Status
...
Page properties | |
---|---|
|
...
JIRA:
Released:
...
|
Motivation
More details can be found in the Flink ML Roadmap Document and in the Flink Model Serving effort specific document.
A related discussion on the list can be found here.
Model Serving Use Cases and solution Architecture
...
This pattern fits well into the overall architecture that we want to implement.
Flink provides 2 ways of implementing low-level joins - key based join based on CoProcessFunction and partitions-based join based on RichCoFlatMapFunction. Although both can be used for required implementation, they provide different SLAs and are applicable for slightly different use cases.
...
Such simple architecture allows to implement a lot of important model serving patterns including the following:
- Guaranteed execution time. Assuming that we have several models with the fastest providing fixed execution time, it is possible to provide a model serving implementation with a fixed upper-limit on execution time, as long as that time is larger than the execution time of the simplest model
- Consensus based model serving. Assuming that we have several models, we can implement model serving where prediction is the one returned by the majority of the models.
- Quality based model serving. Assuming that we have an metric allowing us to evaluate the quality of model serving results, this approach allows us to pick the result with the best quality. It is, of course, possible to combine multiple feature, for example, consensus based model serving with the guaranteed execution time, where the consensus result is used when it completes within a given time interval.
- "Canary" deployment, where some of requests are routed to the "new" executors.
Combining Speculative execution with "real time updatable" model serving, described earlier, leads to the following overall architecture:
...
- List of available models (per data type) - List[String]
SpeculativeExecutionStats describing statistics on the speculative execution:
final case class SpeculativeExecutionStats(
var name: String,
decider : String,
tmout: Long,
since: Long = System.currentTimeMillis(),
usage: Long = 0,
duration: Double = 0.0,
min: Long = Long.MaxValue,
max: Long = Long.MinValue
Implementation
An initial implementation and examples for this Flip, they include just basic model serving implementation without speculative piece. Implementation is provided in both Scala and Java. It Implements both key-base and partition-base joins and Tensorflow support for both "optimized" and saved formats. Implementation is split into 2 part:
Library implementation
This is implementation of the base library independent from the type of the messages that particular solution is using. It is strongly typed, implemented using generics. Library implementation includes 3 modules:
Here Flink Model Serving shared contains protobuf definition (see Pipeline Metadata Definition above) And Flink model serving Java and Flink model serving Scala provides the same implementation in both Java and Scala. Theoretically it is possible to combine the two, but Java and Scala syntax is sufficiently different, so the 2 parallel implementations are provided.
In addition to this both Java and Scala implementation contain a set of unit tests for validating implementation
Flink Model Serving Java
The implementation is contained in the namespace *org.apache.flink.modelserving.java*, which contains 3 namespaces:
- model - code containing definition of model and its transformation implementation
- query - code containing base implementation for the Flink queryable state query
- server - code containing basic implementation of the Flink specific components of the overall implementation
Model implementation is split into generic and tensorflow based implementation, such implementation allows to add other model types support in the future without disrupting the base code. Generic model implementation includes the following classes:
- DataConverter - a set of model transformation methods
- DataToServe - a trait defining generic container for data used for model serving and its behavior
- Model - a trait defining generic model and its behavior (see above)
- ModelFactory - a trait defining generic model factory and its behavior (see above)
- ModelFactoriesResolver - a trait defining generic model factory resolver and its behavior. The purpose of this trait is to get the model factory based on the model type.
- ModelToServe - an intermediate representation of the model
- ModelToServeStats - model usage statistics
- ModelWithType - a combined container for model and its type used by Flink implementation
- ServingResult - generic representation of model serving result
A tensorflow namespace inside model namespace contains 4 classes:
- TensorFlowModel extends Model by adding Tensorflow specific functionality for the case of optimized Tensorflow model
- TensorBundelFlowModel extends Model by adding Tensorflow specific functionality for the case of bundled Tensorflow model
- TField a definition of the field in the tensorfow saved model
- TSignature a definition of the signature in the tensorfow saved model
Query namespace contains a single class - ModelStateQuery - implementing Flink Queryable state client for the model state
Server namespace contains 3 namespaces:
- Keyed - contains DataProcessorKeyed - implementation of the model serving using key based join (see above) and based on Flink's CoProcessFunction
- Partitioned - contains DataProcessorMap - implementation of the model serving using partion based join (see above) and based on Flink's RichCoFlatMapFunction
- Typeshema contains support classes used for type manipulation and includes the following:
- ByteArraySchema - deserialization schema for byte arrays used for reading protobuf based data from Kafka
- ModelTypeSerializer - type serializer used for checkpointing
- ModelSerializerConfigSnapshot - type serializer snapshot configuration for ModelTypeSerializer
- ModelWithTypeSerializer - type serializer used for checkpointing
- ModelWithTypeSerializerConfigSnapshot - type serializer snapshot configuration for ModelWithTypeSerializer
Flink Model Serving Scala
The implementation provides identical functionality to the Java one and is contained in the namespace org.apache.flink.modelserving.scala, which contains 3 namespaces:
- model - code containing definition of model and its transformation implementation
- query - code containing base implementation for the Flink queryable state query
- server - code containing basic implementation of the Flink specific components of the overall implementation
Model implementation is split into generic and tensorflow based implementation, such implementation allows to add other model types support in the future without disrupting the base code. Generic model implementation includes the following classes:
- DataToServe - a trait defining generic container for data used for model serving and its behavior
- Model - a trait defining generic model and its behavior (see above)
- ModelFactory - a trait defining generic model factory and its behavior (see above)
- ModelFactoryResolver - a trait defining generic model factory resolver and its behavior. The purpose of this trait is to get the model factory based on the model type.
ModelToServe defines additional case classes used for the overall implementation and a set of data transformations used for model manipulation and transforming it between different implementations. Additional classes included here include ServingResult - a container for model serving result; ModelToServeStats - model to serve statistics (see above) and ModelToServe - internal generic representation of the model
- ModelWithType - a combined container for model and its type used by Flink implementation
A tensorflow namespace inside model namespace contains 2 abstract classes:
TensorFlowModel extends Model by adding Tensorflow specific functionality for the case of optimized Tensorflow model
- TensorBundelFlowModel extends Model by adding Tensorflow specific functionality for the case of bundled Tensorflow model
Query namespace contains a single class - ModelStateQuery - implementing Flink Queryable state client for the model state
Server namespace contains 3 namespaces:
- Keyed - contains DataProcessorKeyed - implementation of the model serving using key based join (see above) and based on Flink's CoProcessFunction
- Partitioned - contains DataProcessorMap - implementation of the model serving using partion based join (see above) and based on Flink's RichCoFlatMapFunction
- Typeshema contains support classes used for type manipulation and includes the following:
- ByteArraySchema - deserialization schema for byte arrays used for reading protobuf based data from Kafka
- ModelTypeSerializer - type serializer used for checkpointing
- ModelWithTypeSerializer - type serializer used for checkpointing
Example implementation
This is implementation of the a wine quality example based on the above library. Implementation includes 3 modules:
Implementation demonstrates how to use library to build Flink Model serving implementation.
When building a new implementation you first need to define data that is used for model serving. An example is using wine quality example. Data definition is for wine is provided in model serving example shared. We are using protobuf encoding for data here, but other encoding can be used as well. Additionally Shared namespace contains implementation of embedded Kafka server (for local testing) and a Kafka provider periodically publishing both data and model, that can be used for testing the example.
There are two implementations of example - Java and Scala, that works exactly the same but are using corresponding version of Library.
Lets walk through the Scala implementation. It is located in the namespace org.apache.flink.examples.modelserving.scala and is comprised of three namespaces:
- Model
- Query
- Server
Model namespace contains three classes, extending library and implementing specific operations for a given data type.
- WineTensorFlowModel extends TensorFlowModel class by implementing processing specific to Wine quality data
- WineTensorFlowBundeledModel extends TensorFlowBundelModel class by implementing processing specific to Wine quality data
- WineFactoryResolver extends ModelFactoryResolver class by specifying above two classes as available factories
Server namespace implements 2 supporting classes : DataRecord implementing DataToServe trait for Wine type and BadDataHandler - simple data error handler implementation
It also provides complete Flink implementation for both key based join (ModelServingKeyedJob) and partition base join (ModelServingFlatJob).
To run the example first start org.apache.flink.examples.modelserving.client.client.DataProvider class, that will publish both data and model messages to Kafka (to test that publication works correctly you can use org.apache.flink.examples.modelserving.client.client.DataReader; do not forget to stop it before you proceed). Once data provider is running you can start actual model serving (ModelServingKeyedJob or ModelServingFlatJob). If you are using keyed version, you can also run ModelStateQueryJob from the query namespace to see execution statistics (Flink only supports query for keyed join). When running query, do not forget to update jobID with the actual ID of your run