Crayfish: Navigating the Labyrinth of Machine Learning Inference in Stream Processing Systems
Project Structure
.
├── core # Crayfish Java core components and abstractions.
├── crayfish-java # Crayfish adapters (i.e., FLink, Spark Structured Streaming, Kafka Streams).
├── docker # Scripts and configuration for Docker deployment.
├── experiments-driver # Experiments testbed and configurations.
├── input-producer # Input Producer Component. Contains a random input generator.
├── output-consumer # Output Consumer Component. Writes latency measurements to persistent storage.
├── rayfish # Crayfish Ray Adapter.
├── resources # Pre-trained models and training scripts.
└── results-analysis # Notebooks to analyze the results.
Supported Tools
Pre-trained Models
FNN | ResNet50 | ||
---|---|---|---|
Input Size | 28x28 | 224x224x3 | |
Output Size | 10x1 | 1000x1 | |
Parameters Number | 28K | 23M | |
Model Size | ONNX | 113 KB | 97 MB |
Torch | 115 KB | 98 MB | |
H5 | 133 KB | 98 MB | |
SavedModel | 508 KB | 101 MB |
Stream Processors
- Apache Flink 1.15.2
- Apache Kafka Streams 3.2.3
- Spark Structured Streaming 3.3.2
- Ray 2.4
Embedded Serving Frameworks
- ONNX 1.12.1
- DeepLearning4j 1.0.0-M2.1
- TensorFlow Java (SavedModel) 1.13.1
External Serving Frameworks
- TorchServe 0.7.1
- TensorFlow Serving 2.11.1
Quick Start
Environment
- Unix-like environment
- Python 3
- Maven
- Java 8
- Docker installation
Experiments
We offer an option to test the tools locally before deploying to a cluster.
# train ffnn/resnet50 models in different model formats
./resources/train-models.sh
# build images with the trained models
./docker/docker-build.sh
# run all experiments locally (to run a single experiment, see the options below)
./run.sh -e a -ec s -sp a -m a -msm a -msx a -em l
# clean all the running and exited containers and created volumes if you quit before the experiment finishes
./docker/docker-stop.sh
# clean log, results, and models trained by ./resources/train-models.sh
./clean-exp-files.sh
run.sh
has the following options:
Arguments:
[!] NOTE: Configs in 'experiments-driver/configs/exp-configs/[-ec]/[-e]' will be run.
-e | --experiments Independent variable sets to run: a=all, i=input rate, b=batch size, s=scalability, r=bursty rate.
-ec | --experiments-control Controlled variable sets to run: s=small, l=large, d=debug.
- small: Run input rate 256 for the scalability experiment.
[!] NOTE: ResNet50 is recommended for this option due to the large
model size and limited memory.
- large: Run input rate 30000 for the scalability experiment.
- debug: Run simple experiment configs in the debug folder.
-sp | --stream-processors Stream processor to test:
a=all, f=Apache Flink, k=Kafka Streams, s=Spark Streaming, r=Ray.
-m | --models Served models: a=all, f=ffnn, r=resnet50.
-msm | --embedded-model-servers Embedded model serving alternative:
x=none, a=all (w/o noop and nd4j), n=nd4j, d=dl4j, o=onnx, t=tf-savedmodel, k=noop.
[!] NOTE: noop will execute input rate and batch size experiments.
-msx | --external-model-servers External model serving alternative: x=none, a=all, t=tf-serving, s=torchserve.
-em | --execution-mode Execution mode: l=local, c=cluster.
-d | --default-configs Print default configs.
-h | --help Help.
Run on Cluster
To adapt the experiments to be executed on the cluster,
- Build the required docker images on different VMs and train models (commands are from
./docker/docker-build.sh
)- Kafka consumer VM
docker compose build output-consumer --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g)
- Data processor
./resources/train-models.sh docker compose build data-processor --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g)
- Kafka producer VM
docker compose build input-producer --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g)
- External serving VM
./resources/train-models.sh docker compose build external-serving-agent --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g) docker compose build torch-serving docker compose build tf-serving docker compose build ray-serving
- Kafka consumer VM
- Open the respective docker container on different VMs manually. The order matters. (commands are from
./run.sh
)- Kafka producer VM
docker compose run -d --service-ports --use-aliases -u $(id -u):$(id -g) input-producer \ /bin/bash -c "./experiments-driver/scripts/start-daemon.sh -p kp -em 'l' --port $PORT_KP >logs/$EXP_UNIQUE_ID-kp-daemon-logs.txt 2>&1"
- Data processor
docker compose run -d --service-ports --use-aliases -u $(id -u):$(id -g) data-processor \ /bin/bash -c "./experiments-driver/scripts/start-daemon.sh -p dp -em 'l' --port $PORT_DP >logs/$EXP_UNIQUE_ID-dpd-daemon-logs.txt 2>&1"
- External serving VM
docker compose run -u $(id -u):$(id -g) output-consumer \ /bin/bash -c "./experiments-driver/scripts/start-consumer.sh \ -e $EXPERIMENTS_OPT \ -ec $EXPERIMENTS_CTRL_OPT \ -sp $STREAM_PROCESSOR_OPT \ -m $MODELS_OPT \ -msm $EMBEDDED_MODEL_SERVERS_OPT \ -msx $EXTERNAL_MODEL_SERVERS_OPT \ -em $EXECUTION_MODE_OPT \ -eid $EXP_UNIQUE_ID"
- Kafka consumer VM
docker compose run -u $(id -u):$(id -g) output-consumer \ /bin/bash -c "./experiments-driver/scripts/start-consumer.sh \ -e $EXPERIMENTS_OPT \ -ec $EXPERIMENTS_CTRL_OPT \ -sp $STREAM_PROCESSOR_OPT \ -m $MODELS_OPT \ -msm $EMBEDDED_MODEL_SERVERS_OPT \ -msx $EXTERNAL_MODEL_SERVERS_OPT \ -em $EXECUTION_MODE_OPT \ -eid $EXP_UNIQUE_ID"
- Kafka producer VM
- Update the IP addresses for each VM in
experiments-driver/configs/global-configs-cluster.properties
Kafka Overhead Experiments
The experiments measuring the overhead introduced by Kafka in the pipeline do not use the Crayfish pipeline, as they employ a standalone FLink implementation. To run these experiments, the following script can be used:
./experiments-driver/run-standalone.sh -em l -ec l
run-standalone.sh
has the following options:
Arguments:
-ec | --experiments-control Controlled variable sets to run: s=small, l=large, d=debug.
- small: Run input rate 256 for the scalability experiment.
[!] NOTE: ResNet50 is recommended for this option due to the large
model size and limited memory.
- large: Run input rate 30000 for the scalability experiment.
- debug: Run simple experiment configs in the debug folder.
-em | --execution-mode Execution mode: l=local, c=cluster.
Extending Crayfish
Crayfish provides a set of interfaces that allow developers to extend the benchmarking frameworks with other stream processing systems, model serving tools, and pre-trained models. We showcase examples of how to do so for each type of system.
New Stream Processors
New stream processors can be added through abstractions similar to adapters. These adapters need to extend the Crayfish interface and provide functionalities for input reading, model scoring, and output writing alongside logic to set the parallelism of the computation. The adapter needs to provide also abstractions for a stream builder, the generic operator type, and the sink operator. When using the adapter, the model type can be chosen from the list of models supported out of the box or can be a custom model defined by the Crayfish user.
The adapter should extend the Crayfish
abstract class, as below. Next, implement the required methods for building your streaming application.
public class MyCrayfishAdapter extends Crayfish<MyStreamBuilder, MyOperatorType, MySinkType, MyModel> {
// Implement required methods
// ...
@Override
public MyStreamBuilder streamBuilder() {
// Implement the logic to create and return the stream builder specific to the stream processor
}
@Override
public MyOperatorType inputOp(MyStreamBuilder streamBuilder) {
// Implement the logic to create and return a source operator
}
@Override
public MyOperatorType embeddedScoringOp(MyStreamBuilder sb, MyOperatorType input) throws Exception {
// Implement the logic for embedded ML serving
}
@Override
public MyOperatorType externalScoringOp(MyStreamBuilder sb, MyOperatorType input) throws Exception {
// Implement the logic for external ML serving
}
@Override
public CrayfishUtils.Either<MyOperatorType, MySinkType> outputOp(MyStreamBuilder sb,
MyOperatorType output) throws Exception {
// Implement the logic for creating and returning output. Depending on the stream processor intrinsics,
// return either the sink or another abstraction. This will be used in the start() method to start the
// streaming application.
}
@Override
public void start(MyStreamBuilder sb, Properties metaData,
CrayfishUtils.Either<MyOperatorType, MySinkType> out) throws Exception {
// Implement the logic to start the processing
}
@Override
public boolean hasOperatorParallelism() {
// Return true if operator parallelism is supported
}
@Override
public CrayfishUtils.Either<MyOperatorType, MySinkType> setOperatorParallelism(
CrayfishUtils.Either<MyOperatorType, MySinkType> operator, int parallelism) throws Exception {
// Implement the logic to set operator parallelism
// Return null if not supported
}
@Override
public void setDefaultParallelism(MyStreamBuilder sb, Properties metaData, int parallelism) throws Exception {
// Implement the logic to set default parallelism
}
@Override
public Properties addMetadata() throws Exception {
// Implement the logic to add metadata. Some stream processing systems require extra configurations to be
// set via Properties.
// Return null if not required.
}
}
The users can use the new adapter, for instance, to serve DeepLearning4J models, for instance, as follows:
Crayfish adapter = new MyCrayfishAdapter<DL4JModel>(DL4JModel.class, config);
adapter.run();
The adapters for Apache Flink, Kafka Streams, and Spark Structured Streaming can be found under crayfish-java/adapters/.
New Model Serving Tools
Crayfish can be extended to support ML serving tools besides the ones included off-the-shelf. To do so, create a new Java class and extend CrayfishModel
. Then, implement the required methods as below.
public class MyModel extends CrayfishModel {
@Override
public void loadModel(String modelName, String location) throws Exception {
// Implement the logic to load the model
// Use modelName and location to load the model from the specified location
}
@Override
public void build() throws Exception {
// Implement the logic to build the model
// This method is called after loading the model
}
@Override
public CrayfishPrediction apply(CrayfishInputData input) throws Exception {
// Implement the logic to apply the model to the input data
// Use the input data to make predictions
}
}
The users can use the new model, for instance, in conjunction with Spark Structured Streaming, for instance, as follows:
Crayfish adapter = new SparkSSCrayfishAdapter<MyModel>(MyModel.class, config);
adapter.run();
The implementations corresponding to the supported models can be found under core/src/main/java/datatypes/models.
Note that there is no distinction between embedded and external tools; external tools need to send requests to the external serving instance to perform the inference. To do so, Crayfish provides the helper class InferenceRequest
to perform HTTP requests to a given input. The class can be extended to customize gRPC requests as well. Examples can be found under core/src/main/java/request/.
New Pre-trained ML Models
Crayfish supports two models out of the box: FFNN and ResNet50. However, the users can configure the framework to serve other models of choice. To do so, simply create a new directory containing model configurations under experiments-driver/configs/model-configs. Assuming the node model is called new_model
, create a file model-config.yaml
inside experiments-driver/configs/model-configs/new_model
and provide the following information about the model:
model.name: new_model
input.shape: 1, 1 # the input shape of new_model
input.name: input_1 # the name of the input for the model
model.path.dl4j: path/to/dl4j/new_model
model.path.onnx: path/to/onnx/new_model
model.path.tf-savedmodel: path/to/tf-savedmodel/new_model
model.path.torch_jit: path/to/torch/new_model
model.path.torchserve: torchserve:endpoint/newmodel
model.path.tf-serving: tfserving:endpoint/newmodel
These configurations will be passed to the adapter corresponding to the chosen stream processor.