heron/proto/ckptmgr.proto

//  Licensed to the Apache Software Foundation (ASF) under one
//  or more contributor license agreements.  See the NOTICE file
//  distributed with this work for additional information
//  regarding copyright ownership.  The ASF licenses this file
//  to you under the Apache License, Version 2.0 (the
//  "License"); you may not use this file except in compliance
//  with the License.  You may obtain a copy of the License at
//
//    http://www.apache.org/licenses/LICENSE-2.0
//
//  Unless required by applicable law or agreed to in writing,
//  software distributed under the License is distributed on an
//  "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
//  KIND, either express or implied.  See the License for the
//  specific language governing permissions and limitations
//  under the License.

syntax = "proto2";
package heron.proto.ckptmgr;
option java_package = "org.apache.heron.proto.ckptmgr";
option java_outer_classname = "CheckpointManager";

import "common.proto";
import "physical_plan.proto";

// The messages in this file are for stateful processing and providing
// effectively once semantics. Below are the various sequence of events that
// happen in a few(non exhaustive) scenarios
//
// On Startup
//   For topologies needing effectively once, the startup sequence of
// tmaster/stmgr/instance is a little different. Normally, upon
// startup, all stmgrs connect to tmaster and then tmaster computes pplan
// and distributes to stmgrs, which in turn send it to all their
// local instances. As soon as instances get their pplan, they can
// start processing(spouts emitting tuples, etc). However for
// stateful topologies, this is a little different. First
// in addition to the pplan, tmaster needs to send the last
// consistent checkpoint id(if any) to recover from as well.
// Secondly instances need to recover their prev state before
// they can process any tuples.
// WRT messages below, the startup sequence is the following.
// From the StatefulConsistentCheckpoints message stored in the
// state manager, Tmaster picks the latest checkpoint_id and
// sends RestoreTopologyStateRequest to all stmgrs. Each stmgr
// does a GetInstanceStateRequest/GetInstanceStateResponse
// dance with its local checkpoint manager for all of its local
// instances to fetch the state. Upon fetch, stmgr does a
// RestoreInstanceStateRequest/RestoreInstanceStateResponse
// dance with the corresponding instances to restore the
// state. Completion of which, stmgr sends out the 
// RestoreTopologyStateResponse message to the tmaster.
// Once the tmaster receives this from all stmgrs, it sends
// out the StartStmgrStatefulProcessing message to all
// stmgrs which inturn send out StartInstanceStatefulProcessing
// to all their local instances. This gets the topology
// humming
//
// Periodic checkpoints
//  Every so often(as dictated by TOPOLOGY_STATEFUL_CHECKPOINT_INTERVAL)
// tmaster sends out StartStatefulCheckpoint message to all
// stmgrs. Each stmgr sends out InitiateStatefulCheckpoint
// message to each of its local spout instances. Because of the best effort
// nature, this is delivered as a message. After instance does its
// checkpoint, it sends StoreInstanceStateCheckpoint to stmgr to store
// its checkpoint. Stmgr passes this to ckptmgr via
// SaveInstanceStateRequest/SaveInstanceStateResponse dance with its
// local checkpoint manager. Upon getting a successful
// SaveInstanceStateResponse message, stmgr sends out a
// InstanceStateStored message notifying tmaster that the
// state of a particular instance has been saved. Parallelly
// stmgr also sends out DownstreamStatefulCheckpoint message
// to each of the downstream components to do the distributed
// Lamport style checkpointing. Meanwhile at tmaster, when it
// receives InstanceStateStored from all instances for a particular
// checkpoint_id, it has reached globally consistent checkpoint
// and it stores it in the state manager as StatefulConsistentCheckpoint
//
// Failure Recovery
//  Either when stmgr dies, or when an instance dies(which is notified
// by the stmgr using the ResetTopologyState message), tmaster initiates
// the recovery mechanism. The recovery mechanism is exactly the same
// as the initial startup described above

// Message representing information about a globally consistent
// checkpoint
message StatefulConsistentCheckpoint {
  required string checkpoint_id = 1;
  required string packing_plan_id = 2;
  // One can add more meta data about this ckpt later
}

// message stored in the state manager by the tmaster
message StatefulConsistentCheckpoints {
  // An ordered list of the globally consistent checkpoints
  // that have been snapshotted and can be recovered from.
  // 0th element is the most recent one, 1st is the next recent
  // and so on
  repeated StatefulConsistentCheckpoint consistent_checkpoints = 1;
}

/*
 * tmaster <-> stmgr messages
 */

// Message sent to stmgrs by the tmaster to initiate checkpointing
message StartStatefulCheckpoint {
  required string checkpoint_id = 1;
}

// Message broadcasted to stmgr (will then be forwarded to instances) after when checkpoint becomes
// "globally consistent"
message StatefulConsistentCheckpointSaved {
  required StatefulConsistentCheckpoint consistent_checkpoint = 1;
}

// Message sent by tmaster to stmgr asking them to reset their instances
// to this checkpoint
message RestoreTopologyStateRequest {
  required string checkpoint_id = 1;
  // Every Attempt made by Tmaster to restore to a globally consistent
  // checkpoint is tagged with a unique id. The stmgrs include this in their
  // RestoreTopologyStateResponse below. This way tmaster could
  // try restoring to the same checkpoint_id multiple times and still
  // keep track of for which attempt the restore was in the responses
  required int64 restore_txid = 2; 
}

// Message that stmgr sends to tmaster after it restores
// all of its local instances to a checkpoint_id
message RestoreTopologyStateResponse {
  required heron.proto.system.Status status = 1;
  required string checkpoint_id = 2;
  required int64 restore_txid = 3; 
}

// Message sent by stmgr to tmaster asking it to reset the topology
// to some valid checkpoint. This is sent either if stmgr dies
// and comes back up or if an instance dies.
message ResetTopologyState {
  // Was there a dead/recovered stmgr connection that was the reason
  // for this request?
  optional string dead_stmgr = 1;
  // Was there a dead/recovered local instance connection that was the reason
  // for this request?
  optional int32 dead_taskid = 2;
  // This one describes the reason for sending the message.
  // Used primarily for debugging
  required string reason = 3;
}

// Message sent by stmgr to tmaster informing it about
// the fact that we stored a checkpoint belonging
// to the instance
message InstanceStateStored {
  required string checkpoint_id = 1;
  required heron.proto.system.Instance instance = 2;
}

// Message sent by tmaster to stmgr to start processing
// For stateful processing, all the topology components
// should be rolled back to a consistent state before they can
// start processing.
message StartStmgrStatefulProcessing {
  required string checkpoint_id = 1;
}

/*
 * tmaster -> ckptmgr messages
 */

// This is the message that a tmaster sends
// when it wants to register itself
// with checkpoint manager
message RegisterTMasterRequest {
  required string topology_name = 1;
  required string topology_id = 2;
}

// This is the message that checkpoint manager
// sends when it receives the register request
// from tmaster
message RegisterTMasterResponse {
  required heron.proto.system.Status status = 1;
}

// Message sent by tmaster to ckptmgr to cleanup
// old checkpoint state.
message CleanStatefulCheckpointRequest {
  // Any checkpoints older than this can be cleaned
  optional string oldest_checkpoint_preserved = 1;
  optional bool clean_all_checkpoints = 2;
}

// Message sent by ckptmgr to tmaster about the cleanup request
message CleanStatefulCheckpointResponse {
  required heron.proto.system.Status status = 1;
  repeated string cleaned_checkpoint_ids = 2;
}

/*
 * stmgr -> ckptmgr messages
 */

// This message encapsulates the info associated with
// state of an instance/partition. only one of the state_location
// and state should be set
message InstanceStateCheckpoint {
  required string checkpoint_id = 1;
  optional bytes state = 2;
  optional string state_location = 3;
}

// This message encapsulates the info associated with
// checkpoint metadata of a component
message CheckpointComponentMetadata {
  required string component_name = 1;
  required int32 parallelism = 2;
}

// This is the message that a stmgr sends
// when it wants to register itself with
// the checkpoint manager
message RegisterStMgrRequest {
  required string topology_name = 1;
  required string topology_id = 2;
  required string stmgr_id = 3;
  required heron.proto.system.PhysicalPlan physical_plan = 4;
}

// This is the message that checkpoint manager
// sends when it receives the register request
// from stmgr
message RegisterStMgrResponse {
  required heron.proto.system.Status status = 1;
}

// This is the request that StMgr sends to Checkpoint Mgr
// to store instance checkpoint data
message SaveInstanceStateRequest {
  // Information about the instance whose state this is
  required heron.proto.system.Instance instance = 1;
  // TODO(nwang): currently repartition is not supported and each request
  //     contains a single partition.
  required InstanceStateCheckpoint checkpoint = 2;
}

// After successfully saving state, this is what
// Checkpoint Mgr sends back as response
message SaveInstanceStateResponse {
  required heron.proto.system.Status status = 1;
  required heron.proto.system.Instance instance = 2;
  required string checkpoint_id = 3;
}

// This is the request that StMgr sends to Checkpoint Mgr
// to fetch the instance state data
message GetInstanceStateRequest {
  required heron.proto.system.Instance instance = 1;
  required string checkpoint_id = 2;
}

// This is the response that Checkpoint Mgr sends to StMgr
// for its get instance state request
message GetInstanceStateResponse {
  required heron.proto.system.Status status = 1;
  required heron.proto.system.Instance instance = 2;
  required string checkpoint_id = 3;
  optional InstanceStateCheckpoint checkpoint = 4;
}

/*
 * stmgr -> Instance messages
 */

// This is the message that the stmgr sends to its
// local tasks to begin initiating checkpoint
message InitiateStatefulCheckpoint {
  required string checkpoint_id = 1;
}

// This is the message that the instance sends
// when it wants stmgr to store its state
message StoreInstanceStateCheckpoint {
  required InstanceStateCheckpoint state = 1;
}

// This is the message that stmgr sends to its instance
// asking them to restore their state
message RestoreInstanceStateRequest {
  required InstanceStateCheckpoint state = 1;
}

// This is the message that instance sends to its stmgr
// when done with restoring the state
message RestoreInstanceStateResponse {
  required heron.proto.system.Status status = 1;
  required string checkpoint_id = 2;
}

// Message sent by stmgr to Instance to start processing
// This is related to StartStmgrStatefulProcessing message above
// but between stmgr -> instance.
message StartInstanceStatefulProcessing {
  required string checkpoint_id = 1;
}

/*
 * stmgr -> stmgr messages
 */

// Once stmgr receives StoreInstanceStateCheckpoint from an instance,
// it sends this message to all of that instance's 
// downstream components' stmgrs to propagate the checkpoint marker
message DownstreamStatefulCheckpoint {
  required int32 origin_task_id = 1; 
  required int32 destination_task_id = 2; 
  required string checkpoint_id = 3;
}