Status

Current state: Accepted

Discussion thread: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-FLIP-169-DataStream-API-for-Fine-Grained-Resource-Requirements-td51071.html

JIRA:  Unable to render Jira issues macro, execution error.

Released: 1.14.0

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

In FLIP-156 and FLIP-56, we proposed fine-grained resource management to improve resource utilization in some scenarios in which coarse-grained resource management does not work well.

Currently (release-1.13), the SSG-based runtime interface and slot allocation for fine-grained resource requirements have been implemented in FLIP-156 and FLIP-56 respectively, the last piece of delivering fine-grained resource management to users is the User APIs.

In this FLIP, we propose the DataStream API for specifying fine-grained resource requirements in StreamExecutionEnvironment and #slotSharingGroup().

Note: As neither the operator nor the slot sharing group is exposed in Table / SQL API, how to configure fine-grained resource requirements in Table / SQL API is not included in the scope of this FLIP.

Public Interfaces

In this FLIP, we proposed the following two public interface changes:

• Introduce a public class SlotSharinGroup, which contains the name and the resource of a slot sharing group.
• Add setters of SlotSharinGroup in StreamExecutionEnvironment and #slotSharingGroup().

Proposed Changes

SlotSharingGroup

We propose to introduce a public immutable SlotSharingGroup with which users can describe the fine-grained resource requirements for slot sharing groups.

/** Describe the name and the the different resource factors of a slot sharing group. */
@PublicEvolving
public class SlotSharingGroup implements Serializable {

    /** Construct a new builder for the SlotSharingGroup with a specific name. */
    public static Builder newBuilder(String name);

    /** Builder for the {@link SlotSharingGroup}. */
    public static class Builder {
        /** Set the CPU cores. */
        public Builder setCpuCores(double cpuCores);

        /** Set the task heap memory. */
        public Builder setTaskHeapMemory(MemorySize taskHeapMemory);

        /** Set the task heap memory in MB. */
        public Builder setTaskHeapMemoryMB(int taskHeapMemoryMB);

        /** Set the task off-heap memory. */
        public Builder setTaskOffHeapMemory(MemorySize taskOffHeapMemory);

        /** Set the task off-heap memory in MB. */
        public Builder setOffTaskHeapMemoryMB(int taskOffHeapMemoryMB);

        /** Set the task managed memory. */
        public Builder setManagedMemory(MemorySize managedMemory);

        /** Set the task managed memory in MB. */
        public Builder setManagedMemoryMB(int managedMemoryMB);

        /**
         * Add the given external resource. The old value with the same resource name will be
         * replaced if present.
         */
        public Builder setExternalResource(String name, double value);

        /** Build the SlotSharingGroup with current factors. */
        public SlotSharingGroup build();
    }
}

Currently, the resource components to run a task include CPU cores, heap memory, off-heap memory, managed memory, network memory, and external resources. Except for the network memory, which will be automatically set by the runtime in FLINK-15031, all the resource components should be configured by users.

We proposed to introduce a public SlotSharingGroup, with contains the group name as well as the above resource components. User will construct it in a builder pattern.

DataStream API to configure resource requirements

We propose to introduce two DataStream API for configuring SlotSharingGroup: #slotSharingGroup and StreamExecutionEnvironment.

public class SingleOutputStreamOperator {
    /**
     * Sets the slot sharing group of this operation. Parallel instances of operations that are in
     * the same slot sharing group will be co-located in the same TaskManager slot, if possible.
     *
     * <p>Operations inherit the slot sharing group of input operations if all input operations are
     * in the same slot sharing group and no slot sharing group was explicitly specified.
     *
     * <p>Initially an operation is in the default slot sharing group. An operation can be put into
     * the default group explicitly by setting the slot sharing group with name {@code "default"}.
     *
     * <p>Note: the resource spec of the slot sharing group can be replaced by {@link StreamExecutionEnvironment#registerSlotSharingGroup}.
     *
     * @param slotSharingGroup The slot sharing group which contains the name and the resource spec.
     */
    @PublicEvolving
    public SingleOutputStreamOperator<T> slotSharingGroup(SlotSharingGroup slotSharingGroup);
}

public class DataStreamSink {
    /**
     * Sets the slot sharing group of this operation. Parallel instances of operations that are in
     * the same slot sharing group will be co-located in the same TaskManager slot, if possible.
     *
     * <p>Operations inherit the slot sharing group of input operations if all input operations are
     * in the same slot sharing group and no slot sharing group was explicitly specified.
     *
     * <p>Initially an operation is in the default slot sharing group. An operation can be put into
     * the default group explicitly by setting the slot sharing group with name {@code "default"}.
     *
     * <p>Note: the resource spec of the slot sharing group can be replaced by {@link StreamExecutionEnvironment#registerSlotSharingGroup}.
     *
     * @param slotSharingGroup The slot sharing group which contains the name and the resource spec.
     */
    @PublicEvolving
    public SingleOutputStreamOperator<T> slotSharingGroup(SlotSharingGroup slotSharingGroup);
}

Currently, user specifies the SSG of an operator through the #slotSharingGroup(String name) in SingleOutputStreamOperator and DataStreamSink, which only defines the name of SlotSharingGroup. We propose to add another #slotSharingGroup(SlotSharingGroup slotSharingGroup) to it, which can also define the resource spec along with its name.

Note that Flink will check that two SlotSharingGroup with the same name should not have different resource spec. An exception will be thown in violation.

To maintain the backward compatibility with the stale #slotSharingGroup(String name), which has been widely used in previous releases, we also propose a SlotSharingGroup setter in StreamExecutionEnvironment.

public class StreamExecutionEnvironment {
    /**
     * Specify fine-grained resource requirements for slot sharing groups. The existing resource
     * requirement of the same slot sharing group will be replaced.
     *
     * <p>Note that a slot sharing group hints the scheduler that the grouped operators CAN be
     * deployed into a shared slot. There's no guarantee that the scheduler always deploy the
     * grouped operators together. In cases grouped operators are deployed into separate slots, the
     * slot resources will be derived from the specified group requirements.
     */
    @PublicEvolving
    public StreamExecutionEnvironment registerSlotSharingGroup(
            String slotSharingGroup, ResourceSpec resourceSpec);
}

The StreamExecutionEnvironment is the context in which a streaming program is executed. When users want to try fine-grained resource management with their existing jobs, they do not need to touch the existing logic and just need to register a list of SlotSharingGroups into the StreamExecutionEnvironment.

Implementation Plan

Step 1. Introduce the DataStream API

Introduce the SlotSharingGroup and setters in StreamExecutionEnvironment and #slotSharingGroup as shown in section Proposed Changes.

Step 2. Document the fine-grained resource management

Document the fine-grained resource management as a beta feature. Well explains how to use it and what constraints it has.

Known Limitations

Potential resource deadlock in batch jobs

As discussed in FLIP-119, FLIP-156 and FLINK-20865, there might be a resource deadlock when applying fine-grained resource management in batch jobs with PIPELINE edges. Before we fix that issue, this case should be marked as invalid. In this case, Flink will throw an exception with a meaningful and understandable explanation, and tell them to set fine-grained.shuffle-mode.all-blocking, which is an internal config and should be removed after FLINK-20865, to true, with which Flink will force all the edges in this scenario to BLOCKING in compiling stage.

Rejected Alternatives

There is also one reject alternative, ExecutionConfig, where the resource configuration can be located.

ExecutionConfig

The ExecutionConfig includes the configs to define the behavior of the program execution, e.g. the default parallelism, retry strategy in failure, etc.

Adding the DataStream API in ExecutionConfig can also gain good usability and simplify the system. However, it can lead to unnecessary network overhead. As the ExecutionConfig will be serialized into TaskDeploymentDescriptor, the SSG-resource map will also be transferred to the TaskManager with the deployment of each task. Such network overhead is unnecessary and can further aggravate the startup time of large-scale jobs.