Add flow control to SYNC_STATE_SNAPSHOT

This patch changes the peer snapshot retrieval
logic from a non-blocking channel write to a
blocking channel write with a timeout.  This allows
gRPC's own buffering mechanisms to apply back pressure
when sending large state snapshots.

Additionally, this change will suppress some of the
spammy log messages which can flood logs when the
correlation id gets out of sync.

https://jira.hyperledger.org/browse/FAB-380

Change-Id: Icc3d37f2d161f6ac0bae984ca43e2286a45fbb3d
Signed-off-by: Jason Yellick <[email protected]>

Author: Jason Yellick

bddtests/bdd-docker/docker-compose-4-consensus-batch-nosnapshotbuffer.yml

@@ -0,0 +1,18 @@
+vp0:
+  environment:
+    # The combination of the following two environment variables ensures that a state snapshot will be pulled
+    # and that the state snapshot buffer will be exhausted
+    - CORE_PEER_SYNC_STATE_SNAPSHOT_CHANNELSIZE=0
+    - CORE_STATETRANSFER_MAXDELTAS=1
+vp1:
+  environment:
+    - CORE_PEER_SYNC_STATE_SNAPSHOT_CHANNELSIZE=0
+    - CORE_STATETRANSFER_MAXDELTAS=1
+vp2:
+  environment:
+    - CORE_PEER_SYNC_STATE_SNAPSHOT_CHANNELSIZE=0
+    - CORE_STATETRANSFER_MAXDELTAS=1
+vp3:
+  environment:
+    - CORE_PEER_SYNC_STATE_SNAPSHOT_CHANNELSIZE=0
+    - CORE_STATETRANSFER_MAXDELTAS=1

bddtests/peer_basic.feature

@@ -505,9 +505,11 @@ Feature: Network of Peers
         |   docker-compose-4-consensus-batch.yml   |      60      |
 
 
+    #@doNotDecompose
+    #@wip
     @issue_680
+    @fab380
     Scenario Outline: chaincode example02 with 4 peers and 1 membersrvc, issue #680 (State transfer)
-
         Given we compose "<ComposeFile>"
         And I register with CA supplying username "binhn" and secret "7avZQLwcUe9q" on peers:
             | vp0  |
@@ -579,8 +581,9 @@ Feature: Network of Peers
 
 
     Examples: Consensus Options
-        |          ComposeFile                       |   WaitTime   |
-        |   docker-compose-4-consensus-batch.yml     |      60      |
+        |          ComposeFile                                                                            |   WaitTime   |
+        |   docker-compose-4-consensus-batch.yml                                                          |      60      |
+        |   docker-compose-4-consensus-batch.yml docker-compose-4-consensus-batch-nosnapshotbuffer.yml    |      60      |
 
 
     @issue_724

core/peer/handler.go

@@ -29,6 +29,8 @@ import (
 	pb "github.com/hyperledger/fabric/protos"
 )
 
+const DefaultSyncSnapshotTimeout time.Duration = 60 * time.Second
+
 // Handler peer handler implementation.
 type Handler struct {
 	chatMutex                     sync.Mutex
@@ -43,6 +45,8 @@ type Handler struct {
 	snapshotRequestHandler        *syncStateSnapshotRequestHandler
 	syncStateDeltasRequestHandler *syncStateDeltasHandler
 	syncBlocksRequestHandler      *syncBlocksRequestHandler
+	syncSnapshotTimeout           time.Duration
+	lastIgnoredSnapshotCID        *uint64
 }
 
 // NewPeerHandler returns a new Peer handler
@@ -56,6 +60,12 @@ func NewPeerHandler(coord MessageHandlerCoordinator, stream ChatStream, initiate
 	}
 	d.doneChan = make(chan struct{})
 
+	if dur := viper.GetDuration("peer.sync.state.snapshot.writeTimeout"); dur == 0 {
+		d.syncSnapshotTimeout = DefaultSyncSnapshotTimeout
+	} else {
+		d.syncSnapshotTimeout = dur
+	}
+
 	d.snapshotRequestHandler = newSyncStateSnapshotRequestHandler()
 	d.syncStateDeltasRequestHandler = newSyncStateDeltasHandler()
 	d.syncBlocksRequestHandler = newSyncBlocksRequestHandler()
@@ -494,22 +504,26 @@ func (d *Handler) beforeSyncStateSnapshot(e *fsm.Event) {
 			peerLogger.Errorf("Error sending syncStateSnapshot to channel: %v", x)
 		}
 	}()
-	// Use non-blocking send, will WARN and close channel if missed message.
+	// Use blocking send and timeout, will WARN and close channel if write times out
 	d.snapshotRequestHandler.Lock()
 	defer d.snapshotRequestHandler.Unlock()
+	timer := time.NewTimer(d.syncSnapshotTimeout)
 	// Make sure the correlationID matches
 	if d.snapshotRequestHandler.shouldHandle(syncStateSnapshot.Request.CorrelationId) {
 		select {
 		case d.snapshotRequestHandler.channel <- syncStateSnapshot:
-		default:
+		case <-timer.C:
 			// Was not able to write to the channel, in which case the Snapshot stream is incomplete, and must be discarded, closing the channel
 			// without sending the terminating message which would have had an empty byte slice.
-			peerLogger.Warningf("Did NOT send SyncStateSnapshot message to channel for correlationId = %d, sequence = %d, closing channel as the message has been discarded", syncStateSnapshot.Request.CorrelationId, syncStateSnapshot.Sequence)
+			peerLogger.Warningf("Did NOT send SyncStateSnapshot message to channel for correlationId = %d, sequence = %d because we timed out reading, closing channel as the message has been discarded", syncStateSnapshot.Request.CorrelationId, syncStateSnapshot.Sequence)
 			d.snapshotRequestHandler.reset()
 		}
 	} else {
-		//Ignore the message, does not match the current correlationId
-		peerLogger.Warningf("Ignoring SyncStateSnapshot message with correlationId = %d, sequence = %d, as current correlationId = %d", syncStateSnapshot.Request.CorrelationId, syncStateSnapshot.Sequence, d.snapshotRequestHandler.correlationID)
+		if d.lastIgnoredSnapshotCID == nil || *d.lastIgnoredSnapshotCID < syncStateSnapshot.Request.CorrelationId {
+			peerLogger.Warningf("Ignoring SyncStateSnapshot message with correlationId = %d, sequence = %d, as current correlationId = %d, future messages for this (and older ids) will be suppressed", syncStateSnapshot.Request.CorrelationId, syncStateSnapshot.Sequence, d.snapshotRequestHandler.correlationID)
+			d.lastIgnoredSnapshotCID = &syncStateSnapshot.Request.CorrelationId
+			//Ignore the message, does not match the current correlationId
+		}
 	}
 }
 

core/peer/statetransfer/statetransfer.go

@@ -410,7 +410,7 @@ func (sts *coordinatorImpl) syncBlocks(highBlock, lowBlock uint64, highHash []by
 
 func (sts *coordinatorImpl) syncBlockchainToTarget(blockSyncReq *blockSyncReq) {
 
-	logger.Debugf("Processing a blockSyncReq to block %d", blockSyncReq.blockNumber)
+	logger.Debugf("Processing a blockSyncReq to block %d through %d", blockSyncReq.blockNumber, blockSyncReq.reportOnBlock)
 
 	blockchainSize := sts.stack.GetBlockchainSize()
 

peer/core.yaml

@@ -111,9 +111,13 @@ peer:
             snapshot:
                 # Channel size for readonly syncStateSnapshot messages channel
                 # for receiving state deltas for snapshot from oppositie Peer Endpoints.
-                # NOTE: currently messages are not stored and forwarded, but
-                # rather lost if the channel write blocks.
+                # NOTE: when the channel is exhausted, the writes block for up to the
+                # writeTimeout specified below
                 channelSize: 50
+                # Write timeout for the syncStateSnapshot messages
+                # When the channel above is exhausted, messages block before being
+                # discarded for this amount of time
+                writeTimeout: 60s
             deltas:
                 # Channel size for readonly syncStateDeltas messages channel for
                 # receiving state deltas for a syncBlockRange from oppositie