user_id, movie_id, rating, and timestamp
users = df.user_id.values
items = df.movie_id.values
predictions = model(users, items)

我正在考虑将这些代码移植到Flink中，以利用其分布式特性。我的假设是：通过将预测工作负载分布在多个Flink节点上，我应该能够更快地运行整个预测。

所以我写了一份PyFlink的工作。注意，我实现了一个名为predict的UDF来运行预测。

# batch_prediction.py
model = load_model()
settings = EnvironmentSettings.new_instance().use_blink_planner().build()
exec_env = StreamExecutionEnvironment.get_execution_environment()
t_env = StreamTableEnvironment.create(exec_env, environment_settings=settings)
SOURCE_DDL = """
CREATE TABLE source (
user_id INT,
movie_id INT,
rating TINYINT,
event_ms BIGINT
) WITH (
'connector' = 'filesystem',
'format' = 'csv',
'csv.field-delimiter' = 't',
'path' = 'ml-100k/u1.test'
)
"""
SINK_DDL = """
CREATE TABLE sink (
prediction DOUBLE
) WITH (
'connector' = 'print'
)
"""
t_env.execute_sql(SOURCE_DDL)
t_env.execute_sql(SINK_DDL)
t_env.execute_sql(
"INSERT INTO sink SELECT PREDICT(user_id, movie_id) FROM source"
).wait()

这是UDF。

# batch_prediction.py (cont)
@udf(result_type=DataTypes.DOUBLE())
def predict(user, item):
return model([user], [item]).item()
t_env.create_temporary_function("predict", predict)

这项工作做得很好。然而，预测实际上在source表的每一行上运行，这是不可执行的。相反，我想将80000(user_id，movie_id(对拆分为100个批次，每个批次有800行。作业触发model(users, items)函数100次(=批处理的次数(，其中users和items都有800个元素。

我找不到这样做的方法。通过查看文档，矢量化的用户定义函数可能会起作用。

# batch_prediction.py (snippet)
# I add the func_type="pandas"
@udf(result_type=DataTypes.DOUBLE(), func_type="pandas")
def predict(user, item):
...

不幸的是，事实并非如此。

> python batch_prediction.py
...
Traceback (most recent call last):
File "batch_prediction.py", line 55, in <module>
"INSERT INTO sink SELECT PREDICT(user_id, movie_id) FROM source"
File "/usr/local/anaconda3/envs/flink-ml/lib/python3.7/site-packages/pyflink/table/table_result.py", line 76, in wait
get_method(self._j_table_result, "await")()
File "/usr/local/anaconda3/envs/flink-ml/lib/python3.7/site-packages/py4j/java_gateway.py", line 1286, in __call__
answer, self.gateway_client, self.target_id, self.name)
File "/usr/local/anaconda3/envs/flink-ml/lib/python3.7/site-packages/pyflink/util/exceptions.py", line 147, in deco
return f(*a, **kw)
File "/usr/local/anaconda3/envs/flink-ml/lib/python3.7/site-packages/py4j/protocol.py", line 328, in get_return_value
format(target_id, ".", name), value)
py4j.protocol.Py4JJavaError: An error occurred while calling o51.await.
: java.util.concurrent.ExecutionException: org.apache.flink.table.api.TableException: Failed to wait job finish
at java.util.concurrent.CompletableFuture.reportGet(CompletableFuture.java:357)
at java.util.concurrent.CompletableFuture.get(CompletableFuture.java:1908)
at org.apache.flink.table.api.internal.TableResultImpl.awaitInternal(TableResultImpl.java:119)
at org.apache.flink.table.api.internal.TableResultImpl.await(TableResultImpl.java:86)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
at java.lang.reflect.Method.invoke(Method.java:498)
at org.apache.flink.api.python.shaded.py4j.reflection.MethodInvoker.invoke(MethodInvoker.java:244)
at org.apache.flink.api.python.shaded.py4j.reflection.ReflectionEngine.invoke(ReflectionEngine.java:357)
at org.apache.flink.api.python.shaded.py4j.Gateway.invoke(Gateway.java:282)
at org.apache.flink.api.python.shaded.py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:132)
at org.apache.flink.api.python.shaded.py4j.commands.CallCommand.execute(CallCommand.java:79)
at org.apache.flink.api.python.shaded.py4j.GatewayConnection.run(GatewayConnection.java:238)
at java.lang.Thread.run(Thread.java:748)
Caused by: org.apache.flink.table.api.TableException: Failed to wait job finish
at org.apache.flink.table.api.internal.InsertResultIterator.hasNext(InsertResultIterator.java:59)
at org.apache.flink.table.api.internal.TableResultImpl$CloseableRowIteratorWrapper.hasNext(TableResultImpl.java:355)
at org.apache.flink.table.api.internal.TableResultImpl$CloseableRowIteratorWrapper.isFirstRowReady(TableResultImpl.java:368)
at org.apache.flink.table.api.internal.TableResultImpl.lambda$awaitInternal$1(TableResultImpl.java:107)
at java.util.concurrent.CompletableFuture$AsyncRun.run(CompletableFuture.java:1640)
at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624)
... 1 more
Caused by: java.util.concurrent.ExecutionException: org.apache.flink.runtime.client.JobExecutionException: Job execution failed.
at java.util.concurrent.CompletableFuture.reportGet(CompletableFuture.java:357)
at java.util.concurrent.CompletableFuture.get(CompletableFuture.java:1908)
at org.apache.flink.table.api.internal.InsertResultIterator.hasNext(InsertResultIterator.java:57)
... 7 more
Caused by: org.apache.flink.runtime.client.JobExecutionException: Job execution failed.
at org.apache.flink.runtime.jobmaster.JobResult.toJobExecutionResult(JobResult.java:147)
at org.apache.flink.runtime.minicluster.MiniClusterJobClient.lambda$getJobExecutionResult$2(MiniClusterJobClient.java:119)
at java.util.concurrent.CompletableFuture.uniApply(CompletableFuture.java:616)
at java.util.concurrent.CompletableFuture$UniApply.tryFire(CompletableFuture.java:591)
at java.util.concurrent.CompletableFuture.postComplete(CompletableFuture.java:488)
at java.util.concurrent.CompletableFuture.complete(CompletableFuture.java:1975)
at org.apache.flink.runtime.rpc.akka.AkkaInvocationHandler.lambda$invokeRpc$0(AkkaInvocationHandler.java:229)
at java.util.concurrent.CompletableFuture.uniWhenComplete(CompletableFuture.java:774)
at java.util.concurrent.CompletableFuture$UniWhenComplete.tryFire(CompletableFuture.java:750)
at java.util.concurrent.CompletableFuture.postComplete(CompletableFuture.java:488)
at java.util.concurrent.CompletableFuture.complete(CompletableFuture.java:1975)
at org.apache.flink.runtime.concurrent.FutureUtils$1.onComplete(FutureUtils.java:996)
at akka.dispatch.OnComplete.internal(Future.scala:264)
at akka.dispatch.OnComplete.internal(Future.scala:261)
at akka.dispatch.japi$CallbackBridge.apply(Future.scala:191)
at akka.dispatch.japi$CallbackBridge.apply(Future.scala:188)
at scala.concurrent.impl.CallbackRunnable.run(Promise.scala:36)
at org.apache.flink.runtime.concurrent.Executors$DirectExecutionContext.execute(Executors.java:74)
at scala.concurrent.impl.CallbackRunnable.executeWithValue(Promise.scala:44)
at scala.concurrent.impl.Promise$DefaultPromise.tryComplete(Promise.scala:252)
at akka.pattern.PromiseActorRef.$bang(AskSupport.scala:572)
at akka.pattern.PipeToSupport$PipeableFuture$$anonfun$pipeTo$1.applyOrElse(PipeToSupport.scala:22)
at akka.pattern.PipeToSupport$PipeableFuture$$anonfun$pipeTo$1.applyOrElse(PipeToSupport.scala:21)
at scala.concurrent.Future$$anonfun$andThen$1.apply(Future.scala:436)
at scala.concurrent.Future$$anonfun$andThen$1.apply(Future.scala:435)
at scala.concurrent.impl.CallbackRunnable.run(Promise.scala:36)
at akka.dispatch.BatchingExecutor$AbstractBatch.processBatch(BatchingExecutor.scala:55)
at akka.dispatch.BatchingExecutor$BlockableBatch$$anonfun$run$1.apply$mcV$sp(BatchingExecutor.scala:91)
at akka.dispatch.BatchingExecutor$BlockableBatch$$anonfun$run$1.apply(BatchingExecutor.scala:91)
at akka.dispatch.BatchingExecutor$BlockableBatch$$anonfun$run$1.apply(BatchingExecutor.scala:91)
at scala.concurrent.BlockContext$.withBlockContext(BlockContext.scala:72)
at akka.dispatch.BatchingExecutor$BlockableBatch.run(BatchingExecutor.scala:90)
at akka.dispatch.TaskInvocation.run(AbstractDispatcher.scala:40)
at akka.dispatch.ForkJoinExecutorConfigurator$AkkaForkJoinTask.exec(ForkJoinExecutorConfigurator.scala:44)
at akka.dispatch.forkjoin.ForkJoinTask.doExec(ForkJoinTask.java:260)
at akka.dispatch.forkjoin.ForkJoinPool$WorkQueue.runTask(ForkJoinPool.java:1339)
at akka.dispatch.forkjoin.ForkJoinPool.runWorker(ForkJoinPool.java:1979)
at akka.dispatch.forkjoin.ForkJoinWorkerThread.run(ForkJoinWorkerThread.java:107)
Caused by: org.apache.flink.runtime.JobException: Recovery is suppressed by NoRestartBackoffTimeStrategy
at org.apache.flink.runtime.executiongraph.failover.flip1.ExecutionFailureHandler.handleFailure(ExecutionFailureHandler.java:116)
at org.apache.flink.runtime.executiongraph.failover.flip1.ExecutionFailureHandler.getFailureHandlingResult(ExecutionFailureHandler.java:78)
at org.apache.flink.runtime.scheduler.DefaultScheduler.handleTaskFailure(DefaultScheduler.java:224)
at org.apache.flink.runtime.scheduler.DefaultScheduler.maybeHandleTaskFailure(DefaultScheduler.java:217)
at org.apache.flink.runtime.scheduler.DefaultScheduler.updateTaskExecutionStateInternal(DefaultScheduler.java:208)
at org.apache.flink.runtime.scheduler.SchedulerBase.updateTaskExecutionState(SchedulerBase.java:610)
at org.apache.flink.runtime.scheduler.SchedulerNG.updateTaskExecutionState(SchedulerNG.java:89)
at org.apache.flink.runtime.jobmaster.JobMaster.updateTaskExecutionState(JobMaster.java:419)
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
at java.lang.reflect.Method.invoke(Method.java:498)
at org.apache.flink.runtime.rpc.akka.AkkaRpcActor.handleRpcInvocation(AkkaRpcActor.java:286)
at org.apache.flink.runtime.rpc.akka.AkkaRpcActor.handleRpcMessage(AkkaRpcActor.java:201)
at org.apache.flink.runtime.rpc.akka.FencedAkkaRpcActor.handleRpcMessage(FencedAkkaRpcActor.java:74)
at org.apache.flink.runtime.rpc.akka.AkkaRpcActor.handleMessage(AkkaRpcActor.java:154)
at akka.japi.pf.UnitCaseStatement.apply(CaseStatements.scala:26)
at akka.japi.pf.UnitCaseStatement.apply(CaseStatements.scala:21)
at scala.PartialFunction$class.applyOrElse(PartialFunction.scala:123)
at akka.japi.pf.UnitCaseStatement.applyOrElse(CaseStatements.scala:21)
at scala.PartialFunction$OrElse.applyOrElse(PartialFunction.scala:170)
at scala.PartialFunction$OrElse.applyOrElse(PartialFunction.scala:171)
at scala.PartialFunction$OrElse.applyOrElse(PartialFunction.scala:171)
at akka.actor.Actor$class.aroundReceive(Actor.scala:517)
at akka.actor.AbstractActor.aroundReceive(AbstractActor.scala:225)
at akka.actor.ActorCell.receiveMessage(ActorCell.scala:592)
at akka.actor.ActorCell.invoke(ActorCell.scala:561)
at akka.dispatch.Mailbox.processMailbox(Mailbox.scala:258)
at akka.dispatch.Mailbox.run(Mailbox.scala:225)
at akka.dispatch.Mailbox.exec(Mailbox.scala:235)
... 4 more
Caused by: org.apache.flink.streaming.runtime.tasks.AsynchronousException: Caught exception while processing timer.
at org.apache.flink.streaming.runtime.tasks.StreamTask$StreamTaskAsyncExceptionHandler.handleAsyncException(StreamTask.java:1108)
at org.apache.flink.streaming.runtime.tasks.StreamTask.handleAsyncException(StreamTask.java:1082)
at org.apache.flink.streaming.runtime.tasks.StreamTask.invokeProcessingTimeCallback(StreamTask.java:1213)
at org.apache.flink.streaming.runtime.tasks.StreamTask.lambda$null$17(StreamTask.java:1202)
at org.apache.flink.streaming.runtime.tasks.StreamTaskActionExecutor$SynchronizedStreamTaskActionExecutor.runThrowing(StreamTaskActionExecutor.java:92)
at org.apache.flink.streaming.runtime.tasks.mailbox.Mail.run(Mail.java:78)
at org.apache.flink.streaming.runtime.tasks.mailbox.MailboxExecutorImpl.tryYield(MailboxExecutorImpl.java:91)
at org.apache.flink.streaming.runtime.tasks.StreamOperatorWrapper.quiesceTimeServiceAndCloseOperator(StreamOperatorWrapper.java:155)
at org.apache.flink.streaming.runtime.tasks.StreamOperatorWrapper.close(StreamOperatorWrapper.java:130)
at org.apache.flink.streaming.runtime.tasks.OperatorChain.closeOperators(OperatorChain.java:412)
at org.apache.flink.streaming.runtime.tasks.StreamTask.afterInvoke(StreamTask.java:585)
at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:547)
at org.apache.flink.runtime.taskmanager.Task.doRun(Task.java:722)
at org.apache.flink.runtime.taskmanager.Task.run(Task.java:547)
at java.lang.Thread.run(Thread.java:748)
Caused by: TimerException{java.lang.RuntimeException: Failed to close remote bundle}
... 13 more
Caused by: java.lang.RuntimeException: Failed to close remote bundle
at org.apache.flink.streaming.api.runners.python.beam.BeamPythonFunctionRunner.finishBundle(BeamPythonFunctionRunner.java:371)
at org.apache.flink.streaming.api.runners.python.beam.BeamPythonFunctionRunner.flush(BeamPythonFunctionRunner.java:325)
at org.apache.flink.streaming.api.operators.python.AbstractPythonFunctionOperator.invokeFinishBundle(AbstractPythonFunctionOperator.java:291)
at org.apache.flink.table.runtime.operators.python.scalar.arrow.RowDataArrowPythonScalarFunctionOperator.invokeFinishBundle(RowDataArrowPythonScalarFunctionOperator.java:77)
at org.apache.flink.streaming.api.operators.python.AbstractPythonFunctionOperator.checkInvokeFinishBundleByTime(AbstractPythonFunctionOperator.java:285)
at org.apache.flink.streaming.api.operators.python.AbstractPythonFunctionOperator.lambda$open$0(AbstractPythonFunctionOperator.java:134)
at org.apache.flink.streaming.runtime.tasks.StreamTask.invokeProcessingTimeCallback(StreamTask.java:1211)
... 12 more
Caused by: java.lang.NullPointerException
at org.apache.flink.streaming.api.runners.python.beam.BeamPythonFunctionRunner.finishBundle(BeamPythonFunctionRunner.java:369)
... 18 more

错误消息没有多大帮助。有人能帮忙吗？谢谢

注意：源代码可以在这里找到。要运行代码，您将需要Anaconda本地，然后：

conda env create -f environment.yml
conda activate flink-ml

Answer 1

Apache Flink社区对Dian Fu的赞扬。请参见螺纹。

对于Pandas UDF，每个输入参数的输入类型都是Pandas。序列和结果类型也应该是Panda。系列此外，结果的长度应该与输入的长度相同。你能检查一下你的Pandas UDF实现是否是这样吗？

然后我决定为我的UDF添加一个pytest单元测试来验证输入和输出类型。方法如下：

import pandas as pd
from udf_def import predict

def test_predict():
f = predict._func
users = pd.Series([1, 2, 3])
items = pd.Series([1, 4, 9])
preds = f(users, items)
assert isinstance(preds, pd.Series)
assert len(preds) == 3

它揭示了实现错误。然后我修复我的UDF实现以通过单元测试。

from model_def import MatrixFactorization
@udf(result_type=DataTypes.DOUBLE(), func_type="pandas")
def predict(users, items):
n_users, n_items = 943, 1682
model = MatrixFactorization(n_users, n_items)
model.load_state_dict(torch.load("model.pth"))
return pd.Series(model(users, items).detach().numpy())

使用新的UDF实现，我能够运行批处理预测。

> python batch_prediction.py
1> +I(-2.3802385330200195)
1> +I(20.000154495239258)
1> +I(-30.704544067382812)
1> +I(-11.602800369262695)
1> +I(10.998968124389648)
...

更新后的源代码可以在这里找到。