foundationdb/fdbserver/OldTLogServer.actor.cpp

/*
 * OldTLogServer.actor.cpp
 *
 * This source file is part of the FoundationDB open source project
 *
 * Copyright 2013-2018 Apple Inc. and the FoundationDB project authors
 * 
 * Licensed 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.
 */

#include "flow/actorcompiler.h"
#include "flow/Hash3.h"
#include "flow/Stats.h"
#include "flow/UnitTest.h"
#include "fdbclient/NativeAPI.h"
#include "fdbclient/KeyRangeMap.h"
#include "fdbclient/SystemData.h"
#include "WorkerInterface.h"
#include "TLogInterface.h"
#include "flow/Notified.h"
#include "Knobs.h"
#include "IKeyValueStore.h"
#include "flow/ActorCollection.h"
#include "fdbrpc/FailureMonitor.h"
#include "IDiskQueue.h"
#include "fdbrpc/sim_validation.h"
#include "ServerDBInfo.h"
#include "LogSystem.h"
#include "WaitFailure.h"

using std::pair;
using std::make_pair;
using std::min;
using std::max;

namespace oldTLog {
	struct TLogQueueEntryRef {
		Version version;
		Version knownCommittedVersion;
		StringRef messages;
		VectorRef< TagMessagesRef > tags;

		TLogQueueEntryRef() : version(0), knownCommittedVersion(0) {}
		TLogQueueEntryRef(Arena &a, TLogQueueEntryRef const &from)
		  : version(from.version), knownCommittedVersion(from.knownCommittedVersion), messages(a, from.messages), tags(a, from.tags) {
		}

		template <class Ar>
		void serialize(Ar& ar) {
			if( ar.protocolVersion() >= 0x0FDB00A460010001) {
				ar & version & messages & tags & knownCommittedVersion;
			} else if(ar.isDeserializing) {
				ar & version & messages & tags;
				knownCommittedVersion = 0;
			}
		}
		size_t expectedSize() const {
			return messages.expectedSize() + tags.expectedSize();
		}
	};

	typedef Standalone<TLogQueueEntryRef> TLogQueueEntry;

	struct TLogQueue : public IClosable {
	public:
		TLogQueue( IDiskQueue* queue, UID dbgid ) : queue(queue), debugNextReadVersion(1), dbgid(dbgid) {}

		// Each packet in the queue is
		//    uint32_t payloadSize
		//    uint8_t payload[payloadSize]  (begins with uint64_t protocolVersion via IncludeVersion)
		//    uint8_t validFlag

		// TLogQueue is a durable queue of TLogQueueEntry objects with an interface similar to IDiskQueue

		// TLogQueue pushes (but not commits) are atomic - after commit fails to return, a prefix of entire calls to push are durable.  This is
		//    implemented on top of the weaker guarantee of IDiskQueue::commit (that a prefix of bytes is durable) using validFlag and by
		//    padding any incomplete packet with zeros after recovery.

		// Before calling push, pop, or commit, the user must call readNext() until it throws
		//    end_of_stream(). It may not be called again thereafter.
		Future<TLogQueueEntry> readNext() {
			return readNext( this );
		}

		void push( TLogQueueEntryRef const& qe ) {
			ASSERT( version_location.empty() || version_location.lastItem()->key < qe.version );
			BinaryWriter wr( Unversioned() );  // outer framing is not versioned
			wr << uint32_t(0);
			IncludeVersion().write(wr);  // payload is versioned
			wr << qe;
			wr << uint8_t(1);
			*(uint32_t*)wr.getData() = wr.getLength() - sizeof(uint32_t) - sizeof(uint8_t);
			auto loc = queue->push( wr.toStringRef() );
			//TraceEvent("TLogQueueVersionWritten", dbgid).detail("Size", wr.getLength() - sizeof(uint32_t) - sizeof(uint8_t)).detail("Loc", loc);
			version_location[qe.version] = loc;
		}
		void pop( Version upTo ) {
			// Keep only the given and all subsequent version numbers
			// Find the first version >= upTo
			auto v = version_location.lower_bound(upTo);
			if (v == version_location.begin()) return;

			if(v == version_location.end()) {
				v = version_location.lastItem();
			}
			else {
				v.decrementNonEnd();
			}

			queue->pop( v->value );
			version_location.erase( version_location.begin(), v );  // ... and then we erase that previous version and all prior versions
		}
		Future<Void> commit() { return queue->commit(); }

		// Implements IClosable
		virtual Future<Void> getError() { return queue->getError(); }
		virtual Future<Void> onClosed() { return queue->onClosed(); }
		virtual void dispose() { queue->dispose(); delete this; }
		virtual void close() { queue->close(); delete this; }

	private:
		IDiskQueue* queue;
		Map<Version, IDiskQueue::location> version_location;  // For the version of each entry that was push()ed, the end location of the serialized bytes
		Version debugNextReadVersion;
		UID dbgid;

		ACTOR static Future<TLogQueueEntry> readNext( TLogQueue* self ) {
			state TLogQueueEntry result;
			state int zeroFillSize = 0;

			loop {
				Standalone<StringRef> h = wait( self->queue->readNext( sizeof(uint32_t) ) );
				if (h.size() != sizeof(uint32_t)) {
					if (h.size()) {
						TEST( true );  // Zero fill within size field
						int payloadSize = 0;
						memcpy(&payloadSize, h.begin(), h.size());
						zeroFillSize = sizeof(uint32_t)-h.size(); // zero fill the size itself
						zeroFillSize += payloadSize+1;  // and then the contents and valid flag
					}
					break;
				}

				state uint32_t payloadSize = *(uint32_t*)h.begin();
				ASSERT( payloadSize < (100<<20) );

				Standalone<StringRef> e = wait( self->queue->readNext( payloadSize+1 ) );
				if (e.size() != payloadSize+1) {
					TEST( true ); // Zero fill within payload
					zeroFillSize = payloadSize+1 - e.size();
					break;
				}

				if (e[payloadSize]) {
					Arena a = e.arena();
					ArenaReader ar( a, e.substr(0, payloadSize), IncludeVersion() );
					ar >> result;
					ASSERT( result.version >= self->debugNextReadVersion );
					self->debugNextReadVersion = result.version + 1;
					self->version_location[result.version] = self->queue->getNextReadLocation();
					return result;
				}
			}
			if (zeroFillSize) {
				TEST( true );  // Fixing a partial commit at the end of the tlog queue
				for(int i=0; i<zeroFillSize; i++)
					self->queue->push( StringRef((const uint8_t*)"",1) );
			}
			throw end_of_stream();
		}
	};

	struct LengthPrefixedStringRef {
		// Represents a pointer to a string which is prefixed by a 4-byte length
		// A LengthPrefixedStringRef is only pointer-sized (8 bytes vs 12 bytes for StringRef), but the corresponding string is 4 bytes bigger, and
		// substring operations aren't efficient as they are with StringRef.  It's a good choice when there might be lots of references to the same
		// exact string.

		uint32_t* length;

		StringRef toStringRef() const { ASSERT(length); return StringRef( (uint8_t*)(length+1), *length ); }
		int expectedSize() const { ASSERT(length); return *length; }
		uint32_t* getLengthPtr() const { return length; }

		LengthPrefixedStringRef() : length(NULL) {}
		LengthPrefixedStringRef(uint32_t* length) : length(length) {}
	};

	template<class T>
	struct CompareFirst {
		bool operator() (T const& lhs, T const& rhs) const {
			return lhs.first < rhs.first;
		}
	};

	struct TLogData : NonCopyable {
		struct TagData {
			std::deque<std::pair<Version, LengthPrefixedStringRef>> version_messages;
			bool nothing_persistent;				// true means tag is *known* to have no messages in persistentData.  false means nothing.
			bool popped_recently;					// `popped` has changed since last updatePersistentData
			Version popped;				// see popped version tracking contract below
			bool update_version_sizes;

			TagData( Version popped, bool nothing_persistent, bool popped_recently, Tag tag ) : nothing_persistent(nothing_persistent), popped(popped), popped_recently(popped_recently), update_version_sizes(tag != txsTag) {}

			TagData(TagData&& r) noexcept(true) : version_messages(std::move(r.version_messages)), nothing_persistent(r.nothing_persistent), popped_recently(r.popped_recently), popped(r.popped), update_version_sizes(r.update_version_sizes) {}
			void operator= (TagData&& r) noexcept(true) {
				version_messages = std::move(r.version_messages);
				nothing_persistent = r.nothing_persistent;
				popped_recently = r.popped_recently;
				popped = r.popped;
				update_version_sizes = r.update_version_sizes;
			}

			// Erase messages not needed to update *from* versions >= before (thus, messages with toversion <= before)
			ACTOR Future<Void> eraseMessagesBefore( TagData *self, Version before, Counter* bytesErased, TLogData *tlogData, int taskID ) {
				while(!self->version_messages.empty() && self->version_messages.front().first < before) {
					Version version = self->version_messages.front().first;
					std::pair<int, int> &sizes = tlogData->version_sizes[version];
					int64_t messagesErased = 0;

					while(!self->version_messages.empty() && self->version_messages.front().first == version) {
						auto const& m = self->version_messages.front();
						++messagesErased;

						if(self->update_version_sizes) {
							sizes.first -= m.second.expectedSize();
						}

						self->version_messages.pop_front();
					}

					*bytesErased += (messagesErased * sizeof(std::pair<Version, LengthPrefixedStringRef>) * SERVER_KNOBS->VERSION_MESSAGES_OVERHEAD_FACTOR_1024THS) >> 10;
					Void _ = wait(yield(taskID));
				}

				return Void();
			}

			Future<Void> eraseMessagesBefore(Version before, Counter* bytesErased, TLogData *tlogData, int taskID) {
				return eraseMessagesBefore(this, before, bytesErased, tlogData, taskID);
			}
		};

		/*
		Popped version tracking contract needed by log system to implement ILogCursor::popped():

		 - Log server tracks for each (possible) tag a popped_version
			Impl: TagData::popped (in memory) and persistTagPoppedKeys (in persistentData)
		 - popped_version(tag) is <= the maximum version for which log server (or a predecessor) is ever asked to pop the tag
			Impl: Only increased by tLogPop() in response to either a pop request or recovery from a predecessor
		 - popped_version(tag) is > the maximum version for which log server is unable to peek messages due to previous pops (on this server or a predecessor)
			Impl: Increased by tLogPop() atomically with erasing messages from memory; persisted by updatePersistentData() atomically with erasing messages from store; messages are not erased from queue where popped_version is not persisted
		 - LockTLogReply returns all tags which either have messages, or which have nonzero popped_versions
			Impl: tag_data is present for all such tags
		 - peek(tag, v) returns the popped_version for tag if that is greater than v
			Impl: Check tag_data->popped (after all waits)
		*/

		struct peekTrackerData {
			std::map<int, Promise<Version>> sequence_version;
			double lastUpdate;
		};

		std::map<UID, peekTrackerData> peekTracker;

		UID dbgid;
		bool coreStarted;
		bool stopped;
		DBRecoveryCount recoveryCount;

		IKeyValueStore* persistentData;
		IDiskQueue* rawPersistentQueue;
		TLogQueue *persistentQueue;
		VersionMetricHandle persistentDataVersion, persistentDataDurableVersion;  // The last version number in the portion of the log (written|durable) to persistentData
		NotifiedVersion version, queueCommittedVersion, queueCommitEnd;
		Version queueCommitBegin, queueCommittingVersion;
		int64_t diskQueueCommitBytes;
		AsyncVar<bool> largeDiskQueueCommitBytes; //becomes true when diskQueueCommitBytes is greater than MAX_QUEUE_COMMIT_BYTES
		Version prevVersion, knownCommittedVersion;

		Deque<std::pair<Version, Standalone<VectorRef<uint8_t>>>> messageBlocks;
		Map< Tag, TagData > tag_data;

		Map<Version, std::pair<int,int>> version_sizes;

		int64_t instanceID;
		CounterCollection cc;
		Counter bytesInput;
		Counter bytesDurable;

		Reference<AsyncVar<ServerDBInfo>> dbInfo;
		Future<Void> updatePersist; //SOMEDAY: integrate the recovery and update storage so that only one of them is committing to persistant data.
		PromiseStream<Future<Void>> addActor;

		TLogData(UID dbgid, IKeyValueStore* persistentData, IDiskQueue * persistentQueue, Reference<AsyncVar<ServerDBInfo>> const& dbInfo)
			: dbgid(dbgid),
			  persistentData(persistentData), rawPersistentQueue(persistentQueue), persistentQueue(new TLogQueue(persistentQueue, dbgid)),
			  prevVersion(0), knownCommittedVersion(0),
			  dbInfo(dbInfo),
			  updatePersist(Void()),
			  instanceID(g_random->randomUniqueID().first()),
			  cc("TLog", dbgid.toString()),
			  bytesInput("bytesInput", cc),
			  bytesDurable("bytesDurable", cc),
			  // These are initialized differently on init() or recovery
			  recoveryCount(), coreStarted(false), stopped(false), queueCommitBegin(0), queueCommitEnd(0), diskQueueCommitBytes(0), largeDiskQueueCommitBytes(false), queueCommittingVersion(0)
		{
			persistentDataVersion.init(LiteralStringRef("TLog.PersistentDataVersion"), cc.id);
			persistentDataDurableVersion.init(LiteralStringRef("TLog.PersistentDataDurableVersion"), cc.id);
			version.initMetric(LiteralStringRef("TLog.Version"), cc.id);
			queueCommittedVersion.initMetric(LiteralStringRef("TLog.QueueCommittedVersion"), cc.id);

			specialCounter(cc, "version", [this](){ return this->version.get(); });
			specialCounter(cc, "kvstoreBytesUsed", [this](){ return this->persistentData->getStorageBytes().used; });
			specialCounter(cc, "kvstoreBytesFree", [this](){ return this->persistentData->getStorageBytes().free; });
			specialCounter(cc, "kvstoreBytesAvailable", [this](){ return this->persistentData->getStorageBytes().available; });
			specialCounter(cc, "kvstoreBytesTotal", [this](){ return this->persistentData->getStorageBytes().total; });
			specialCounter(cc, "queueDiskBytesUsed", [this](){ return this->rawPersistentQueue->getStorageBytes().used; });
			specialCounter(cc, "queueDiskBytesFree", [this](){ return this->rawPersistentQueue->getStorageBytes().free; });
			specialCounter(cc, "queueDiskBytesAvailable", [this](){ return this->rawPersistentQueue->getStorageBytes().available; });
			specialCounter(cc, "queueDiskBytesTotal", [this](){ return this->rawPersistentQueue->getStorageBytes().total; });
		}

		LogEpoch epoch() const { return recoveryCount; }
	};

	ACTOR Future<Void> tLogLock( TLogData* self, ReplyPromise< TLogLockResult > reply ) {
		state Version stopVersion = self->version.get();

		TEST(true); // TLog stopped by recovering master
		TEST( self->stopped );
		TEST( !self->stopped );

		TraceEvent("TLogStop", self->dbgid).detail("Ver", stopVersion).detail("isStopped", self->stopped);

		self->stopped = true;

		// Lock once the current version has been committed
		Void _ = wait( self->queueCommittedVersion.whenAtLeast( stopVersion ) );

		ASSERT(stopVersion == self->version.get());

		TLogLockResult result;
		result.end = stopVersion;
		result.knownCommittedVersion = self->knownCommittedVersion;
		for( auto & tag : self->tag_data )
			result.tags.push_back( tag.key );

		reply.send( result );
		return Void();
	}

	KeyRange prefixRange( KeyRef prefix ) {
		Key end = prefix;
		UNSTOPPABLE_ASSERT( end.size() && end.end()[-1] != 0xFF );
		++const_cast<uint8_t&>( end.end()[-1] );
		return KeyRangeRef( prefix, end );
	}

	////// Persistence format (for self->persistentData)

	// Immutable keys
	static const KeyValueRef persistFormat( LiteralStringRef( "Format" ), LiteralStringRef("FoundationDB/LogServer/2/2") );
	static const KeyRangeRef persistFormatReadableRange( LiteralStringRef("FoundationDB/LogServer/2/2"), LiteralStringRef("FoundationDB/LogServer/2/3") );
	static const KeyRef persistID = LiteralStringRef( "ID" );
	static const KeyRef persistRecoveryCountKey = LiteralStringRef("DbRecoveryCount");

	// Updated on updatePersistentData()
	static const KeyRef persistCurrentVersionKey = LiteralStringRef("version");
	static const KeyRange persistTagMessagesKeys = prefixRange(LiteralStringRef("TagMsg/"));
	static const KeyRange persistTagPoppedKeys = prefixRange(LiteralStringRef("TagPop/"));

	// Only present during network recovery process
	static const KeyValueRef persistRecoveryInProgress( LiteralStringRef("RecoveryInProgress"), LiteralStringRef("1") );

	static Key persistTagMessagesKey( Tag tag, Version version ) {
		BinaryWriter wr( Unversioned() );
		wr.serializeBytes(persistTagMessagesKeys.begin);
		wr << tag;
		wr << bigEndian64( version );
		return wr.toStringRef();
	}

	static Key persistTagPoppedKey( Tag tag ) {
		BinaryWriter wr(Unversioned());
		wr.serializeBytes( persistTagPoppedKeys.begin );
		wr << tag;
		return wr.toStringRef();
	}

	static Value persistTagPoppedValue( Version popped ) {
		return BinaryWriter::toValue( popped, Unversioned() );
	}

	static Tag decodeTagPoppedKey( KeyRef key ) {
		Tag s;
		BinaryReader rd( key.removePrefix(persistTagPoppedKeys.begin), Unversioned() );
		rd >> s;
		return s;
	}

	static Version decodeTagPoppedValue( ValueRef value ) {
		return BinaryReader::fromStringRef<Version>( value, Unversioned() );
	}

	static StringRef stripTagMessagesKey( StringRef key ) {
		return key.substr( sizeof(Tag) + persistTagMessagesKeys.begin.size() );
	}

	static Version decodeTagMessagesKey( StringRef key ) {
		return bigEndian64( BinaryReader::fromStringRef<Version>( stripTagMessagesKey(key), Unversioned() ) );
	}

	static Standalone<StringRef> decodeTagMessagesKeyTag( StringRef key ) {
		key = key.removePrefix( persistTagMessagesKeys.begin );  // <tag>\x00\xff<version>
		BinaryWriter wr( Unversioned() );
		for(auto c = key.begin(); c != key.end(); ++c) {
			if (*c)
				wr << *c;
			else {
				ASSERT( c+1 != key.end() );
				if (c[1] == 0xff) {
					wr << uint8_t(0);
					c++;
				} else if (c[1] == 0)
					break;
				else
					throw internal_error();
			}
		}
		return wr.toStringRef();
	}

	void validate( TLogData* self, bool force = false ) {
	}

	void updatePersistentPopped( TLogData* self, Tag tag, TLogData::TagData& data ) {
		if (!data.popped_recently) return;
		self->persistentData->set(KeyValueRef( persistTagPoppedKey(tag), persistTagPoppedValue(data.popped) ));
		data.popped_recently = false;

		if (data.nothing_persistent) return;

		self->persistentData->clear( KeyRangeRef(
			persistTagMessagesKey( tag, Version(0) ),
			persistTagMessagesKey( tag, data.popped ) ) );
		if (data.popped > self->persistentDataVersion)
			data.nothing_persistent = true;
		//TraceEvent("TLogPopWrite", self->dbgid).detail("Tag", tag).detail("To", data.popped);
	}

	ACTOR Future<Void> updatePersistentData( TLogData* self, Version newPersistentDataVersion ) {
		// PERSIST: Changes self->persistentDataVersion and writes and commits the relevant changes
		ASSERT( newPersistentDataVersion <= self->version.get() );
		ASSERT( newPersistentDataVersion <= self->queueCommittedVersion.get() );
		ASSERT( newPersistentDataVersion > self->persistentDataVersion );
		ASSERT( self->persistentDataVersion == self->persistentDataDurableVersion );

		//TraceEvent("updatePersistentData", self->dbgid).detail("seq", newPersistentDataSeq);

		state bool anyData = false;
		state Map<Tag, TLogData::TagData>::iterator tag;
		// For all existing tags
		for(tag = self->tag_data.begin(); tag != self->tag_data.end(); ++tag) {
			state Version currentVersion = 0;
			// Clear recently popped versions from persistentData if necessary
			updatePersistentPopped( self, tag->key, tag->value );
			// Transfer unpopped messages with version numbers less than newPersistentDataVersion to persistentData
			state std::deque<std::pair<Version, LengthPrefixedStringRef>>::iterator msg = tag->value.version_messages.begin();
			while(msg != tag->value.version_messages.end() && msg->first <= newPersistentDataVersion) {
				currentVersion = msg->first;
				anyData = true;
				tag->value.nothing_persistent = false;
				BinaryWriter wr( Unversioned() );

				for(; msg != tag->value.version_messages.end() && msg->first == currentVersion; ++msg)
					wr << msg->second.toStringRef();

				self->persistentData->set( KeyValueRef( persistTagMessagesKey( tag->key, currentVersion ), wr.toStringRef() ) );

				Future<Void> f = yield(TaskUpdateStorage);
				if(!f.isReady()) {
					Void _ = wait(f);
					msg = std::upper_bound(tag->value.version_messages.begin(), tag->value.version_messages.end(), std::make_pair(currentVersion, LengthPrefixedStringRef()), CompareFirst<std::pair<Version, LengthPrefixedStringRef>>());
				}
			}

			Void _ = wait(yield(TaskUpdateStorage));
		}

		self->persistentData->set( KeyValueRef( persistCurrentVersionKey, BinaryWriter::toValue(newPersistentDataVersion, Unversioned()) ) );
		self->persistentDataVersion = newPersistentDataVersion;

		Void _ = wait( self->persistentData->commit() ); // SOMEDAY: This seems to be running pretty often, should we slow it down???
		Void _ = wait( delay(0, TaskUpdateStorage) );

		// Now that the changes we made to persistentData are durable, erase the data we moved from memory and the queue, increase bytesDurable accordingly, and update persistentDataDurableVersion.

		TEST(anyData);  // TLog moved data to persistentData
		self->persistentDataDurableVersion = newPersistentDataVersion;

		for(tag = self->tag_data.begin(); tag != self->tag_data.end(); ++tag) {
			Void _ = wait(tag->value.eraseMessagesBefore( newPersistentDataVersion+1, &self->bytesDurable, self, TaskUpdateStorage ));
			Void _ = wait(yield(TaskUpdateStorage));
		}

		self->version_sizes.erase(self->version_sizes.begin(), self->version_sizes.lower_bound(self->persistentDataDurableVersion));

		Void _ = wait(yield(TaskUpdateStorage));

		while(!self->messageBlocks.empty() && self->messageBlocks.front().first <= newPersistentDataVersion) {
			self->bytesDurable += self->messageBlocks.front().second.size() * SERVER_KNOBS->TLOG_MESSAGE_BLOCK_OVERHEAD_FACTOR;
			self->messageBlocks.pop_front();
			Void _ = wait(yield(TaskUpdateStorage));
		}

		ASSERT(self->bytesDurable.getValue() <= self->bytesInput.getValue());

		if( self->queueCommitEnd.get() > 0 )
			self->persistentQueue->pop( newPersistentDataVersion+1 ); // SOMEDAY: this can cause a slow task (~0.5ms), presumably from erasing too many versions. Should we limit the number of versions cleared at a time?

		return Void();
	}

	// This function (and updatePersistentData, which is called by this function) run at a low priority and can soak up all CPU resources.
	// For this reason, they employ aggressive use of yields to avoid causing slow tasks that could introduce latencies for more important
	// work (e.g. commits).
	ACTOR Future<Void> updateStorage( TLogData* self ) {
		Void _ = wait(delay(0, TaskUpdateStorage));
		loop {
			state Version prevVersion = 0;
			state Version nextVersion = 0;
			state int totalSize = 0;

			state Map<Version, std::pair<int, int>>::iterator sizeItr = self->version_sizes.begin();
			while( totalSize < SERVER_KNOBS->UPDATE_STORAGE_BYTE_LIMIT && sizeItr != self->version_sizes.end()
					&& (self->bytesInput.getValue() - self->bytesDurable.getValue() - totalSize >= SERVER_KNOBS->TLOG_SPILL_THRESHOLD || sizeItr->value.first == 0) )
			{
				Void _ = wait( yield(TaskUpdateStorage) );

				++sizeItr;
				nextVersion = sizeItr == self->version_sizes.end() ? self->version.get() : sizeItr->key;

				state Map<Tag, TLogData::TagData>::iterator tag;
				for(tag = self->tag_data.begin(); tag != self->tag_data.end(); ++tag) {
					auto it = std::lower_bound(tag->value.version_messages.begin(), tag->value.version_messages.end(), std::make_pair(prevVersion, LengthPrefixedStringRef()), CompareFirst<std::pair<Version, LengthPrefixedStringRef>>());
					for(; it != tag->value.version_messages.end() && it->first < nextVersion; ++it) {
						totalSize += it->second.expectedSize();
					}

					Void _ = wait(yield(TaskUpdateStorage));
				}

				prevVersion = nextVersion;
			}

			nextVersion = std::max<Version>(nextVersion, self->persistentDataVersion);

			TraceEvent("UpdateStorageVer", self->dbgid).detail("nextVersion", nextVersion).detail("persistentDataVersion", self->persistentDataVersion).detail("totalSize", totalSize);

			Void _ = wait( self->queueCommittedVersion.whenAtLeast( nextVersion ) );
			Void _ = wait( delay(0, TaskUpdateStorage) );

			if (nextVersion > self->persistentDataVersion) {
				self->updatePersist = updatePersistentData(self, nextVersion);
				Void _ = wait( self->updatePersist );
			}

			if( totalSize < SERVER_KNOBS->UPDATE_STORAGE_BYTE_LIMIT ) {
				Void _ = wait( delay(BUGGIFY ? SERVER_KNOBS->BUGGIFY_TLOG_STORAGE_MIN_UPDATE_INTERVAL : SERVER_KNOBS->TLOG_STORAGE_MIN_UPDATE_INTERVAL, TaskUpdateStorage) );
			}
			else {
				//recovery wants to commit to persistant data when updatePersistentData is not active, this delay ensures that immediately after
				//updatePersist returns another one has not been started yet.
				Void _ = wait( delay(0.0, TaskUpdateStorage) );
			}
		}
	}

	void commitMessages( TLogData* self, Version version, Arena arena, StringRef messages, VectorRef< TagMessagesRef > tags) {
		// SOMEDAY: This method of copying messages is reasonably memory efficient, but it's still a lot of bytes copied.  Find a
		// way to do the memory allocation right as we receive the messages in the network layer.

		int64_t addedBytes = 0;
		int64_t expectedBytes = 0;

		if(!messages.size()) {
			return;
		}

		StringRef messages1;  // the first block of messages, if they aren't all stored contiguously.  otherwise empty

		// Grab the last block in the blocks list so we can share its arena
		// We pop all of the elements of it to create a "fresh" vector that starts at the end of the previous vector
		Standalone<VectorRef<uint8_t>> block;
		if(self->messageBlocks.empty()) {
			block = Standalone<VectorRef<uint8_t>>();
			block.reserve(block.arena(), std::max<int64_t>(SERVER_KNOBS->TLOG_MESSAGE_BLOCK_BYTES, messages.size()));
		}
		else {
			block = self->messageBlocks.back().second;
		}

		block.pop_front(block.size());

		// If the current batch of messages doesn't fit entirely in the remainder of the last block in the list
		if(messages.size() + block.size() > block.capacity()) {
			// Find how many messages will fit
			LengthPrefixedStringRef r((uint32_t*)messages.begin());
			uint8_t const* end = messages.begin() + block.capacity() - block.size();
			while(r.toStringRef().end() <= end) {
				r = LengthPrefixedStringRef( (uint32_t*)r.toStringRef().end() );
			}

			// Fill up the rest of this block
			int bytes = (uint8_t*)r.getLengthPtr()-messages.begin();
			if (bytes) {
				TEST(true); // Splitting commit messages across multiple blocks
				messages1 = StringRef(block.end(), bytes);
				block.append(block.arena(), messages.begin(), bytes);
				self->messageBlocks.push_back( std::make_pair(version, block) );
				addedBytes += int64_t(block.size()) * SERVER_KNOBS->TLOG_MESSAGE_BLOCK_OVERHEAD_FACTOR;
				messages = messages.substr(bytes);
			}

			// Make a new block
			block = Standalone<VectorRef<uint8_t>>();
			block.reserve(block.arena(), std::max<int64_t>(SERVER_KNOBS->TLOG_MESSAGE_BLOCK_BYTES, messages.size()));
		}

		// Copy messages into block
		ASSERT(messages.size() <= block.capacity() - block.size());
		block.append(block.arena(), messages.begin(), messages.size());
		self->messageBlocks.push_back( std::make_pair(version, block) );
		addedBytes += int64_t(block.size()) * SERVER_KNOBS->TLOG_MESSAGE_BLOCK_OVERHEAD_FACTOR;
		messages = StringRef(block.end()-messages.size(), messages.size());

		for(auto tag = tags.begin(); tag != tags.end(); ++tag) {
			int64_t tagMessages = 0;

			auto tsm = self->tag_data.find(tag->tag);
			if (tsm == self->tag_data.end()) {
				tsm = self->tag_data.insert( mapPair(std::move(Tag(tag->tag)), TLogData::TagData(Version(0), true, true, tag->tag) ), false );
			}

			if (version >= tsm->value.popped) {
				for(int m = 0; m < tag->messageOffsets.size(); ++m) {
					int offs = tag->messageOffsets[m];
					uint8_t const* p = offs < messages1.size() ? messages1.begin() + offs : messages.begin() + offs - messages1.size();
					tsm->value.version_messages.push_back(std::make_pair(version, LengthPrefixedStringRef((uint32_t*)p)));
					if(tsm->value.version_messages.back().second.expectedSize() > SERVER_KNOBS->MAX_MESSAGE_SIZE) {
						TraceEvent(SevWarnAlways, "LargeMessage").detail("Size", tsm->value.version_messages.back().second.expectedSize());
					}
					if (tag->tag != txsTag)
						expectedBytes += tsm->value.version_messages.back().second.expectedSize();

					++tagMessages;
				}
			}

			// The factor of VERSION_MESSAGES_OVERHEAD is intended to be an overestimate of the actual memory used to store this data in a std::deque.
			// In practice, this number is probably something like 528/512 ~= 1.03, but this could vary based on the implementation.
			// There will also be a fixed overhead per std::deque, but its size should be trivial relative to the size of the TLog
			// queue and can be thought of as increasing the capacity of the queue slightly.
			addedBytes += (tagMessages * sizeof(std::pair<Version, LengthPrefixedStringRef>) * SERVER_KNOBS->VERSION_MESSAGES_OVERHEAD_FACTOR_1024THS) >> 10;
		}

		self->version_sizes[version] = make_pair(expectedBytes, expectedBytes);
		self->bytesInput += addedBytes;

		//TraceEvent("TLogPushed", self->dbgid).detail("Bytes", addedBytes).detail("MessageBytes", messages.size()).detail("Tags", tags.size()).detail("expectedBytes", expectedBytes).detail("mCount", mCount).detail("tCount", tCount);
	}

	Version poppedVersion( TLogData* self, Tag tag) {
		auto mapIt = self->tag_data.find(tag);
		if (mapIt == self->tag_data.end())
			return Version(0);
		return mapIt->value.popped;
	}

	std::deque<std::pair<Version, LengthPrefixedStringRef>> & get_version_messages( TLogData* self, Tag tag ) {
		auto mapIt = self->tag_data.find(tag);
		if (mapIt == self->tag_data.end()) {
			static std::deque<std::pair<Version, LengthPrefixedStringRef>> empty;
			return empty;
		}
		return mapIt->value.version_messages;
	};

	ACTOR Future<Void> tLogPop( TLogData* self, TLogPopRequest req ) {
		auto ti = self->tag_data.find(req.tag);
		if (ti == self->tag_data.end()) {
			ti = self->tag_data.insert( mapPair(std::move(Tag(req.tag)), TLogData::TagData(req.to, true, true, req.tag)) );
		} else if (req.to > ti->value.popped) {
			ti->value.popped = req.to;
			ti->value.popped_recently = true;
			//if (to.epoch == self->epoch())
			if ( req.to > self->persistentDataDurableVersion )
				Void _ = wait(ti->value.eraseMessagesBefore( req.to, &self->bytesDurable, self, TaskTLogPop ));
			//TraceEvent("TLogPop", self->dbgid).detail("Tag", req.tag).detail("To", req.to);
		}

		req.reply.send(Void());
		return Void();
	}

	void peekMessagesFromMemory( TLogData* self, TLogPeekRequest const& req, BinaryWriter& messages, Version& endVersion ) {
		ASSERT( !messages.getLength() );

		auto& deque = get_version_messages(self, req.tag);
		//TraceEvent("tLogPeekMem", self->dbgid).detail("Tag", printable(req.tag1)).detail("pDS", self->persistentDataSequence).detail("pDDS", self->persistentDataDurableSequence).detail("Oldest", map1.empty() ? 0 : map1.begin()->key ).detail("OldestMsgCount", map1.empty() ? 0 : map1.begin()->value.size());

		Version begin = std::max( req.begin, self->persistentDataDurableVersion+1 );
		auto it = std::lower_bound(deque.begin(), deque.end(), std::make_pair(begin, LengthPrefixedStringRef()), CompareFirst<std::pair<Version, LengthPrefixedStringRef>>());

		Version currentVersion = -1;
		for(; it != deque.end(); ++it) {
			if(it->first != currentVersion) {
				if (messages.getLength() >= SERVER_KNOBS->DESIRED_TOTAL_BYTES) {
					endVersion = it->first;
					//TraceEvent("tLogPeekMessagesReached2", self->dbgid);
					break;
				}

				currentVersion = it->first;
				messages << int32_t(-1) << currentVersion;
			}

			messages << it->second.toStringRef();
		}
	}

	ACTOR Future<Void> tLogPeekMessages( TLogData* self, TLogPeekRequest req ) {
		state BinaryWriter messages(Unversioned());
		state BinaryWriter messages2(Unversioned());
		state int sequence = -1;
		state UID peekId;

		if(req.sequence.present()) {
			try {
				peekId = req.sequence.get().first;
				sequence = req.sequence.get().second;
				if(sequence > 0) {
					auto& trackerData = self->peekTracker[peekId];
					trackerData.lastUpdate = now();
					Version ver = wait(trackerData.sequence_version[sequence].getFuture());
					req.begin = ver;
					Void _ = wait(yield());
				}
			} catch( Error &e ) {
				if(e.code() == error_code_timed_out) {
					req.reply.sendError(timed_out());
					return Void();
				} else {
					throw;
				}
			}
		}

		if( req.returnIfBlocked && self->version.get() < req.begin ) {
			req.reply.sendError(end_of_stream());
			return Void();
		}

		//TraceEvent("tLogPeekMessages0", self->dbgid).detail("reqBeginEpoch", req.begin.epoch).detail("reqBeginSeq", req.begin.sequence).detail("epoch", self->epoch()).detail("persistentDataSeq", self->persistentDataSequence).detail("Tag1", printable(req.tag1)).detail("Tag2", printable(req.tag2));
		// Wait until we have something to return that the caller doesn't already have
		if( self->version.get() < req.begin ) {
			Void _ = wait( self->version.whenAtLeast( req.begin ) );
			Void _ = wait( delay(SERVER_KNOBS->TLOG_PEEK_DELAY, g_network->getCurrentTask()) );
		}

		state Version endVersion = self->version.get() + 1;

		//grab messages from disk
		//TraceEvent("tLogPeekMessages", self->dbgid).detail("reqBeginEpoch", req.begin.epoch).detail("reqBeginSeq", req.begin.sequence).detail("epoch", self->epoch()).detail("persistentDataSeq", self->persistentDataSequence).detail("Tag1", printable(req.tag1)).detail("Tag2", printable(req.tag2));
		if( req.begin <= self->persistentDataDurableVersion ) {
			// Just in case the durable version changes while we are waiting for the read, we grab this data from memory.  We may or may not actually send it depending on
			// whether we get enough data from disk.
			// SOMEDAY: Only do this if an initial attempt to read from disk results in insufficient data and the required data is no longer in memory
			// SOMEDAY: Should we only send part of the messages we collected, to actually limit the size of the result?

			peekMessagesFromMemory( self, req, messages2, endVersion );

			Standalone<VectorRef<KeyValueRef>> kvs = wait(
				self->persistentData->readRange(KeyRangeRef(
					persistTagMessagesKey(req.tag, req.begin),
					persistTagMessagesKey(req.tag, self->persistentDataDurableVersion + 1)), SERVER_KNOBS->DESIRED_TOTAL_BYTES, SERVER_KNOBS->DESIRED_TOTAL_BYTES));

			//TraceEvent("TLogPeekResults", self->dbgid).detail("ForAddress", req.reply.getEndpoint().address).detail("Tag1Results", s1).detail("Tag2Results", s2).detail("Tag1ResultsLim", kv1.size()).detail("Tag2ResultsLim", kv2.size()).detail("Tag1ResultsLast", kv1.size() ? printable(kv1[0].key) : "").detail("Tag2ResultsLast", kv2.size() ? printable(kv2[0].key) : "").detail("Limited", limited).detail("NextEpoch", next_pos.epoch).detail("NextSeq", next_pos.sequence).detail("NowEpoch", self->epoch()).detail("NowSeq", self->sequence.getNextSequence());

			for (auto &kv : kvs) {
				auto ver = decodeTagMessagesKey(kv.key);
				messages << int32_t(-1) << ver;
				messages.serializeBytes(kv.value);
			}

			if (kvs.expectedSize() >= SERVER_KNOBS->DESIRED_TOTAL_BYTES)
				endVersion = decodeTagMessagesKey(kvs.end()[-1].key) + 1;
			else
				messages.serializeBytes( messages2.toStringRef() );
		} else {
			peekMessagesFromMemory( self, req, messages, endVersion );
			//TraceEvent("TLogPeekResults", self->dbgid).detail("ForAddress", req.reply.getEndpoint().address).detail("MessageBytes", messages.getLength()).detail("NextEpoch", next_pos.epoch).detail("NextSeq", next_pos.sequence).detail("NowSeq", self->sequence.getNextSequence());
		}

		Version poppedVer = poppedVersion(self, req.tag);

		TLogPeekReply reply;
		reply.maxKnownVersion = self->version.get();
		if(poppedVer > req.begin) {
			reply.popped = poppedVer;
			reply.end = poppedVer;
		} else {
			reply.messages = messages.toStringRef();
			reply.end = endVersion;
		}
		//TraceEvent("TlogPeek", self->dbgid).detail("endVer", reply.end).detail("msgBytes", reply.messages.expectedSize()).detail("ForAddress", req.reply.getEndpoint().address);

		if(req.sequence.present()) {
			auto& trackerData = self->peekTracker[peekId];
			trackerData.lastUpdate = now();
			auto& sequenceData = trackerData.sequence_version[sequence+1];
			if(sequenceData.isSet()) {
				if(sequenceData.getFuture().get() != reply.end) {
					TEST(true); //tlog peek second attempt ended at a different version
					req.reply.sendError(timed_out());
					return Void();
				}
			} else {
				sequenceData.send(reply.end);
			}
		}

		req.reply.send( reply );
		return Void();
	}

	ACTOR Future<Void> doQueueCommit( TLogData* self ) {
		state Version ver = self->version.get();
		state Version commitNumber = self->queueCommitBegin+1;
		self->queueCommitBegin = commitNumber;
		self->queueCommittingVersion = ver;

		Future<Void> c = self->persistentQueue->commit();
		self->diskQueueCommitBytes = 0;
		self->largeDiskQueueCommitBytes.set(false);

		Void _ = wait(c);
		Void _ = wait(self->queueCommitEnd.whenAtLeast(commitNumber-1));

		//Calling check_yield instead of yield to avoid a destruction ordering problem in simulation
		if(g_network->check_yield(g_network->getCurrentTask())) {
			Void _ = wait(delay(0, g_network->getCurrentTask()));
		}

		ASSERT( ver > self->queueCommittedVersion.get() );

		self->queueCommittedVersion.set(ver);
		self->queueCommitEnd.set(commitNumber);

		TraceEvent("TLogCommitDurable", self->dbgid).detail("Version", ver);

		return Void();
	}

	ACTOR Future<Void> commitQueue( TLogData* self ) {
		loop {
			Void _ = wait( self->version.whenAtLeast( std::max(self->queueCommittingVersion, self->queueCommittedVersion.get()) + 1 ) );
			while( self->queueCommitBegin != self->queueCommitEnd.get() && !self->largeDiskQueueCommitBytes.get() ) {
				Void _ = wait( self->queueCommitEnd.whenAtLeast(self->queueCommitBegin) || self->largeDiskQueueCommitBytes.onChange() );
			}
			self->addActor.send(doQueueCommit(self));
		}
	}

	ACTOR Future<Void> tLogCommit(
			TLogData* self,
			TLogCommitRequest req,
			PromiseStream<Void> warningCollectorInput ) {

		state Optional<UID> tlogDebugID;
		if(req.debugID.present())
		{
			tlogDebugID = g_nondeterministic_random->randomUniqueID();
			g_traceBatch.addAttach("CommitAttachID", req.debugID.get().first(), tlogDebugID.get().first());
			g_traceBatch.addEvent("CommitDebug", tlogDebugID.get().first(), "TLog.tLogCommit.BeforeWaitForVersion");
		}

		self->knownCommittedVersion = std::max(self->knownCommittedVersion, req.knownCommittedVersion);

		Void _ = wait( self->version.whenAtLeast( req.prevVersion ) );

		//Calling check_yield instead of yield to avoid a destruction ordering problem in simulation
		if(g_network->check_yield(g_network->getCurrentTask())) {
			Void _ = wait(delay(0, g_network->getCurrentTask()));
		}

		if(self->stopped) {
			req.reply.sendError( tlog_stopped() );
			return Void();
		}

		if (self->version.get() == req.prevVersion) {  // Not a duplicate (check relies on no waiting between here and self->version.set() below!)
			if(req.debugID.present())
				g_traceBatch.addEvent("CommitDebug", tlogDebugID.get().first(), "TLog.tLogCommit.Before");

			TraceEvent("TLogCommit", self->dbgid).detail("Version", req.version);
			commitMessages(self, req.version, req.arena, req.messages, req.tags);

			// Log the changes to the persistent queue, to be committed by commitQueue()
			TLogQueueEntryRef qe;
			qe.version = req.version;
			qe.knownCommittedVersion = req.knownCommittedVersion;
			qe.messages = req.messages;
			qe.tags = req.tags;
			self->persistentQueue->push( qe );

			self->diskQueueCommitBytes += qe.expectedSize();
			if( self->diskQueueCommitBytes > SERVER_KNOBS->MAX_QUEUE_COMMIT_BYTES ) {
				self->largeDiskQueueCommitBytes.set(true);
			}

			// Notifies the commitQueue actor to commit persistentQueue, and also unblocks tLogPeekMessages actors
			self->prevVersion = self->version.get();
			self->version.set( req.version );

			if(req.debugID.present())
				g_traceBatch.addEvent("CommitDebug", tlogDebugID.get().first(), "TLog.tLogCommit.AfterTLogCommit");
		}
		// Send replies only once all prior messages have been received and committed.
		Void _ = wait( timeoutWarning( self->queueCommittedVersion.whenAtLeast( req.version ), 0.1, warningCollectorInput ) );

		if(req.debugID.present())
			g_traceBatch.addEvent("CommitDebug", tlogDebugID.get().first(), "TLog.tLogCommit.After");

		req.reply.send( Void() );
		return Void();
	}

	ACTOR Future<Void> initPersistentState( TLogData* self ) {
		// PERSIST: Initial setup of persistentData for a brand new tLog for a new database
		IKeyValueStore *storage = self->persistentData;
		storage->set( persistFormat );
		storage->set( KeyValueRef( persistID, BinaryWriter::toValue( self->dbgid, Unversioned() ) ) );
		storage->set( KeyValueRef( persistCurrentVersionKey, BinaryWriter::toValue(self->version.get(), Unversioned()) ) );
		storage->set( KeyValueRef( persistRecoveryCountKey, BinaryWriter::toValue(self->recoveryCount, Unversioned()) ) );

		TraceEvent("TLogInitCommit", self->dbgid).detail("Version", self->version.get());
		Void _ = wait( storage->commit() );
		return Void();
	}

	ACTOR Future<Void> restorePersistentState( TLogData* self, Promise<DBRecoveryCount> outRecoveryCount, bool processQueue, TLogInterface myInterface ) {
		state double startt = now();
		// PERSIST: Read basic state from persistentData; replay persistentQueue but don't erase it
		IKeyValueStore *storage = self->persistentData;

		TraceEvent("TLogRestorePersistentState", self->dbgid).detail("pq", processQueue);

		state Future<Optional<Value>> fFormat = storage->readValue(persistFormat.key);
		state Future<Optional<Value>> fID = storage->readValue(persistID);
		state Future<Optional<Value>> fVer = storage->readValue(persistCurrentVersionKey);
		state Future<Optional<Value>> fRecoveryCount = storage->readValue(persistRecoveryCountKey);
		state Future<Optional<Value>> fRecoveryInProgress = storage->readValue( persistRecoveryInProgress.key );

		// FIXME: metadata in queue?

		Void _ = wait( waitForAll( (vector<Future<Optional<Value>>>(), fFormat, fID, fVer, fRecoveryCount, fRecoveryInProgress) ) );

		if (fFormat.get().present() && !persistFormatReadableRange.contains( fFormat.get().get() )) {
			TraceEvent(SevError, "UnsupportedDBFormat", self->dbgid).detail("Format", printable(fFormat.get().get())).detail("Expected", persistFormat.value.toString());
			throw worker_recovery_failed();
		}

		if (fRecoveryInProgress.get().present()) {
			TEST(true); // We must have rebooted during network recovery; the master recovery that depended on us will fail and we can permanently delete our (incomplete) storage
			TraceEvent("RestartedDuringNetworkRecovery", self->dbgid);
			throw worker_removed();
		}

		if (!fFormat.get().present()) {
			Standalone<VectorRef<KeyValueRef>> v = wait( self->persistentData->readRange( KeyRangeRef(StringRef(), LiteralStringRef("\xff")), 1 ) );
			if (!v.size()) {
				TEST(true); // The DB is completely empty, so it was never initialized.  Delete it.
				throw worker_removed();
			} else {
				// This should never happen
				TraceEvent(SevError, "NoDBFormatKey", self->dbgid).detail("FirstKey", printable(v[0].key));
				ASSERT( false );
				throw worker_recovery_failed();
			}
		}


		ASSERT( self->dbgid == BinaryReader::fromStringRef<UID>(fID.get().get(), Unversioned()) );

		Version ver = BinaryReader::fromStringRef<Version>( fVer.get().get(), Unversioned() );
		self->persistentDataVersion = ver;
		self->persistentDataDurableVersion = ver;
		self->version.set( ver );

		TraceEvent("TLogRestorePersistentStateVer", self->dbgid).detail("ver", self->version.get());

		self->recoveryCount = BinaryReader::fromStringRef<uint64_t>( fRecoveryCount.get().get(), Unversioned() );

		outRecoveryCount.send( self->recoveryCount );  // This might cancel this actor (if the recovery count is ancient) and destroy self
		Void _ = wait(Future<Void>(Void()));           // ... so check for cancellation

		// Restore popped keys.  Pop operations that took place after the last (committed) updatePersistentDataVersion might be lost, but
		// that is fine because we will get the corresponding data back, too.
		state KeyRange tagKeys = persistTagPoppedKeys;
		loop {
			Standalone<VectorRef<KeyValueRef>> data = wait( self->persistentData->readRange( tagKeys, BUGGIFY ? 3 : 1<<30, 1<<20 ) );
			if (!data.size()) break;
			((KeyRangeRef&)tagKeys) = KeyRangeRef( keyAfter(data.back().key, tagKeys.arena()), tagKeys.end );

			for(auto &kv : data) {
				Tag tag = decodeTagPoppedKey(kv.key);
				Version popped = decodeTagPoppedValue(kv.value);
				TraceEvent("TLogRestorePop", self->dbgid).detail("Tag", tag).detail("To", popped);
				ASSERT( self->tag_data.find(tag) == self->tag_data.end() );
				self->tag_data.insert( mapPair( std::move(Tag(tag)), TLogData::TagData( popped, false, false, tag )) );
			}
		}

		// PERSIST: Apply changes from queue
		if (processQueue) {
			state Version lastVer = 0;
			state double recoverMemoryLimit = SERVER_KNOBS->TARGET_BYTES_PER_TLOG + SERVER_KNOBS->SPRING_BYTES_TLOG;
			if (BUGGIFY) recoverMemoryLimit = SERVER_KNOBS->BUGGIFY_RECOVER_MEMORY_LIMIT;
			try {
				loop {
					TLogQueueEntry qe = wait( self->persistentQueue->readNext() );
					//TraceEvent("TLogRecoveredQE", self->dbgid).detail("ver", qe.version).detail("MessageBytes", qe.messages.size()).detail("Tags", qe.tags.size())
					//	.detail("Tag0", qe.tags.size() ? qe.tags[0].tag : invalidTag);

					ASSERT( qe.version > lastVer );
					lastVer = qe.version;
					self->knownCommittedVersion = std::max(self->knownCommittedVersion, qe.knownCommittedVersion);
					if( qe.version > self->version.get() ) {
						commitMessages(self, qe.version, qe.arena(), qe.messages, qe.tags);
						self->version.set( qe.version );
						self->queueCommittedVersion.set( qe.version );

						if (self->bytesInput.getValue() - self->bytesDurable.getValue() > recoverMemoryLimit) {
							TEST(true);  // Flush excess data during TLog queue recovery
							TraceEvent("FlushLargeQueueDuringRecovery", self->dbgid).detail("BytesInput", self->bytesInput.getValue()).detail("BytesDurable", self->bytesDurable.getValue()).detail("Version", self->version.get()).detail("PVer", self->persistentDataVersion);

							while(self->persistentDataDurableVersion != self->version.get()) {
								Version nextVersion;
								int totalSize = 0;
								std::vector<std::pair<std::deque<std::pair<Version, LengthPrefixedStringRef>>::iterator, std::deque<std::pair<Version, LengthPrefixedStringRef>>::iterator>> iters;
								for(auto tag = self->tag_data.begin(); tag != self->tag_data.end(); ++tag)
									iters.push_back(std::make_pair(tag->value.version_messages.begin(), tag->value.version_messages.end()));

								while( totalSize < SERVER_KNOBS->UPDATE_STORAGE_BYTE_LIMIT ) {
									nextVersion = self->version.get();
									for( auto &it : iters )
										if(it.first != it.second)
											nextVersion = std::min( nextVersion, it.first->first + 1 );

									if(nextVersion == self->version.get())
										break;

									for( auto &it : iters ) {
										while (it.first != it.second && it.first->first < nextVersion) {
											totalSize += it.first->second.expectedSize();
											++it.first;
										}
									}
								}

								Void _ = wait( updatePersistentData(self, nextVersion ) );
							}
						}
					}
				}
			} catch (Error& e) {
				if (e.code() != error_code_end_of_stream) throw;
			}
		}

		TraceEvent("TLogRestorePersistentStateDone", self->dbgid)
			.detail("pq", processQueue).detail("version", self->version.get()).detail("durableVer", self->persistentDataDurableVersion)
			.detail("Took", now()-startt);
		TEST( now()-startt >= 1.0 );  // TLog recovery took more than 1 second
		TEST( processQueue );  // TLog recovered from disk queue

		return Void();
	}

	void getQueuingMetrics( TLogData* self, TLogQueuingMetricsRequest const& req ) {
		TLogQueuingMetricsReply reply;
		reply.localTime = now();
		reply.instanceID = self->instanceID;
		reply.bytesInput = self->bytesInput.getValue();
		reply.bytesDurable = self->bytesDurable.getValue();
		reply.storageBytes = self->persistentData->getStorageBytes();
		reply.v = self->prevVersion;
		req.reply.send( reply );
	}

	ACTOR Future<Void> respondToRecovered( TLogInterface tli, Future<Void> recovery ) {
		Void _ = wait( recovery );

		loop {
			TLogRecoveryFinishedRequest req = waitNext( tli.recoveryFinished.getFuture() );
			req.reply.send(Void());
		}
	}

	ACTOR Future<Void> cleanupPeekTrackers( TLogData* self ) {
		loop {
			double minExpireTime = SERVER_KNOBS->PEEK_TRACKER_EXPIRATION_TIME;
			auto it = self->peekTracker.begin();
			while(it != self->peekTracker.end()) {
				double expireTime = SERVER_KNOBS->PEEK_TRACKER_EXPIRATION_TIME - now()-it->second.lastUpdate;
				if(expireTime < 1.0e-6) {
					for(auto seq : it->second.sequence_version) {
						if(!seq.second.isSet()) {
							seq.second.sendError(timed_out());
						}
					}
					it = self->peekTracker.erase(it);
				} else {
					minExpireTime = std::min(minExpireTime, expireTime);
					++it;
				}
			}

			Void _ = wait( delay(minExpireTime) );
		}
	}

	ACTOR Future<Void> serveTLogInterface( TLogData* self, TLogInterface tli, PromiseStream<Void> warningCollectorInput ) {
		loop choose {
			when( TLogPeekRequest req = waitNext( tli.peekMessages.getFuture() ) ) {
				self->addActor.send( tLogPeekMessages( self, req ) );
			}
			when( TLogPopRequest req = waitNext( tli.popMessages.getFuture() ) ) {
				self->addActor.send( tLogPop( self, req ) );
			}
			when( TLogCommitRequest req = waitNext( tli.commit.getFuture() ) ) {
				TEST(self->stopped); // TLogCommitRequest while stopped
				if (!self->stopped)
					self->addActor.send( tLogCommit( self, req, warningCollectorInput ) );
				else
					req.reply.sendError( tlog_stopped() );
			}
			when( ReplyPromise< TLogLockResult > reply = waitNext( tli.lock.getFuture() ) ) {
				self->addActor.send( tLogLock(self, reply) );
			}
			when (TLogQueuingMetricsRequest req = waitNext(tli.getQueuingMetrics.getFuture())) {
				getQueuingMetrics(self, req);
			}
			when (TLogConfirmRunningRequest req = waitNext(tli.confirmRunning.getFuture())){
				if (req.debugID.present() ) {
					UID tlogDebugID = g_nondeterministic_random->randomUniqueID();
					g_traceBatch.addAttach("TransactionAttachID", req.debugID.get().first(), tlogDebugID.first());
					g_traceBatch.addEvent("TransactionDebug", tlogDebugID.first(), "TLogServer.TLogConfirmRunningRequest");
				}
				if (!self->stopped)
					req.reply.send(Void());
				else
					req.reply.sendError( tlog_stopped() );
			}
		}
	}

	ACTOR Future<Void> tLogCore( TLogData* self, TLogInterface tli, Future<Void> recovery ) {
		state PromiseStream<Void> warningCollectorInput;
		state Future<Void> warningCollector = timeoutWarningCollector( warningCollectorInput.getFuture(), 1.0, "TLogQueueCommitSlow", self->dbgid );
		state Future<Void> error = actorCollection( self->addActor.getFuture() );

		self->addActor.send( updateStorage(self) );
		self->addActor.send( commitQueue(self) );
		self->addActor.send( waitFailureServer(tli.waitFailure.getFuture()) );
		self->addActor.send( respondToRecovered(tli, recovery) );
		self->addActor.send( traceCounters("TLogMetrics", self->dbgid, SERVER_KNOBS->STORAGE_LOGGING_DELAY, &self->cc, self->dbgid.toString() + "/TLogMetrics"));
		self->addActor.send( cleanupPeekTrackers(self) );

		if( recovery.isValid() && !recovery.isReady()) {
			self->addActor.send( recovery );
		}

		self->coreStarted = true;

		Void _ = wait( serveTLogInterface(self, tli, warningCollectorInput) || error );
		throw internal_error();
	};

	ACTOR Future<Void> checkEmptyQueue(TLogData* self) {
		TraceEvent("TLogCheckEmptyQueueBegin", self->dbgid);
		try {
			TLogQueueEntry r = wait( self->persistentQueue->readNext() );
			throw internal_error();
		} catch (Error& e) {
			if (e.code() != error_code_end_of_stream) throw;
			TraceEvent("TLogCheckEmptyQueueEnd", self->dbgid);
			return Void();
		}
	}

	ACTOR Future<Void> recoverTagFromLogSystem( TLogData* self, Version beginVersion, Version endVersion, Tag tag, Reference<AsyncVar<int>> uncommittedBytes, Reference<AsyncVar<Reference<ILogSystem>>> logSystem ) {
		state Future<Void> dbInfoChange = Void();
		state Reference<ILogSystem::IPeekCursor> r;
		state Version tagAt = beginVersion;
		state Version tagPopped = 0;
		state Version lastVer = 0;

		TraceEvent("LogRecoveringTagBegin", self->dbgid).detail("Tag", tag).detail("recoverAt", endVersion);

		while (tagAt <= endVersion) {
			loop {
				choose {
					when(Void _ = wait( r ? r->getMore() : Never() ) ) {
						break;
					}
					when( Void _ = wait( dbInfoChange ) ) {
						if(r) tagPopped = std::max(tagPopped, r->popped());
						if( logSystem->get() )
							r = logSystem->get()->peek( tagAt, tag );
						else
							r = Reference<ILogSystem::IPeekCursor>();
						dbInfoChange = logSystem->onChange();
					}
				}
			}

			TraceEvent("LogRecoveringTagResults", self->dbgid).detail("Tag", tag);

			Version ver = 0;
			BinaryWriter wr( Unversioned() );
			int writtenBytes = 0;
			while (true) {
				bool foundMessage = r->hasMessage();
				//TraceEvent("LogRecoveringMsg").detail("Tag", tag).detail("foundMessage", foundMessage).detail("ver", r->version().toString());
				if (!foundMessage || r->version().version != ver) {
					ASSERT(r->version().version > lastVer);
					if (ver) {
						//TraceEvent("LogRecoveringTagVersion", self->dbgid).detail("Tag", tag).detail("Ver", ver).detail("Bytes", wr.getLength());
						writtenBytes += 100 + wr.getLength();
						self->persistentData->set( KeyValueRef( persistTagMessagesKey( tag, ver ), wr.toStringRef() ) );
					}
					lastVer = ver;
					ver = r->version().version;
					wr = BinaryWriter( Unversioned() );
					if (!foundMessage || ver > endVersion)
						break;
				}

				// FIXME: This logic duplicates stuff in LogPushData::addMessage(), and really would be better in PeekResults or somewhere else.  Also unnecessary copying.
				StringRef msg = r->getMessage();
				wr << uint32_t( msg.size() + sizeof(uint32_t) ) << r->version().sub;
				wr.serializeBytes( msg );
				r->nextMessage();
			}

			tagAt = r->version().version;

			if(writtenBytes)
				uncommittedBytes->set(uncommittedBytes->get() + writtenBytes);

			while(uncommittedBytes->get() >= SERVER_KNOBS->LARGE_TLOG_COMMIT_BYTES) {
				Void _ = wait(uncommittedBytes->onChange());
			}
		}
		if(r) tagPopped = std::max(tagPopped, r->popped());

		auto tsm = self->tag_data.find(tag);
		if (tsm == self->tag_data.end()) {
			self->tag_data.insert( mapPair(std::move(Tag(tag)), TLogData::TagData(tagPopped, false, true, tag)) );
		}

		Void _ = wait(tLogPop( self, TLogPopRequest(tagPopped, tag) ));

		updatePersistentPopped( self, tag, self->tag_data.find(tag)->value );
		return Void();
	}

	ACTOR Future<Void> updateLogSystem(TLogData* self, LogSystemConfig recoverFrom, Reference<AsyncVar<Reference<ILogSystem>>> logSystem) {
		loop {
			TraceEvent("TLogUpdate", self->dbgid).detail("recoverFrom", recoverFrom.toString()).detail("dbInfo", self->dbInfo->get().logSystemConfig.toString());
			if( self->dbInfo->get().logSystemConfig.isEqualIds(recoverFrom) ) {
				logSystem->set(ILogSystem::fromLogSystemConfig( self->dbgid, self->dbInfo->get().myLocality, self->dbInfo->get().logSystemConfig ));
			} else if( self->dbInfo->get().logSystemConfig.isNextGenerationOf(recoverFrom) && std::count( self->dbInfo->get().logSystemConfig.tLogs.begin(), self->dbInfo->get().logSystemConfig.tLogs.end(), self->dbgid ) ) {
				logSystem->set(ILogSystem::fromOldLogSystemConfig( self->dbgid, self->dbInfo->get().myLocality, self->dbInfo->get().logSystemConfig ));
			} else {
				logSystem->set(Reference<ILogSystem>());
			}
			Void _ = wait( self->dbInfo->onChange() );
		}
	}

	ACTOR Future<Void> recoverFromLogSystem( TLogData* self, LogSystemConfig recoverFrom, Version recoverAt, Version knownCommittedVersion, std::vector<Tag> recoverTags, Promise<Void> copyComplete ) {
		state Future<Void> committing = Void();
		state double lastCommitT = now();
		state Reference<AsyncVar<int>> uncommittedBytes = Reference<AsyncVar<int>>(new AsyncVar<int>());
		state std::vector<Future<Void>> recoverFutures;
		state Reference<AsyncVar<Reference<ILogSystem>>> logSystem = Reference<AsyncVar<Reference<ILogSystem>>>(new AsyncVar<Reference<ILogSystem>>());
		state Future<Void> updater = updateLogSystem(self, recoverFrom, logSystem);

		for(auto tag : recoverTags )
			recoverFutures.push_back(recoverTagFromLogSystem(self, knownCommittedVersion, recoverAt, tag, uncommittedBytes, logSystem));

		state Future<Void> copyDone = waitForAll(recoverFutures);
		state Future<Void> recoveryDone = Never();
		state Future<Void> commitTimeout = delay(SERVER_KNOBS->LONG_TLOG_COMMIT_TIME);

		loop {
			choose {
				when(Void _ = wait(copyDone)) {
					recoverFutures.clear();
					for(auto tag : recoverTags )
						recoverFutures.push_back(recoverTagFromLogSystem(self, 0, knownCommittedVersion, tag, uncommittedBytes, logSystem));
					copyDone = Never();
					recoveryDone =  waitForAll(recoverFutures);

					Void __ = wait( committing );
					Void __ = wait( self->updatePersist );
					committing = self->persistentData->commit();
					commitTimeout = delay(SERVER_KNOBS->LONG_TLOG_COMMIT_TIME);
					uncommittedBytes->set(0);
					Void __ = wait( committing );
					TraceEvent("TLogCommitCopyData", self->dbgid);

					if(!copyComplete.isSet())
						copyComplete.send(Void());
				}
				when(Void _ = wait(recoveryDone)) { break; }
				when(Void _ = wait(commitTimeout)) {
					TEST(true); // We need to commit occasionally if this process is long to avoid running out of memory.
					// We let one, but not more, commits pipeline with the network transfer
					Void __ = wait( committing );
					Void __ = wait( self->updatePersist );
					committing = self->persistentData->commit();
					commitTimeout = delay(SERVER_KNOBS->LONG_TLOG_COMMIT_TIME);
					uncommittedBytes->set(0);
					TraceEvent("TLogCommitRecoveryData", self->dbgid).detail("MemoryUsage", DEBUG_DETERMINISM ? 0 : getMemoryUsage());
				}
				when(Void _ = wait(uncommittedBytes->onChange())) {
					if(uncommittedBytes->get() >= SERVER_KNOBS->LARGE_TLOG_COMMIT_BYTES)
						commitTimeout = Void();
				}
			}
		}

		Void _ = wait( committing );
		Void _ = wait( self->updatePersist );
		Void _ = wait( self->persistentData->commit() );

		TraceEvent("TLogRecoveryComplete", self->dbgid).detail("Locality", self->dbInfo->get().myLocality.toString());
		TEST(true);  // tLog restore from old log system completed

		return Void();
	}

	ACTOR Future<Void> tLogStart( TLogData* self, LogSystemConfig recoverFrom, Version recoverAt, Version knownCommittedVersion, std::vector<Tag> recoverTags, bool recoverFromDisk,
		TLogInterface tli, ReplyPromise<TLogInterface> outInterface, Promise<DBRecoveryCount> outRecoveryCount ) {
		state Future<Void> recovery = Void();
		if (recoverFrom.logSystemType == 1) {
			ASSERT(false);
		} else if (recoverFrom.logSystemType == 2) {
			Void _ = wait( checkEmptyQueue(self) );

			self->persistentDataVersion = recoverAt;
			self->persistentDataDurableVersion = recoverAt; // durable is a white lie until initPersistentState() commits the store
			self->queueCommittedVersion.set( recoverAt );
			self->version.set( recoverAt );

			Void _ = wait( initPersistentState( self ) );

			state Promise<Void> copyComplete;
			recovery = recoverFromLogSystem( self, recoverFrom, recoverAt, knownCommittedVersion, recoverTags, copyComplete );
			Void _ = wait(copyComplete.getFuture());
		} else if (recoverFromDisk) {
			Void _ = wait( restorePersistentState( self, outRecoveryCount, true, tli ) );
			TEST(true);  // tLog restore from disk completed
		} else {
			// Brand new tlog, initialization has already been done by caller
			Void _ = wait( checkEmptyQueue(self) );
			Void _ = wait( initPersistentState( self ) );
		}

		TraceEvent("TLogReady", self->dbgid);

		validate(self);
		//dump(self);

		outInterface.send( tli );

		Void _ = wait( tLogCore( self, tli, recovery ) );
		throw internal_error();  // tLogCore() shouldn't return without an error
	}

	ACTOR Future<Void> rejoinMasters( TLogData* self, TLogInterface tli, Future<DBRecoveryCount> fRecoveryCount ) {
		state DBRecoveryCount recoveryCount = wait( fRecoveryCount );
		state UID lastMasterID(0,0);
		loop {
			auto const& inf = self->dbInfo->get();
			bool isDisplaced = inf.recoveryCount >= recoveryCount && inf.recoveryState != 0 &&
				 !std::count( inf.logSystemConfig.tLogs.begin(), inf.logSystemConfig.tLogs.end(), tli.id() ) &&
				 !std::count( inf.priorCommittedLogServers.begin(), inf.priorCommittedLogServers.end(), tli.id() );
			for(int i = 0; i < inf.logSystemConfig.oldTLogs.size() && isDisplaced; i++) {
				isDisplaced = !std::count( inf.logSystemConfig.oldTLogs[i].tLogs.begin(), inf.logSystemConfig.oldTLogs[i].tLogs.end(), tli.id() );
			}
			if ( isDisplaced )
			{
				TraceEvent("TLogDisplaced", tli.id()).detail("Reason", "DBInfoDoesNotContain");
				if (BUGGIFY) Void _ = wait( delay( SERVER_KNOBS->BUGGIFY_WORKER_REMOVED_MAX_LAG * g_random->random01() ) );
				throw worker_removed();
			}

			if (self->dbInfo->get().master.id() != lastMasterID) {
				// The TLogRejoinRequest is needed to establish communications with a new master, which doesn't have our TLogInterface
				TLogRejoinRequest req;
				req.myInterface = tli;
				TraceEvent("TLogRejoining", self->dbgid).detail("Master", self->dbInfo->get().master.id());
				choose {
					when ( bool success = wait( brokenPromiseToNever( self->dbInfo->get().master.tlogRejoin.getReply( req ) ) ) ) {
						if (success)
							lastMasterID = self->dbInfo->get().master.id();
					}
					when ( Void _ = wait( self->dbInfo->onChange() ) ) { }
				}
			} else
				Void _ = wait( self->dbInfo->onChange() );
		}
	}

	// Restore from disk
	ACTOR Future<Void> tLog( IKeyValueStore* persistentData, IDiskQueue* persistentQueue, TLogInterface tli, Reference<AsyncVar<ServerDBInfo>> db ) {
		state TLogData self( tli.id(), persistentData, persistentQueue, db );
		state Promise<DBRecoveryCount> recoveryCount;
		state Future<Void> removed = rejoinMasters(&self, tli, recoveryCount.getFuture());

		Void _ = wait( tLogStart( &self, LogSystemConfig(), Version(0), Version(0), std::vector<Tag>(), true, tli, ReplyPromise<TLogInterface>(), recoveryCount ) || removed );
		throw internal_error();  // tLogStart doesn't return without an error
	}
}