Files
ternfs-XTXMarkets/cpp/shard/Shard.cpp
2025-12-02 21:24:24 +00:00

2477 lines
109 KiB
C++

// Copyright 2025 XTX Markets Technologies Limited
//
// SPDX-License-Identifier: GPL-2.0-or-later
#include "Shard.hpp"
#include <atomic>
#include <cstdint>
#include <memory>
#include <ostream>
#include <string>
#include <unordered_map>
#include <vector>
#include <fcntl.h>
#include "Assert.hpp"
#include "Bincode.hpp"
#include "BlockServicesCacheDB.hpp"
#include "CDCKey.hpp"
#include "Common.hpp"
#include "Crypto.hpp"
#include "Env.hpp"
#include "ErrorCount.hpp"
#include "Exception.hpp"
#include "LogsDB.hpp"
#include "Loop.hpp"
#include "Metrics.hpp"
#include "Msgs.hpp"
#include "MsgsGen.hpp"
#include "MultiplexedChannel.hpp"
#include "PeriodicLoop.hpp"
#include "Protocol.hpp"
#include "Random.hpp"
#include "ShardDB.hpp"
#include "ShardKey.hpp"
#include "SharedRocksDB.hpp"
#include "RegistryClient.hpp"
#include "SPSC.hpp"
#include "Time.hpp"
#include "Timings.hpp"
#include "UDPSocketPair.hpp"
#include "Xmon.hpp"
struct ShardReq {
uint32_t protocol;
ShardReqMsg msg;
TernTime receivedAt;
IpPort clientAddr;
int sockIx; // which sock to use to reply
};
struct ShardResp {
ShardRespMsg msg;
IpPort clientAddr;
};
struct ProxyLogsDBRequest {
IpPort clientAddr;
int sockIx; // which sock to use to reply
LogsDBRequest request;
};
struct ProxyLogsDBResponse {
LogsDBResponse response;
};
// TODO make options
const int WRITER_QUEUE_SIZE = 8192;
const int READER_QUEUE_SIZE = 1024;
const int MAX_RECV_MSGS = 100;
std::ostream& operator<<(std::ostream& out, const std::vector<FullShardInfo>& shards) {
out << "[";
for (const auto& s : shards) {
out << s << ", ";
}
return out << "]";
}
struct ShardShared {
const ShardOptions& options;
// network
std::array<UDPSocketPair, 1> socks; // in an array to play with UDPReceiver<>
// server to -> reader/writer communication
MultiSPSCWaiter writeQueuesWaiter;
SPSC<LogsDBRequest, true> logsDBRequestQueue;
SPSC<LogsDBResponse, true> logsDBResponseQueue;
SPSC<ProxyLogsDBRequest, true> proxyLogsDBRequestQueue;
SPSC<ProxyLogsDBResponse, true> proxyLogsDBResponseQueue;
SPSC<ShardReq, true> writerShardReqQueue;
SPSC<ShardResp, true> writerShardRespQueue;
SPSC<ProxyShardRespMsg, true> writerProxyShardRespQueue;
std::vector<std::unique_ptr<SPSC<ShardReq>>> readerRequestsQueues;
// databases and caches
SharedRocksDB& sharedDB;
LogsDB& logsDB;
ShardDB& shardDB;
BlockServicesCacheDB& blockServicesCache;
// statistics
std::array<Timings, maxShardMessageKind+1> timings;
std::array<ErrorCount, maxShardMessageKind+1> errors;
std::atomic<uint64_t> receivedLogsDBReqs;
std::atomic<uint64_t> receivedLogsDBResps;
std::atomic<uint64_t> receivedProxyLogsDBReqs;
std::atomic<uint64_t> receivedProxyLogsDBResps;
std::atomic<uint64_t> receivedWriterReqs;
std::atomic<uint64_t> receivedWriterProxyResps;
std::atomic<uint64_t> receivedReaderReqs;
std::array<std::atomic<uint64_t>, 2> receivedRequests; // how many requests we got at once from each socket
// how many requests we dropped
std::atomic<uint64_t> droppedLogsDBReqs;
std::atomic<uint64_t> droppedLogsDBResps;
std::atomic<uint64_t> droppedProxyLogsDBReqs;
std::atomic<uint64_t> droppedProxyLogsDBResps;
std::atomic<uint64_t> droppedWriteReqs;
std::atomic<uint64_t> droppedProxyWriteResps;
std::atomic<uint64_t> droppedReadReqs;
// we should get up to date information from registry before we start serving any requests
// this is populated by ShardRegisterer
std::atomic<bool> isInitiated;
std::atomic<bool> isBlockServiceCacheInitiated;
ShardShared() = delete;
ShardShared(const ShardOptions& options_, SharedRocksDB& sharedDB_, BlockServicesCacheDB& blockServicesCache_, ShardDB& shardDB_, LogsDB& logsDB_, UDPSocketPair&& sock) :
options(options_),
socks({std::move(sock)}),
logsDBRequestQueue(WRITER_QUEUE_SIZE, writeQueuesWaiter),
logsDBResponseQueue(WRITER_QUEUE_SIZE, writeQueuesWaiter),
proxyLogsDBRequestQueue(WRITER_QUEUE_SIZE, writeQueuesWaiter),
proxyLogsDBResponseQueue(WRITER_QUEUE_SIZE, writeQueuesWaiter),
writerShardReqQueue(WRITER_QUEUE_SIZE, writeQueuesWaiter),
writerShardRespQueue(WRITER_QUEUE_SIZE, writeQueuesWaiter),
writerProxyShardRespQueue(WRITER_QUEUE_SIZE, writeQueuesWaiter),
sharedDB(sharedDB_),
logsDB(logsDB_),
shardDB(shardDB_),
blockServicesCache(blockServicesCache_),
receivedLogsDBReqs(0),
receivedLogsDBResps(0),
receivedProxyLogsDBReqs(0),
receivedProxyLogsDBResps(0),
receivedWriterReqs(0),
receivedWriterProxyResps(0),
receivedReaderReqs(0),
droppedLogsDBReqs(0),
droppedLogsDBResps(0),
droppedProxyLogsDBReqs(0),
droppedProxyLogsDBResps(0),
droppedWriteReqs(0),
droppedProxyWriteResps(0),
droppedReadReqs(0),
isInitiated(false),
isBlockServiceCacheInitiated(false),
_replicas(nullptr),
_leadersAtOtherLocations(std::make_shared<const std::vector<FullShardInfo>>())
{
for (uint16_t i = 0; i < options_.numReaders; ++i) {
readerRequestsQueues.emplace_back(std::make_unique<SPSC<ShardReq>>(READER_QUEUE_SIZE));
}
for (ShardMessageKind kind : allShardMessageKind) {
timings[(int)kind] = Timings::Standard();
}
for (auto& x: receivedRequests) {
x = 0;
}
}
const UDPSocketPair& sock() const {
return socks[0];
}
inline std::shared_ptr<const std::array<AddrsInfo, LogsDB::REPLICA_COUNT>> getReplicas() const {
return std::atomic_load(&_replicas);
}
inline void setReplicas(const std::array<AddrsInfo, LogsDB::REPLICA_COUNT>&& replicas) {
std::atomic_store(&_replicas, std::make_shared<const std::array<AddrsInfo, LogsDB::REPLICA_COUNT>>(replicas));
}
inline std::shared_ptr<const std::vector<FullShardInfo>> getLeadersAtOtherLocations() const {
return std::atomic_load(&_leadersAtOtherLocations);
}
inline void setLeadersAtOtherLocations(const std::vector<FullShardInfo>&& leadersAtOtherLocations) {
std::atomic_store(&_leadersAtOtherLocations, std::make_shared<const std::vector<FullShardInfo>>(leadersAtOtherLocations));
}
private:
// replication information
std::shared_ptr<const std::array<AddrsInfo, LogsDB::REPLICA_COUNT>> _replicas; // used for replicating shard within location
std::shared_ptr<const std::vector<FullShardInfo>> _leadersAtOtherLocations; // used for cross location replication
};
static bool bigRequest(ShardMessageKind kind) {
return unlikely(
kind == ShardMessageKind::ADD_SPAN_INITIATE ||
kind == ShardMessageKind::ADD_SPAN_CERTIFY
);
}
static bool bigResponse(ShardMessageKind kind) {
return unlikely(
kind == ShardMessageKind::READ_DIR ||
kind == ShardMessageKind::ADD_SPAN_INITIATE ||
kind == ShardMessageKind::LOCAL_FILE_SPANS ||
kind == ShardMessageKind::FILE_SPANS ||
kind == ShardMessageKind::VISIT_DIRECTORIES ||
kind == ShardMessageKind::VISIT_FILES ||
kind == ShardMessageKind::VISIT_TRANSIENT_FILES ||
kind == ShardMessageKind::BLOCK_SERVICE_FILES ||
kind == ShardMessageKind::FULL_READ_DIR
);
}
static void packShardResponse(
Env& env,
ShardShared& shared,
const AddrsInfo& srcAddr,
UDPSender& sender,
bool dropArtificially,
const ShardReq& req,
const ShardRespMsg& msg
) {
auto respKind = msg.body.kind();
auto reqKind = req.msg.body.kind();
auto elapsed = ternNow() - req.receivedAt;
shared.timings[(int)reqKind].add(elapsed);
shared.errors[(int)reqKind].add( respKind != ShardMessageKind::ERROR ? TernError::NO_ERROR : msg.body.getError());
if (unlikely(dropArtificially)) {
LOG_DEBUG(env, "artificially dropping response %s", msg.id);
return;
}
ALWAYS_ASSERT(req.clientAddr.port != 0);
if (respKind != ShardMessageKind::ERROR) {
LOG_DEBUG(env, "successfully processed request %s with kind %s in %s", msg.id, reqKind, elapsed);
if (bigResponse(respKind)) {
if (unlikely(env._shouldLog(LogLevel::LOG_TRACE))) {
LOG_TRACE(env, "resp: %s", msg);
} else {
LOG_DEBUG(env, "resp: <omitted>");
}
} else {
LOG_DEBUG(env, "resp: %s", msg);
}
} else {
LOG_DEBUG(env, "request %s failed with error %s in %s", reqKind, msg.body.getError(), elapsed);
}
sender.prepareOutgoingMessage(env, srcAddr, req.sockIx, req.clientAddr,
[&msg](BincodeBuf& buf) {
msg.pack(buf);
});
LOG_DEBUG(env, "will send response for req id %s kind %s to %s", msg.id, reqKind, req.clientAddr);
}
template<typename T>
static void packCheckPointedShardResponse(
Env& env,
ShardShared& shared,
const AddrsInfo& srcAddr,
UDPSender& sender,
bool dropArtificially,
const ShardReq& req,
const T& msg,
const AES128Key& key
) {
auto respKind = msg.body.resp.kind();
auto reqKind = req.msg.body.kind();
auto elapsed = ternNow() - req.receivedAt;
shared.timings[(int)reqKind].add(elapsed);
shared.errors[(int)reqKind].add( respKind != ShardMessageKind::ERROR ? TernError::NO_ERROR : msg.body.resp.getError());
if (unlikely(dropArtificially)) {
LOG_DEBUG(env, "artificially dropping response %s", msg.id);
return;
}
ALWAYS_ASSERT(req.clientAddr.port != 0);
if (respKind != ShardMessageKind::ERROR) {
LOG_DEBUG(env, "successfully processed request %s with kind %s in %s", msg.id, reqKind, elapsed);
if (bigResponse(respKind)) {
if (unlikely(env._shouldLog(LogLevel::LOG_TRACE))) {
LOG_TRACE(env, "resp: %s", msg);
} else {
LOG_DEBUG(env, "resp: <omitted>");
}
} else {
LOG_DEBUG(env, "resp: %s", msg);
}
} else {
LOG_DEBUG(env, "request %s failed with error %s in %s", reqKind, msg.body.resp.getError(), elapsed);
}
sender.prepareOutgoingMessage(env, srcAddr, req.sockIx, req.clientAddr,
[&msg, &key](BincodeBuf& buf) {
msg.pack(buf, key);
});
LOG_DEBUG(env, "will send response for req id %s kind %s to %s", msg.id, reqKind, req.clientAddr);
}
static constexpr std::array<uint32_t, 7> ShardProtocols = {
SHARD_REQ_PROTOCOL_VERSION,
SHARD_RESP_PROTOCOL_VERSION,
CDC_TO_SHARD_REQ_PROTOCOL_VERSION,
LOG_REQ_PROTOCOL_VERSION,
LOG_RESP_PROTOCOL_VERSION,
PROXY_SHARD_RESP_PROTOCOL_VERSION,
PROXY_SHARD_REQ_PROTOCOL_VERSION};
using ShardChannel = MultiplexedChannel<ShardProtocols.size(), ShardProtocols>;
struct ShardServer : Loop {
private:
// init data
ShardShared& _shared;
// run data
AES128Key _expandedCDCKey;
AES128Key _expandedShardKey;
// log entries buffers
std::vector<LogsDBRequest> _logsDBRequests;
std::vector<LogsDBResponse> _logsDBResponses;
std::vector<ProxyLogsDBRequest> _proxyLogsDBRequests;
std::vector<ProxyLogsDBResponse> _proxyLogsDBResponses;
std::vector<ShardReq> _writeReqs;
std::vector<ProxyShardRespMsg> _proxyResponses;
std::vector<ShardResp> _waitResponses;
// read requests buffer
std::vector<ShardReq> _readRequests;
std::vector<std::unique_ptr<SPSC<ShardReq>>>::iterator _readQueueIt;
std::unique_ptr<UDPReceiver<1>> _receiver;
std::unique_ptr<ShardChannel> _channel;
public:
ShardServer(Logger& logger, std::shared_ptr<XmonAgent>& xmon, ShardShared& shared) :
Loop(logger, xmon, "server"),
_shared(shared),
_readQueueIt(_shared.readerRequestsQueues.begin())
{
auto convertProb = [this](const std::string& what, double prob, uint64_t& iprob) {
if (prob != 0.0) {
LOG_INFO(_env, "Will drop %s%% of %s packets", prob*100.0, what);
iprob = prob * 10'000.0;
ALWAYS_ASSERT(iprob > 0 && iprob < 10'000);
}
};
expandKey(ShardKey, _expandedShardKey);
expandKey(CDCKey, _expandedCDCKey);
_receiver = std::make_unique<UDPReceiver<1>>(UDPReceiverConfig{.perSockMaxRecvMsg = MAX_RECV_MSGS, .maxMsgSize = MAX_UDP_MTU});
_channel = std::make_unique<ShardChannel>();
}
virtual ~ShardServer() = default;
private:
void _handleLogsDBResponse(UDPMessage& msg) {
LOG_DEBUG(_env, "received LogsDBResponse from %s", msg.clientAddr);
LogRespMsg* respMsg = nullptr;
auto replicaId = _getReplicaId(msg.clientAddr);
bool proxyLocation = true;
if (replicaId != LogsDB::REPLICA_COUNT) {
auto& resp = _logsDBResponses.emplace_back();
resp.replicaId = replicaId;
respMsg = &resp.msg;
proxyLocation = false;
} else {
// try matching to other locations
auto leadersPtr = _shared.getLeadersAtOtherLocations();
for (auto& leader : *leadersPtr) {
if (leader.locationId == 0 && leader.addrs.contains(msg.clientAddr)) {
auto& resp = _proxyLogsDBResponses.emplace_back();
respMsg = &resp.response.msg;
break;
}
}
}
if (respMsg == nullptr) {
LOG_DEBUG(_env, "We can't match this address to replica or other leader. Dropping");
return;
}
try {
respMsg->unpack(msg.buf, _expandedShardKey);
} catch (const BincodeException& err) {
LOG_ERROR(_env, "Could not parse LogsDBResponse: %s", err.what());
if (proxyLocation) {
_proxyLogsDBResponses.pop_back();
} else {
_logsDBResponses.pop_back();
}
return;
}
LOG_DEBUG(_env, "Received response %s for requests id %s from replica id %s", respMsg->body.kind(), respMsg->id, replicaId);
}
void _handleLogsDBRequest(UDPMessage& msg) {
LOG_DEBUG(_env, "received LogsDBRequest from %s", msg.clientAddr);
LogReqMsg* reqMsg = nullptr;
bool proxyLocation = true;
auto replicaId = _getReplicaId(msg.clientAddr);
if (replicaId != LogsDB::REPLICA_COUNT) {
auto& req = _logsDBRequests.emplace_back();
req.replicaId = replicaId;
reqMsg = &req.msg;
proxyLocation = false;
} else {
// try matching to other locations
auto leadersPtr = _shared.getLeadersAtOtherLocations();
for (auto& leader : *leadersPtr) {
if (leader.addrs.contains(msg.clientAddr)) {
auto& req = _proxyLogsDBRequests.emplace_back();
req.clientAddr = msg.clientAddr;
req.sockIx = msg.socketIx;
reqMsg = &req.request.msg;
break;
}
}
}
if (reqMsg == nullptr) {
LOG_DEBUG(_env, "We can't match this address to replica or other leader. Dropping");
return;
}
try {
reqMsg->unpack(msg.buf, _expandedShardKey);
} catch (const BincodeException& err) {
LOG_ERROR(_env, "Could not parse LogsDBRequest: %s", err.what());
if (proxyLocation) {
_proxyLogsDBRequests.pop_back();
} else {
_logsDBRequests.pop_back();
}
return;
}
LOG_DEBUG(_env, "Received request %s with requests id %s from replica id %s", reqMsg->body.kind(), reqMsg->id, replicaId);
}
uint8_t _getReplicaId(const IpPort& clientAddress) {
auto replicasPtr = _shared.getReplicas();
if (!replicasPtr) {
return LogsDB::REPLICA_COUNT;
}
for (ReplicaId replicaId = 0; replicaId.u8 < replicasPtr->size(); ++replicaId.u8) {
if (replicasPtr->at(replicaId.u8).contains(clientAddress)) {
return replicaId.u8;
}
}
return LogsDB::REPLICA_COUNT;
}
void _handleShardRequest(UDPMessage& msg, uint32_t protocol) {
LOG_DEBUG(_env, "received message from %s", msg.clientAddr);
if (unlikely(!_shared.options.isLeader())) {
LOG_DEBUG(_env, "not leader, dropping request %s", msg.clientAddr);
return;
}
ShardReqMsg req;
try {
switch (protocol) {
case CDC_TO_SHARD_REQ_PROTOCOL_VERSION:
{
CdcToShardReqMsg signedReq;
signedReq.unpack(msg.buf, _expandedCDCKey);
req.id = signedReq.id;
req.body = std::move(signedReq.body);
}
break;
case SHARD_REQ_PROTOCOL_VERSION:
req.unpack(msg.buf);
if (isPrivilegedRequestKind((uint8_t)req.body.kind())) {
LOG_ERROR(_env, "Received unauthenticated request %s from %s", req.body.kind(), msg.clientAddr);
return;
}
break;
case PROXY_SHARD_REQ_PROTOCOL_VERSION:
{
ProxyShardReqMsg signedReq;
signedReq.unpack(msg.buf, _expandedShardKey);
req.id = signedReq.id;
req.body = std::move(signedReq.body);
}
break;
default:
ALWAYS_ASSERT(false, "Unknown protocol version");
}
} catch (const BincodeException& err) {
LOG_ERROR(_env, "Could not parse: %s", err.what());
RAISE_ALERT(_env, "could not parse request from %s, dropping it.", msg.clientAddr);
return;
}
auto t0 = ternNow();
LOG_DEBUG(_env, "received request id %s, kind %s, from %s", req.id, req.body.kind(), msg.clientAddr);
if (bigRequest(req.body.kind())) {
if (unlikely(_env._shouldLog(LogLevel::LOG_TRACE))) {
LOG_TRACE(_env, "parsed request: %s", req);
} else {
LOG_DEBUG(_env, "parsed request: <omitted>");
}
} else {
LOG_DEBUG(_env, "parsed request: %s", req);
}
auto& entry = readOnlyShardReq(req.body.kind()) ? _readRequests.emplace_back() : _writeReqs.emplace_back();
entry.sockIx = msg.socketIx;
entry.clientAddr = msg.clientAddr;
entry.receivedAt = t0;
entry.protocol = protocol;
entry.msg = std::move(req);
}
void _handleShardResponse(UDPMessage& msg, uint32_t protocol) {
LOG_DEBUG(_env, "received message from %s", msg.clientAddr);
if (unlikely(!_shared.options.isLeader())) {
LOG_DEBUG(_env, "not leader, dropping response %s", msg.clientAddr);
return;
}
switch (protocol) {
case SHARD_RESP_PROTOCOL_VERSION: {
ShardResp& resp = _waitResponses.emplace_back();
try {
resp.msg.unpack(msg.buf);
ALWAYS_ASSERT(resp.msg.body.kind() == ShardMessageKind::WAIT_STATE_APPLIED);
resp.clientAddr = msg.clientAddr;
} catch (const BincodeException& err) {
LOG_ERROR(_env, "Could not parse: %s", err.what());
RAISE_ALERT(_env, "could not parse request from %s, dropping it.", msg.clientAddr);
_waitResponses.pop_back();
return;
}
LOG_DEBUG(_env, "parsed shard response from %s: %s", msg.clientAddr, resp.msg);
break;
}
case PROXY_SHARD_RESP_PROTOCOL_VERSION: {
ProxyShardRespMsg& resp = _proxyResponses.emplace_back();
try {
resp.unpack(msg.buf, _expandedShardKey);
} catch (const BincodeException& err) {
LOG_ERROR(_env, "Could not parse: %s", err.what());
RAISE_ALERT(_env, "could not parse request from %s, dropping it.", msg.clientAddr);
_proxyResponses.pop_back();
return;
}
LOG_DEBUG(_env, "parsed shard response from %s: %s", msg.clientAddr, resp);
break;
}
default:
ALWAYS_ASSERT(false, "Unknown protocol version");
}
}
public:
virtual void step() override {
if (unlikely(!_shared.isInitiated.load(std::memory_order_acquire) || !_shared.isBlockServiceCacheInitiated.load(std::memory_order_acquire))) {
(100_ms).sleepRetry();
return;
}
_logsDBRequests.clear();
_logsDBResponses.clear();
_proxyLogsDBRequests.clear();
_proxyLogsDBResponses.clear();
_writeReqs.clear();
_proxyResponses.clear();
_readRequests.clear();
_waitResponses.clear();
if (unlikely(!_channel->receiveMessages(_env, _shared.socks, *_receiver))) {
return;
}
for (auto& msg : _channel->protocolMessages(LOG_RESP_PROTOCOL_VERSION)) {
_handleLogsDBResponse(msg);
}
for (auto& msg : _channel->protocolMessages(LOG_REQ_PROTOCOL_VERSION)) {
_handleLogsDBRequest(msg);
}
for (auto& msg: _channel->protocolMessages(PROXY_SHARD_RESP_PROTOCOL_VERSION)) {
_handleShardResponse(msg, PROXY_SHARD_RESP_PROTOCOL_VERSION);
}
for (auto& msg: _channel->protocolMessages(SHARD_RESP_PROTOCOL_VERSION)) {
_handleShardResponse(msg, SHARD_RESP_PROTOCOL_VERSION);
}
std::array<size_t, 2> shardMsgCount{0, 0};
for (auto& msg: _channel->protocolMessages(PROXY_SHARD_REQ_PROTOCOL_VERSION)) {
_handleShardRequest(msg, PROXY_SHARD_REQ_PROTOCOL_VERSION);
++shardMsgCount[msg.socketIx];
}
for (auto& msg : _channel->protocolMessages(CDC_TO_SHARD_REQ_PROTOCOL_VERSION)) {
_handleShardRequest(msg, CDC_TO_SHARD_REQ_PROTOCOL_VERSION);
}
for (auto& msg : _channel->protocolMessages(SHARD_REQ_PROTOCOL_VERSION)) {
_handleShardRequest(msg, SHARD_REQ_PROTOCOL_VERSION);
++shardMsgCount[msg.socketIx];
}
for (size_t i = 0; i < _shared.receivedRequests.size(); ++i) {
_shared.receivedRequests[i] = _shared.receivedRequests[i]*0.95 + ((double)shardMsgCount[i])*0.05;
}
// write out write requests to queues
_shared.writerShardRespQueue.push(_waitResponses);
auto pushed = _shared.logsDBRequestQueue.push(_logsDBRequests);
_shared.receivedLogsDBReqs.fetch_add(_logsDBRequests.size(), std::memory_order_relaxed);
_shared.droppedLogsDBReqs.fetch_add(_logsDBRequests.size() - pushed, std::memory_order_relaxed);
pushed = _shared.logsDBResponseQueue.push(_logsDBResponses);
_shared.receivedLogsDBResps.fetch_add(_logsDBResponses.size(), std::memory_order_relaxed);
_shared.droppedLogsDBResps.fetch_add(_logsDBResponses.size() - pushed, std::memory_order_relaxed);
pushed = _shared.proxyLogsDBRequestQueue.push(_proxyLogsDBRequests);
_shared.receivedProxyLogsDBReqs.fetch_add(_proxyLogsDBRequests.size(), std::memory_order_relaxed);
_shared.droppedProxyLogsDBReqs.fetch_add(_proxyLogsDBRequests.size() - pushed, std::memory_order_relaxed);
pushed = _shared.proxyLogsDBResponseQueue.push(_proxyLogsDBResponses);
_shared.receivedProxyLogsDBResps.fetch_add(_proxyLogsDBResponses.size(), std::memory_order_relaxed);
_shared.droppedProxyLogsDBResps.fetch_add(_proxyLogsDBResponses.size() - pushed, std::memory_order_relaxed);
pushed = _shared.writerProxyShardRespQueue.push(_proxyResponses);
_shared.receivedWriterProxyResps.fetch_add(_proxyResponses.size(), std::memory_order_relaxed);
_shared.droppedProxyWriteResps.fetch_add(_proxyResponses.size() - pushed, std::memory_order_relaxed);
pushed = _shared.writerShardReqQueue.push(_writeReqs);
_shared.receivedWriterReqs.fetch_add(_writeReqs.size(), std::memory_order_relaxed);
_shared.droppedWriteReqs.fetch_add(_writeReqs.size() - pushed, std::memory_order_relaxed);
// write out read requests to queue
auto readIt = _readRequests.begin();
auto readQueueIt = _readQueueIt;
for(;;) {
readIt = (*readQueueIt)->push(readIt, _readRequests.end());
if (++readQueueIt == _shared.readerRequestsQueues.end()) {
readQueueIt = _shared.readerRequestsQueues.begin();
}
if (readIt == _readRequests.end()) {
break;
}
if (readQueueIt == _readQueueIt) {
break;
}
}
_readQueueIt = readQueueIt;
_shared.receivedReaderReqs.fetch_add(_readRequests.size(), std::memory_order_relaxed);
_shared.droppedReadReqs.fetch_add(std::distance(readIt, _readRequests.end()), std::memory_order_relaxed);
}
};
struct InFlightRequestKey {
uint64_t requestId;
uint64_t portIp;
InFlightRequestKey(uint64_t requestId_, IpPort clientAddr) :
requestId(requestId_)
{
sockaddr_in addr;
clientAddr.toSockAddrIn(addr);
portIp = ((uint64_t) addr.sin_port << 32) | ((uint64_t) addr.sin_addr.s_addr);
}
bool operator==(const InFlightRequestKey& other) const {
return requestId == other.requestId && portIp == other.portIp;
}
};
template <>
struct std::hash<InFlightRequestKey> {
std::size_t operator()(const InFlightRequestKey& key) const {
return std::hash<uint64_t>{}(key.requestId ^ key.portIp);
}
};
struct ProxyShardReq {
ShardReq req;
TernTime lastSent;
TernTime created;
TernTime gotLogIdx;
TernTime finished;
};
struct CrossRegionLocationState {
TernTime lastSent;
uint64_t waitReqId;
};
struct CrossRegionWaitInfo {
LogIdx idx;
ShardReq originalRequest;
CdcToShardRespMsg responseToSend;
std::unordered_map<uint8_t, CrossRegionLocationState> locationStates; // location ID -> state
TernTime createdAt;
};
struct ShardWriter : Loop {
private:
const std::string _basePath;
ShardShared& _shared;
AES128Key _expandedShardKey;
AES128Key _expandedCDCKey;
// outgoing network
UDPSender _sender;
RandomGenerator _packetDropRand;
uint64_t _outgoingPacketDropProbability; // probability * 10,000
// work queue
const size_t _maxWorkItemsAtOnce;
LogsDB& _logsDB;
bool _isLogsDBLeader;
uint64_t _currentLogIndex;
LogIdx _knownLastReleased; // used by leaders in proxy location to determine if they need to catch up
Duration _nextTimeout;
// buffers for separated items from work queue
std::vector<LogsDBRequest> _logsDBRequests; // requests from other replicas
std::vector<LogsDBResponse> _logsDBResponses; // responses from other replicas
std::vector<ShardReq> _shardRequests; // requests from clients or proxy locations
std::vector<ProxyShardRespMsg> _shardResponses; // responses from shard leader in primary location
std::vector<ProxyLogsDBRequest> _proxyLogsDBRequests; // catchup requests from logs db leader in proxy location
// or write requests from logs db leader in primary location
std::vector<ProxyLogsDBResponse> _proxyLogsDBResponses; // responses from logs db leader in primary location
std::vector<ShardResp> _waitResponses; // responses to wait state applied messages
// buffers for outgoing local losgsdb requests/responses
std::vector<LogsDBRequest *> _logsDBOutRequests;
std::vector<LogsDBResponse> _logsDBOutResponses;
std::vector<LogsDBLogEntry> _logsDBEntries; // used for batching writes/reads to logsdb
std::vector<ShardLogEntry> _shardEntries; // used for batching logEntries
std::unordered_map<uint64_t, ShardLogEntry> _inFlightEntries; // used while primary leader to track in flight entries we produced
static constexpr Duration PROXIED_REUQEST_TIMEOUT = 100_ms;
std::unordered_map<uint64_t, ProxyShardReq> _proxyShardRequests; // outstanding proxied shard requests
std::unordered_map<uint64_t, std::pair<LogIdx, TernTime>> _proxyCatchupRequests; // outstanding logsdb catchup requests to primary leader
std::unordered_map<uint64_t, std::pair<ShardRespContainer, ProxyShardReq>> _proxiedResponses; // responses from primary location that we need to send back to client
std::vector<ProxyLogsDBRequest> _proxyReadRequests; // currently processing proxied read requests
std::vector<LogIdx> _proxyReadRequestsIndices; // indices from above reuqests as LogsDB api needs them
std::vector<ShardReq> _linkReadRequests; //TODO: make LogsDB support reads from different thread
static constexpr size_t CATCHUP_BUFFER_SIZE = 1 << 17;
std::array<std::pair<uint64_t,LogsDBLogEntry>,CATCHUP_BUFFER_SIZE> _catchupWindow;
uint64_t _catchupWindowIndex; // index into the catchup window
// The enqueued requests, but indexed by req id + ip + port. We
// store this so that we can drop repeated requests which are
// still queued, and which will therefore be processed in due
// time anyway. This relies on clients having unique req ids. It's
// kinda unsafe anyway (if clients get restarted), but it's such
// a useful optimization for now that we live with it.
std::unordered_set<InFlightRequestKey> _inFlightRequestKeys;
std::map<std::pair<LogIdx, uint64_t>, ShardReq> _waitStateRequests; // log idx + request id for requests waiting for logsdb to advance
// Cross-region wait tracking for CDC writes
static constexpr Duration CROSS_REGION_WAIT_RETRY_INTERVAL = 200_ms;
static constexpr Duration SHARD_DOWN_THRESHOLD = 10_mins;
std::unordered_map<uint64_t, CrossRegionWaitInfo> _crossRegionWaitResponses; // wait req id -> info
std::unordered_map<uint64_t, uint64_t> _waitReqIdToResponseId;
uint64_t _requestIdCounter;
std::unordered_map<uint64_t, ShardReq> _logIdToShardRequest; // used to track which log entry was generated by which request
std::shared_ptr<const std::array<AddrsInfo, LogsDB::REPLICA_COUNT>> _replicaInfo;
virtual void sendStop() override {
_shared.logsDBRequestQueue.close();
_shared.logsDBResponseQueue.close();
_shared.proxyLogsDBRequestQueue.close();
_shared.proxyLogsDBResponseQueue.close();
_shared.writerShardReqQueue.close();
_shared.writerShardRespQueue.close();
_shared.writerProxyShardRespQueue.close();
}
public:
ShardWriter(Logger& logger, std::shared_ptr<XmonAgent>& xmon, ShardShared& shared) :
Loop(logger, xmon, "writer"),
_basePath(shared.options.logsDBOptions.dbDir),
_shared(shared),
_sender(UDPSenderConfig{.maxMsgSize = MAX_UDP_MTU}),
_packetDropRand(ternNow().ns),
_outgoingPacketDropProbability(0),
_maxWorkItemsAtOnce(LogsDB::IN_FLIGHT_APPEND_WINDOW * 10),
_logsDB(shared.logsDB),
_isLogsDBLeader(false),
_currentLogIndex(_shared.shardDB.lastAppliedLogEntry()),
_knownLastReleased(_logsDB.getLastReleased()),
_nextTimeout(0),
_catchupWindowIndex(0),
_requestIdCounter(RandomGenerator().generate64())
{
expandKey(ShardKey, _expandedShardKey);
expandKey(CDCKey, _expandedCDCKey);
auto convertProb = [this](const std::string& what, double prob, uint64_t& iprob) {
if (prob != 0.0) {
LOG_INFO(_env, "Will drop %s%% of %s packets", prob*100.0, what);
iprob = prob * 10'000.0;
ALWAYS_ASSERT(iprob > 0 && iprob < 10'000);
}
};
memset(_catchupWindow.data(), 0, _catchupWindow.size()*sizeof(decltype(_catchupWindow)::value_type));
convertProb("outgoing", _shared.options.serverOptions.simulateOutgoingPacketDrop, _outgoingPacketDropProbability);
_shardEntries.reserve(LogsDB::IN_FLIGHT_APPEND_WINDOW);
}
virtual ~ShardWriter() = default;
void _sendProxyAndCatchupRequests() {
if (!_shared.options.isProxyLocation()) {
return;
}
auto leadersAtOtherLocations = _shared.getLeadersAtOtherLocations();
const AddrsInfo* primaryLeaderAddress = nullptr;
for(auto& leader : *leadersAtOtherLocations ) {
if (leader.locationId == 0) {
primaryLeaderAddress = &leader.addrs;
break;
}
}
if (primaryLeaderAddress == nullptr) {
LOG_DEBUG(_env, "don't have primary leader address yet");
return;
}
// catchup requests first as progressing state is more important then sending new requests
auto now = ternNow();
for(auto& req : _proxyCatchupRequests) {
if (now - req.second.second < PROXIED_REUQEST_TIMEOUT) {
continue;
}
req.second.second = now;
LogReqMsg reqMsg;
reqMsg.id = req.first;
reqMsg.body.setLogRead().idx = req.second.first;
_sender.prepareOutgoingMessage(
_env,
_shared.sock().addr(),
*primaryLeaderAddress,
[&reqMsg, this](BincodeBuf& buf) {
reqMsg.pack(buf, _expandedShardKey);
});
}
for(auto& req : _proxyShardRequests) {
if(now - req.second.lastSent < PROXIED_REUQEST_TIMEOUT) {
continue;
}
req.second.lastSent = now;
ProxyShardReqMsg reqMsg;
reqMsg.id = req.first;
switch (req.second.req.msg.body.kind()) {
case ShardMessageKind::ADD_SPAN_INITIATE:
{
auto& addSpanReq = reqMsg.body.setAddSpanAtLocationInitiate();
addSpanReq.locationId = _shared.options.logsDBOptions.location;
addSpanReq.req.reference = NULL_INODE_ID;
addSpanReq.req.req = req.second.req.msg.body.getAddSpanInitiate();
break;
}
case ShardMessageKind::ADD_SPAN_INITIATE_WITH_REFERENCE:
{
auto& addSpanReq = reqMsg.body.setAddSpanAtLocationInitiate();
addSpanReq.locationId = _shared.options.logsDBOptions.location;
addSpanReq.req = req.second.req.msg.body.getAddSpanInitiateWithReference();
break;
}
default:
reqMsg.body = req.second.req.msg.body;
}
_sender.prepareOutgoingMessage(
_env,
_shared.sock().addr(),
*primaryLeaderAddress,
[&reqMsg, this](BincodeBuf& buf) {
reqMsg.pack(buf, _expandedShardKey);
});
}
}
void _sendWaitStateAppliedResponses() {
auto lastApplied = _logsDB.getLastReleased();
for(auto it = _waitStateRequests.begin(); it != _waitStateRequests.end(); ) {
if (lastApplied < it->first.first.u64) {
break;
}
auto& req = it->second;
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
ShardRespMsg resp;
resp.id = req.msg.id;
resp.body.setWaitStateApplied();
_inFlightRequestKeys.erase(InFlightRequestKey{req.msg.id, req.clientAddr});
packShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, req, resp);
it = _waitStateRequests.erase(it);
}
}
void _sendCrossRegionWaitResponses() {
if (_crossRegionWaitResponses.empty()) {
return;
}
auto now = ternNow();
auto leadersAtOtherLocations = _shared.getLeadersAtOtherLocations();
for (auto it = _crossRegionWaitResponses.begin(); it != _crossRegionWaitResponses.end(); ) {
auto& waitInfo = it->second;
bool shouldSendResponse = false;
// Check if any locations need retry or are down
for (auto locIt = waitInfo.locationStates.begin(); locIt != waitInfo.locationStates.end(); ) {
uint8_t locationId = locIt->first;
auto& locState = locIt->second;
bool locationFound = false;
bool locationDown = false;
AddrsInfo locationAddrs;
for (const auto& leader : *leadersAtOtherLocations) {
if (leader.locationId == locationId) {
locationFound = true;
locationAddrs = leader.addrs;
auto timeSinceLastSeen = now - leader.lastSeen;
if (timeSinceLastSeen >= SHARD_DOWN_THRESHOLD) {
locationDown = true;
LOG_INFO(_env, "Location %s detected as down (no response for %s), removing from wait list for LogIdx %s",
locationId, timeSinceLastSeen, waitInfo.idx.u64);
}
break;
}
}
if (!locationFound || locationDown) {
_waitReqIdToResponseId.erase(locState.waitReqId);
locIt = waitInfo.locationStates.erase(locIt);
continue;
}
auto timeSinceLastSent = now - locState.lastSent;
if (timeSinceLastSent >= CROSS_REGION_WAIT_RETRY_INTERVAL) {
uint64_t waitReqId = locState.waitReqId;
ShardReqMsg waitReq;
waitReq.id = waitReqId;
auto& waitBody = waitReq.body.setWaitStateApplied();
waitBody.idx.u64 = waitInfo.idx.u64;
_sender.prepareOutgoingMessage(
_env,
_shared.sock().addr(),
locationAddrs,
[&waitReq](BincodeBuf& buf) {
waitReq.pack(buf);
});
LOG_DEBUG(_env, "Retrying WaitStateAppliedReq (id=%s) for LogIdx %s to location %s",
waitReqId, waitInfo.idx.u64, locationId);
locState.lastSent = now;
}
++locIt;
}
if (waitInfo.locationStates.empty()) {
LOG_DEBUG(_env, "All secondary locations have applied LogIdx %s, sending response for request %s",
waitInfo.idx.u64, waitInfo.originalRequest.msg.id);
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
packCheckPointedShardResponse(_env, _shared, _shared.sock().addr(), _sender,
dropArtificially, waitInfo.originalRequest,
waitInfo.responseToSend, _expandedCDCKey);
ALWAYS_ASSERT(_inFlightRequestKeys.erase(InFlightRequestKey{waitInfo.originalRequest.msg.id,
waitInfo.originalRequest.clientAddr}) == 1);
it = _crossRegionWaitResponses.erase(it);
} else {
++it;
}
}
}
void _tryReplicateToOtherLocations() {
// if we are not leader or in proxy location already we will not try to replicate
if (!_isLogsDBLeader || _shared.options.isProxyLocation()) {
return;
}
auto leadersAtOtherLocations = _shared.getLeadersAtOtherLocations();
if (leadersAtOtherLocations->empty()) {
return;
}
auto lastReleased = _logsDB.getLastReleased();
// for each entry we fire off a no response best effort write to other locations
// if some of the leaders are down or messages is lost they will catch up through othe means
for (auto& logsDBEntry : _logsDBEntries) {
LogReqMsg msg;
msg.id = 0;
auto& write = msg.body.setLogWrite();
write.idx = logsDBEntry.idx;
write.value.els = std::move(logsDBEntry.value);
write.lastReleased = lastReleased;
for (const auto& receiver : *leadersAtOtherLocations) {
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
if (unlikely(dropArtificially)) {
LOG_DEBUG(_env, "artificially dropping replication write %s to shard %s at location %s", write.idx, receiver.id, receiver.locationId);
return;
}
_sender.prepareOutgoingMessage(
_env,
_shared.sock().addr(),
receiver.addrs,
[&msg, this](BincodeBuf& buf) {
msg.pack(buf, _expandedShardKey);
});
LOG_DEBUG(_env, "sending replication write %s to shard %s at location %s", write.idx, receiver.id, receiver.locationId);
}
}
}
void _applyLogEntries() {
ALWAYS_ASSERT(_logsDBEntries.empty());
auto lastContinuousIdx = _logsDB.getLastContinuous();
while (_currentLogIndex < lastContinuousIdx.u64) {
_logsDB.readEntries(_logsDBEntries);
ALWAYS_ASSERT(!_logsDBEntries.empty());
for (auto& logsDBEntry : _logsDBEntries) {
++_currentLogIndex;
ALWAYS_ASSERT(_currentLogIndex == logsDBEntry.idx);
ALWAYS_ASSERT(logsDBEntry.value.size() > 0);
BincodeBuf buf((char*)&logsDBEntry.value.front(), logsDBEntry.value.size());
ShardLogEntry shardEntry;
shardEntry.unpack(buf);
ALWAYS_ASSERT(_currentLogIndex == shardEntry.idx);
if (_isLogsDBLeader) {
{
//this is sanity check confirming what we got through the log and deserialized
//exactly matches what we serialized and pushed to log
//this is only ever true in primary location
auto it = _inFlightEntries.find(shardEntry.idx.u64);
ALWAYS_ASSERT(it != _inFlightEntries.end());
ALWAYS_ASSERT(shardEntry == it->second);
ALWAYS_ASSERT(shardEntry.idx == it->second.idx);
_inFlightEntries.erase(it);
}
auto it = _logIdToShardRequest.find(shardEntry.idx.u64);
if (it == _logIdToShardRequest.end()) {
// if we are primary location, all writes should be triggered by requests
ALWAYS_ASSERT(_shared.options.isProxyLocation());
// we are proxy location there are writes not initiated by us and we behave like follower
// we are not leader, we can not do any checks and there is no response to send
ShardRespContainer _;
_shared.shardDB.applyLogEntry(logsDBEntry.idx.u64, shardEntry, _);
continue;
}
auto& request = it->second;
if (likely(request.msg.id)) {
LOG_DEBUG(_env, "applying log entry for request %s kind %s from %s", request.msg.id, request.msg.body.kind(), request.clientAddr);
} else {
LOG_DEBUG(_env, "applying request-less log entry");
}
// first handle case where client does not care about response
if (request.msg.id == 0) {
ShardRespContainer resp;
_shared.shardDB.applyLogEntry(logsDBEntry.idx.u64, shardEntry, resp);
if (unlikely(resp.kind() == ShardMessageKind::ERROR)) {
RAISE_ALERT(_env, "could not apply request-less log entry: %s", resp.getError());
}
_logIdToShardRequest.erase(it);
continue;
}
// depending on protocol we need different kind of responses
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
bool shouldWait = false;
switch(request.protocol){
case SHARD_REQ_PROTOCOL_VERSION:
{
ShardRespMsg resp;
_shared.shardDB.applyLogEntry(logsDBEntry.idx.u64, shardEntry, resp.body);
resp.id = request.msg.id;
auto it = _proxiedResponses.find(logsDBEntry.idx.u64);
if (it != _proxiedResponses.end()) {
ALWAYS_ASSERT(_shared.options.isProxyLocation());
it->second.second.finished = ternNow();
logSlowProxyReq(it->second.second);
resp.body = std::move(it->second.first);
_proxiedResponses.erase(it);
}
if (resp.body.kind() == ShardMessageKind::ADD_SPAN_AT_LOCATION_INITIATE) {
ShardRespContainer tmpResp;
switch (request.msg.body.kind()) {
case ShardMessageKind::ADD_SPAN_INITIATE:
{
auto& addResp = tmpResp.setAddSpanInitiate();
addResp.blocks = std::move(resp.body.getAddSpanAtLocationInitiate().resp.blocks);
resp.body.setAddSpanInitiate().blocks = std::move(addResp.blocks);
break;
}
case ShardMessageKind::ADD_SPAN_INITIATE_WITH_REFERENCE:
{
auto& addResp = tmpResp.setAddSpanInitiateWithReference();
addResp.resp.blocks = std::move(resp.body.getAddSpanAtLocationInitiate().resp.blocks);
resp.body.setAddSpanInitiateWithReference().resp.blocks = std::move(addResp.resp.blocks);
break;
}
case ShardMessageKind::ADD_SPAN_AT_LOCATION_INITIATE:
{
break;
}
default:
ALWAYS_ASSERT(false, "Unexpected reponse kind %s for requests kind %s", resp.body.kind(), request.msg.body.kind() );
}
}
packShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, request, resp);
}
break;
case CDC_TO_SHARD_REQ_PROTOCOL_VERSION:
{
CdcToShardRespMsg resp;
resp.body.checkPointIdx = shardEntry.idx;
resp.id = request.msg.id;
_shared.shardDB.applyLogEntry(logsDBEntry.idx.u64, shardEntry, resp.body.resp);
// For CDC requests, wait for all secondary leaders to apply
// before sending response
auto now = ternNow();
auto leadersAtOtherLocations = _shared.getLeadersAtOtherLocations();
if ((!leadersAtOtherLocations->empty()) && resp.body.resp.kind() == ShardMessageKind::CREATE_DIRECTORY_INODE) {
CrossRegionWaitInfo waitInfo;
waitInfo.idx = shardEntry.idx;
waitInfo.createdAt = now;
uint64_t responseId = request.msg.id;
for (const auto& leader : *leadersAtOtherLocations) {
if (leader.locationId == 0) continue;
auto timeSinceLastSeen = now - leader.lastSeen;
if (timeSinceLastSeen >= SHARD_DOWN_THRESHOLD) {
LOG_INFO(_env, "Skipping wait for secondary leader at location %s (down for %s)",
leader.locationId, timeSinceLastSeen);
continue;
}
shouldWait = true;
uint64_t waitReqId = ++_requestIdCounter;
CrossRegionLocationState locState;
locState.lastSent = now;
locState.waitReqId = waitReqId;
waitInfo.locationStates[leader.locationId] = locState;
_waitReqIdToResponseId[waitReqId] = responseId;
ShardReqMsg waitReq;
waitReq.id = waitReqId;
auto& waitBody = waitReq.body.setWaitStateApplied();
waitBody.idx.u64 = shardEntry.idx.u64;
_sender.prepareOutgoingMessage(
_env,
_shared.sock().addr(),
leader.addrs,
[&waitReq](BincodeBuf& buf) {
waitReq.pack(buf);
});
LOG_DEBUG(_env, "Sent WaitStateAppliedReq (id=%s) for LogIdx %s to location %s",
waitReqId, shardEntry.idx.u64, leader.locationId);
}
if (shouldWait) {
waitInfo.originalRequest = std::move(request);
waitInfo.responseToSend = std::move(resp);
_crossRegionWaitResponses[responseId] = std::move(waitInfo);
}
}
if (!shouldWait) {
packCheckPointedShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, request, resp, _expandedCDCKey);
}
}
break;
case PROXY_SHARD_REQ_PROTOCOL_VERSION:
{
ProxyShardRespMsg resp;
resp.body.checkPointIdx = shardEntry.idx;
resp.id = request.msg.id;
_shared.shardDB.applyLogEntry(logsDBEntry.idx.u64, shardEntry, resp.body.resp);
packCheckPointedShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, request, resp, _expandedShardKey);
}
break;
}
if (!shouldWait) {
ALWAYS_ASSERT(_inFlightRequestKeys.erase(InFlightRequestKey{request.msg.id, request.clientAddr}) == 1);
}
_logIdToShardRequest.erase(it);
} else {
// we are not leader, we can not do any checks and there is no response to send
ShardRespContainer _;
_shared.shardDB.applyLogEntry(logsDBEntry.idx.u64, shardEntry, _);
}
}
// we send new LogsDB entry to leaders in other locations
_tryReplicateToOtherLocations();
_logsDBEntries.clear();
}
}
void _processCathupReads() {
if (_proxyReadRequests.empty()) {
return;
}
std::sort(
_proxyReadRequests.begin(),
_proxyReadRequests.end(),
[](const auto& lhs, const auto& rhs) {
return lhs.request.msg.body.getLogRead().idx < rhs.request.msg.body.getLogRead().idx;
}
);
ALWAYS_ASSERT(_logsDBEntries.empty());
ALWAYS_ASSERT(_proxyReadRequestsIndices.empty());
_proxyReadRequestsIndices.reserve(_proxyReadRequests.size());
for (auto& request : _proxyReadRequests) {
_proxyReadRequestsIndices.emplace_back(request.request.msg.body.getLogRead().idx);
}
_logsDB.readIndexedEntries(_proxyReadRequestsIndices, _logsDBEntries);
ALWAYS_ASSERT(_logsDBEntries.size() == _proxyReadRequestsIndices.size());
for (size_t i = 0; i < _proxyReadRequestsIndices.size(); ++i) {
auto& request = _proxyReadRequests[i];
if (_logsDBEntries[i].idx == 0) {
LOG_ERROR(_env, "We failed reading a catchup entry for index %s", request.request.msg.body.getLogRead().idx);
continue;
}
LogRespMsg resp;
resp.id = request.request.msg.id;
auto& readResp = resp.body.setLogRead();
readResp.result = TernError::NO_ERROR;
readResp.value.els = _logsDBEntries[i].value;
_sender.prepareOutgoingMessage(
_env,
_shared.sock().addr(),
request.sockIx,
request.clientAddr,
[&resp,this](BincodeBuf& buf) {
resp.pack(buf, _expandedShardKey);
});
}
_proxyReadRequests.clear();
_proxyReadRequestsIndices.clear();
_logsDBEntries.clear();
}
void _processLinkReadRequest(const ShardReq& req) {
ShardRespMsg resp;
resp.id = req.msg.id;
auto& respBody = resp.body.setGetLinkEntries();
auto& reqBody = req.msg.body.getGetLinkEntries();
respBody.nextIdx = std::max(reqBody.fromIdx, _logsDB.getHeadIdx());
std::vector<LogIdx> indexes;
std::vector<LogsDBLogEntry> readEntries;
indexes.reserve(128);
readEntries.reserve(128);
size_t processed = 0;
const size_t maxProcess = 1024;
auto mtu = std::min(reqBody.mtu == 0 ? (int)DEFAULT_UDP_MTU : (int)reqBody.mtu, (int)MAX_UDP_MTU);
int budget = mtu - (int)ShardRespMsg::STATIC_SIZE - (int)GetLinkEntriesResp::STATIC_SIZE;
const size_t maxEntriesByMtu = budget / (int)LinkEntry::STATIC_SIZE;
while (respBody.nextIdx < _logsDB.getLastReleased() && respBody.entries.els.size() < maxEntriesByMtu && processed < maxProcess) {
LogIdx startIdx = respBody.nextIdx;
readEntries.clear();
indexes.clear();
for (int i = 0; i < 128 && processed < maxProcess; ++i) {
if (_logsDB.getLastReleased() < startIdx) {
break;
}
indexes.push_back(startIdx);
++startIdx;
++processed;
}
if (indexes.empty()) {
break;
}
_logsDB.readIndexedEntries(indexes, readEntries);
for (auto& entry: readEntries) {
if (entry.idx == 0) {
++respBody.nextIdx;
continue;
}
respBody.nextIdx = entry.idx;
ShardLogEntry logEntry;
BincodeBuf buf((char*)entry.value.data(), entry.value.size());
logEntry.unpack(buf);
if (logEntry.body.kind() != ShardLogEntryKind::LINK_FILE) {
continue;
}
auto& link = logEntry.body.getLinkFile();
LinkEntry newEntry;
newEntry.id = link.fileId;
newEntry.ownerId = link.ownerId;
newEntry.time = logEntry.time;
respBody.entries.els.emplace_back(newEntry);
if (respBody.entries.els.size() >= maxEntriesByMtu) {
break;
}
}
}
packShardResponse(_env, _shared, _shared.sock().addr(), _sender, false, req, resp);
}
void _processCatchupWindow() {
if (!_isLogsDBLeader || !_shared.options.isProxyLocation()) {
return;
}
ALWAYS_ASSERT(_shardEntries.empty());
ALWAYS_ASSERT(_knownLastReleased.u64 >= _currentLogIndex + _inFlightEntries.size());
// first we move any continuous entries from cathup window and schedule them for replication
auto maxToWrite = LogsDB::IN_FLIGHT_APPEND_WINDOW - _inFlightEntries.size();
auto expectedLogIdx = _currentLogIndex + _inFlightEntries.size() + 1;
while(maxToWrite > 0) {
--maxToWrite;
if (_catchupWindow[_catchupWindowIndex%_catchupWindow.size()].second.idx == 0) {
// missing entry
break;
}
auto& logsDBEntry = _catchupWindow[_catchupWindowIndex%_catchupWindow.size()].second;
// LogIndex should match
ALWAYS_ASSERT(expectedLogIdx == logsDBEntry.idx);
// there should be no request in flight for it
ALWAYS_ASSERT(_catchupWindow[_catchupWindowIndex%_catchupWindow.size()].first == 0);
BincodeBuf buf((char*)&logsDBEntry.value.front(), logsDBEntry.value.size());
auto& shardEntry = _shardEntries.emplace_back();
shardEntry.unpack(buf);
shardEntry.idx = expectedLogIdx;
expectedLogIdx++;
_catchupWindowIndex++;
//clear catchup window element
logsDBEntry.idx = 0;
logsDBEntry.value.clear();
}
auto maxCatchupElements = std::min(_knownLastReleased.u64 - (_currentLogIndex + _inFlightEntries.size() + _shardEntries.size()), _catchupWindow.size());
// schedule any catchup requests for missing entries
for (uint64_t i = 0 ; i < maxCatchupElements; ++i) {
auto &catchupEntry = _catchupWindow[(_catchupWindowIndex+i)%_catchupWindow.size()];
if (catchupEntry.first != 0 || catchupEntry.second.idx != 0) {
// already requested or already have it
continue;
}
if (_proxyCatchupRequests.size() == LogsDB::CATCHUP_WINDOW * 4) {
break;
}
catchupEntry.first = ++_requestIdCounter;
_proxyCatchupRequests.insert({catchupEntry.first, {expectedLogIdx+i, 0}});
}
}
void logSlowProxyReq(const ProxyShardReq& req) {
auto elapsed = req.finished - req.created;
if (elapsed > 5_sec) {
LOG_ERROR(_env, "slow proxy requests, elapsed %s, waiting response %s, waiting log apply %s", elapsed, req.gotLogIdx - req.created, req.finished-req.gotLogIdx);
}
}
void logsDBStep() {
_logsDB.processIncomingMessages(_logsDBRequests,_logsDBResponses);
_knownLastReleased = std::max(_knownLastReleased,_logsDB.getLastReleased());
_nextTimeout = _logsDB.getNextTimeout();
auto now = ternNow();
// check for leadership state change and clean up internal state
if (unlikely(_isLogsDBLeader != _logsDB.isLeader())) {
if (_logsDB.isLeader()) {
// we should not add anything to these if not leader
ALWAYS_ASSERT(_inFlightEntries.empty());
ALWAYS_ASSERT(_proxyCatchupRequests.empty());
ALWAYS_ASSERT(_inFlightRequestKeys.empty());
ALWAYS_ASSERT(_proxyShardRequests.empty());
ALWAYS_ASSERT(_proxiedResponses.empty());
// during leader election relase point might have moved and there could be entries we have not applied yet
_applyLogEntries();
_knownLastReleased = _logsDB.getLastReleased();
} else {
LOG_INFO(_env, "We are no longer leader in LogsDB");
_inFlightEntries.clear();
_proxyCatchupRequests.clear();
_inFlightRequestKeys.clear();
_proxyShardRequests.clear();
_proxiedResponses.clear();
for(auto& entry : _catchupWindow) {
entry.first = 0;
entry.second.idx = 0;
}
_catchupWindowIndex = 0;
}
_isLogsDBLeader = _logsDB.isLeader();
}
if (!_isLogsDBLeader) {
// clear things we don't serve or expect as a follower
_shardRequests.clear();
_shardResponses.clear();
_proxyLogsDBRequests.clear();
_proxyLogsDBResponses.clear();
_waitResponses.clear();
}
_linkReadRequests.clear();
if (!_shared.options.isProxyLocation()) {
// we don't expect and should not process any shard responses if in primary location
_shardResponses.clear();
}
// there could be log entries just released which we should apply
_applyLogEntries();
// if we are leader we should alyways have latest state applied
ALWAYS_ASSERT(!_isLogsDBLeader || _currentLogIndex == _logsDB.getLastReleased());
ShardReq snapshotReq;
// prepare log entries for requests or proxy them to primary leader
ALWAYS_ASSERT(_shardEntries.empty());
for (auto& req : _shardRequests) {
if (req.msg.id != 0 && _inFlightRequestKeys.contains(InFlightRequestKey{req.msg.id, req.clientAddr})) {
// we already have a request in flight with this id from this client
continue;
}
switch (req.msg.body.kind()) {
case ShardMessageKind::SHARD_SNAPSHOT:
snapshotReq = std::move(req);
continue;
case ShardMessageKind::GET_LINK_ENTRIES:
_linkReadRequests.emplace_back(std::move(req));
continue;
case ShardMessageKind::WAIT_STATE_APPLIED:
{
auto& waitReq = req.msg.body.getWaitStateApplied();
auto logIdx = LogIdx{waitReq.idx.u64};
if (logIdx.u64 <= _logsDB.getLastReleased().u64) {
// already applied
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
ShardRespMsg resp;
resp.id = req.msg.id;
resp.body.setWaitStateApplied();
packShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, req, resp);
} else {
_inFlightRequestKeys.insert(InFlightRequestKey{req.msg.id, req.clientAddr});
_waitStateRequests.emplace(std::make_pair(logIdx, req.msg.id), std::move(req));
}
}
continue;
default:
break;
}
if (_shared.options.isProxyLocation()) {
if (req.msg.id != 0) {
_inFlightRequestKeys.insert(InFlightRequestKey{req.msg.id, req.clientAddr});
}
// we send them out later along with timed out ones
_proxyShardRequests.insert({++_requestIdCounter, ProxyShardReq{std::move(req), 0, now, 0, 0}});
continue;
}
if (_inFlightEntries.size() + _shardEntries.size() == LogsDB::IN_FLIGHT_APPEND_WINDOW) {
// we have reached the limit of in flight entries anything else will be dropped anyway
// clients will retry
continue;
}
auto& entry = _shardEntries.emplace_back();
auto err = _shared.shardDB.prepareLogEntry(req.msg.body, entry);
if (unlikely(err != TernError::NO_ERROR)) {
_shardEntries.pop_back(); // back out the log entry
LOG_ERROR(_env, "error preparing log entry for request: %s from: %s err: %s", req.msg, req.clientAddr, err);
// depending on protocol we need different kind of responses
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
switch(req.protocol){
case SHARD_REQ_PROTOCOL_VERSION:
{
ShardRespMsg resp;
resp.id = req.msg.id;
resp.body.setError() = err;
packShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, req, resp);
}
break;
case CDC_TO_SHARD_REQ_PROTOCOL_VERSION:
{
CdcToShardRespMsg resp;
resp.body.checkPointIdx = _logsDB.getLastReleased();
resp.id = req.msg.id;
resp.body.resp.setError() = err;
packCheckPointedShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, req, resp, _expandedCDCKey);
}
break;
case PROXY_SHARD_REQ_PROTOCOL_VERSION:
{
ProxyShardRespMsg resp;
resp.body.checkPointIdx = _logsDB.getLastReleased();
resp.id = req.msg.id;
resp.body.resp.setError() = err;
packCheckPointedShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, req, resp, _expandedShardKey);
}
break;
}
}
entry.idx = _currentLogIndex + _inFlightEntries.size() + _shardEntries.size();
// requests with id 0 are "one off, don't want response, will not retry" so we assume that if we receive it twice we need to execute it twice
if (req.msg.id != 0) {
_inFlightRequestKeys.insert(InFlightRequestKey{req.msg.id, req.clientAddr});
}
_logIdToShardRequest.insert({entry.idx.u64, std::move(req)});
}
for(auto& waitResp : _waitResponses) {
ALWAYS_ASSERT(waitResp.msg.body.kind() == ShardMessageKind::WAIT_STATE_APPLIED);
auto waitReqIt = _waitReqIdToResponseId.find(waitResp.msg.id);
if (waitReqIt != _waitReqIdToResponseId.end()) {
uint64_t responseId = waitReqIt->second;
auto waitIt = _crossRegionWaitResponses.find(responseId);
if (waitIt != _crossRegionWaitResponses.end()) {
// Match source address to location ID
auto leadersAtOtherLocations = _shared.getLeadersAtOtherLocations();
uint8_t respondingLocationId = 255; // Invalid location ID
for (const auto& leader : *leadersAtOtherLocations) {
if (leader.locationId == 0) {
continue;
}
if (leader.addrs.contains(waitResp.clientAddr)) {
respondingLocationId = leader.locationId;
break;
}
}
auto locIt = waitIt->second.locationStates.find(respondingLocationId);
if (locIt == waitIt->second.locationStates.end()) {
// it's possible leader address change and we could not match it
// we just ignore this response as we will retry later
continue;
}
auto& locState = locIt->second;
waitIt->second.locationStates.erase(locIt);
LOG_DEBUG(_env, "Received WaitStateAppliedResp for LogIdx %s from location %s, %s locations remaining",
waitIt->second.idx.u64, respondingLocationId,
waitIt->second.locationStates.size());
_waitReqIdToResponseId.erase(waitReqIt);
}
}
}
for(auto& resp: _shardResponses) {
auto it = _proxyShardRequests.find(resp.id);
if (it == _proxyShardRequests.end()) {
// it is possible we already replied to this request and removed it from the map
continue;
}
auto& req = it->second.req;
it->second.gotLogIdx = now;
// it is possible we already applied the log entry, forward the response
if (resp.body.checkPointIdx <= _logsDB.getLastContinuous()) {
if (likely(req.msg.id)) {
LOG_DEBUG(_env, "applying log entry for request %s kind %s from %s", req.msg.id, req.msg.body.kind(), req.clientAddr);
} else {
LOG_DEBUG(_env, "applying request-less log entry");
// client does not care about response
_proxyShardRequests.erase(it);
continue;
}
it->second.finished = now;
logSlowProxyReq(it->second);
// depending on protocol we need different kind of responses
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
ALWAYS_ASSERT(req.protocol == SHARD_REQ_PROTOCOL_VERSION);
{
ShardRespMsg forwarded_resp;
forwarded_resp.id = req.msg.id;
forwarded_resp.body = std::move(resp.body.resp);
if (forwarded_resp.body.kind() == ShardMessageKind::ADD_SPAN_AT_LOCATION_INITIATE) {
ShardRespContainer tmpResp;
switch (req.msg.body.kind()) {
case ShardMessageKind::ADD_SPAN_INITIATE:
{
auto& addResp = tmpResp.setAddSpanInitiate();
addResp.blocks = std::move(forwarded_resp.body.getAddSpanAtLocationInitiate().resp.blocks);
forwarded_resp.body.setAddSpanInitiate().blocks = std::move(addResp.blocks);
break;
}
case ShardMessageKind::ADD_SPAN_INITIATE_WITH_REFERENCE:
{
auto& addResp = tmpResp.setAddSpanInitiateWithReference();
addResp.resp.blocks = std::move(forwarded_resp.body.getAddSpanAtLocationInitiate().resp.blocks);
forwarded_resp.body.setAddSpanInitiateWithReference().resp.blocks = std::move(addResp.resp.blocks);
break;
}
case ShardMessageKind::ADD_SPAN_AT_LOCATION_INITIATE:
{
break;
}
default:
ALWAYS_ASSERT(false, "Unexpected reponse kind %s for requests kind %s", forwarded_resp.body.kind(), req.msg.body.kind() );
}
}
packShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, req, forwarded_resp);
}
ALWAYS_ASSERT(_inFlightRequestKeys.erase(InFlightRequestKey{req.msg.id, req.clientAddr}) == 1);
_proxyShardRequests.erase(it);
continue;
}
if (resp.body.checkPointIdx.u64 > _knownLastReleased.u64 + 1000) {
LOG_ERROR(_env, "new last known released on proxy response receive, lag of %s",resp.body.checkPointIdx.u64 - _knownLastReleased.u64 );
}
_knownLastReleased = std::max(_knownLastReleased, resp.body.checkPointIdx);
if (resp.body.checkPointIdx.u64 - _logsDB.getLastReleased().u64 > 1000) {
LOG_ERROR(_env, "large log lag on proxy response receive %s", resp.body.checkPointIdx.u64 - _logsDB.getLastReleased().u64);
}
// we have the response but we will not reply immediately as we need to wait for logsDB to apply the log entry to guarantee
// read your own writes
_logIdToShardRequest.insert({resp.body.checkPointIdx.u64, std::move(req)});
_proxiedResponses.insert({resp.body.checkPointIdx.u64, std::pair<ShardRespContainer, ProxyShardReq>({std::move(resp.body.resp), std::move(it->second)})});
_proxyShardRequests.erase(it);
}
for (auto &resp : _proxyLogsDBResponses) {
auto it = _proxyCatchupRequests.find(resp.response.msg.id);
if (it == _proxyCatchupRequests.end()) {
continue;
}
if (resp.response.msg.body.kind() != LogMessageKind::LOG_READ) {
LOG_ERROR(_env, "Unexpected LogsDB proxy response kind: %s", resp.response.msg.body.kind());
continue;
}
auto logIdx = it->second.first;
// if we had an outstanding request this log index should be something we have not already written
// are not currently replicating within location (inFlightEntries)
ALWAYS_ASSERT(logIdx.u64 > _currentLogIndex);
ALWAYS_ASSERT(logIdx.u64 - _currentLogIndex > _inFlightEntries.size());
_proxyCatchupRequests.erase(it);
auto offsetFromLastReplication = logIdx.u64 - _currentLogIndex - _inFlightEntries.size() - 1;
// we never request things out of catchup window
ALWAYS_ASSERT(offsetFromLastReplication < _catchupWindow.size());
auto catchupWindowOffset = (_catchupWindowIndex + offsetFromLastReplication) % _catchupWindow.size();
// request should match and it should only be procesed once
ALWAYS_ASSERT(_catchupWindow[catchupWindowOffset].first == resp.response.msg.id);
ALWAYS_ASSERT(_catchupWindow[catchupWindowOffset].second.idx == 0);
_catchupWindow[catchupWindowOffset].first = 0;
_catchupWindow[catchupWindowOffset].second.idx = logIdx;
_catchupWindow[catchupWindowOffset].second.value = std::move(resp.response.msg.body.getLogRead().value.els);
}
for (auto &req : _proxyLogsDBRequests) {
switch(req.request.msg.body.kind()) {
case LogMessageKind::LOG_WRITE:
{
auto& write= req.request.msg.body.getLogWrite();
if (_knownLastReleased + 1000 < write.lastReleased) {
LOG_ERROR(_env, "received large update for _lastKnownRelaease %s", write.lastReleased.u64 - _knownLastReleased.u64);
}
_knownLastReleased = std::max(_knownLastReleased, write.lastReleased);
if (write.idx.u64 <= _currentLogIndex + _inFlightEntries.size()) {
LOG_ERROR(_env, "we have received write from leader but we are already replicating it. This should never happen");
break;
}
if (write.idx.u64 > _currentLogIndex + _inFlightEntries.size() + _catchupWindow.size()) {
// write outside of catchup window, we can't store it yet so we ignore. lastReleased was updated
break;
}
auto offsetFromLastReplication = write.idx.u64 - _currentLogIndex - _inFlightEntries.size() - 1;
auto catchupWindowOffset = (_catchupWindowIndex + offsetFromLastReplication) % _catchupWindow.size();
if (_catchupWindow[catchupWindowOffset].first != 0) {
// we have outstanding request for this entry
auto it = _proxyCatchupRequests.find(_catchupWindow[catchupWindowOffset].first);
ALWAYS_ASSERT(it != _proxyCatchupRequests.end());
ALWAYS_ASSERT(it->second.first == write.idx);
_proxyCatchupRequests.erase(it);
_catchupWindow[catchupWindowOffset].first = 0;
}
_catchupWindow[catchupWindowOffset].second.idx = write.idx;
_catchupWindow[catchupWindowOffset].second.value = std::move(write.value.els);
break;
}
case LogMessageKind::LOG_READ:
_proxyReadRequests.emplace_back(std::move(req));
break;
default:
LOG_ERROR(_env, "Unexpected LogsDB proxy request kind: %s", req.request.msg.body.kind());
break;
}
}
_processCathupReads();
_processCatchupWindow();
if (!_shardEntries.empty()) {
ALWAYS_ASSERT(_isLogsDBLeader);
ALWAYS_ASSERT(_logsDBEntries.empty());
_logsDBEntries.reserve(_shardEntries.size());
std::array<uint8_t, MAX_UDP_MTU> data;
for(auto& entry : _shardEntries) {
auto& logsDBEntry = _logsDBEntries.emplace_back();;
BincodeBuf buf((char*)&data[0], MAX_UDP_MTU);
entry.pack(buf);
logsDBEntry.value.assign(buf.data, buf.cursor);
}
auto err = _logsDB.appendEntries(_logsDBEntries);
ALWAYS_ASSERT(err == TernError::NO_ERROR);
for (size_t i = 0; i < _shardEntries.size(); ++i) {
ALWAYS_ASSERT(_logsDBEntries[i].idx == _shardEntries[i].idx);
_inFlightEntries.emplace(_shardEntries[i].idx.u64, std::move(_shardEntries[i]));
}
_logsDBEntries.clear();
}
// Log if not active is not chaty but it's messages are higher priority as they make us progress state under high load.
// We want to have priority when sending out
_logsDB.getOutgoingMessages(_logsDBOutRequests, _logsDBOutResponses);
for (auto& response : _logsDBOutResponses) {
packLogsDBResponse(response);
}
for (auto request : _logsDBOutRequests) {
packLogsDBRequest(*request);
}
_sendProxyAndCatchupRequests();
_sendWaitStateAppliedResponses();
_sendCrossRegionWaitResponses();
_shared.shardDB.flush(true);
for (auto& req : _linkReadRequests) {
_processLinkReadRequest(req);
}
// not needed as we just flushed and apparently it does actually flush again
// _logsDB.flush(true);
// snapshot is processed once and last as it will initiate a flush
if (unlikely(snapshotReq.msg.body.kind() == ShardMessageKind::SHARD_SNAPSHOT)) {
ShardRespMsg resp;
resp.id = snapshotReq.msg.id;
auto err = _shared.sharedDB.snapshot(_basePath +"/snapshot-" + std::to_string(snapshotReq.msg.body.getShardSnapshot().snapshotId));
if (err == TernError::NO_ERROR) {
resp.body.setShardSnapshot();
} else {
resp.body.setError() = err;
}
packShardResponse(_env, _shared, _shared.sock().addr(), _sender, false, snapshotReq, resp);
}
_sender.sendMessages(_env, _shared.sock());
}
const AddrsInfo* addressFromReplicaId(ReplicaId id) {
if (!_replicaInfo) {
return nullptr;
}
auto& addr = (*_replicaInfo)[id.u8];
if (addr[0].port == 0) {
return nullptr;
}
return &addr;
}
void packLogsDBResponse(LogsDBResponse& response) {
auto addrInfoPtr = addressFromReplicaId(response.replicaId);
if (unlikely(addrInfoPtr == nullptr)) {
LOG_DEBUG(_env, "No information for replica id %s. dropping response", response.replicaId);
return;
}
auto& addrInfo = *addrInfoPtr;
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
if (unlikely(dropArtificially)) {
LOG_DEBUG(_env, "artificially dropping response %s", response.msg.id);
return;
}
_sender.prepareOutgoingMessage(
_env,
_shared.sock().addr(),
addrInfo,
[&response,this](BincodeBuf& buf) {
response.msg.pack(buf, _expandedShardKey);
});
LOG_DEBUG(_env, "will send response for req id %s kind %s to %s", response.msg.id, response.msg.body.kind(), addrInfo);
}
void packLogsDBRequest(LogsDBRequest& request) {
auto addrInfoPtr = addressFromReplicaId(request.replicaId);
if (unlikely(addrInfoPtr == nullptr)) {
LOG_DEBUG(_env, "No information for replica id %s. dropping request", request.replicaId);
return;
}
auto& addrInfo = *addrInfoPtr;
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
if (unlikely(dropArtificially)) {
LOG_DEBUG(_env, "artificially dropping request %s", request.msg.id);
return;
}
_sender.prepareOutgoingMessage(
_env,
_shared.sock().addr(),
addrInfo,
[&request,this](BincodeBuf& buf) {
request.msg.pack(buf, _expandedShardKey);
});
LOG_DEBUG(_env, "will send request for req id %s kind %s to %s", request.msg.id, request.msg.body.kind(), addrInfo);
}
virtual void step() override {
_logsDBRequests.clear();
_logsDBResponses.clear();
_logsDBOutRequests.clear();
_logsDBOutResponses.clear();
_shardRequests.clear();
_shardResponses.clear();
_waitResponses.clear();
_shardEntries.clear();
_proxyLogsDBRequests.clear();
_proxyLogsDBResponses.clear();
_proxyReadRequests.clear();
_proxyReadRequestsIndices.clear();
_replicaInfo = _shared.getReplicas();
auto hasWork = _shared.writeQueuesWaiter.wait(_nextTimeout);
auto remainingPullBudget = _maxWorkItemsAtOnce;
if (unlikely(!hasWork && _shared.logsDBRequestQueue.isClosed())) {
// queue is closed, stop
stop();
return;
}
auto start = ternNow();
// the order here determins priority of processing when there is more work than we can handle at once
// we pull wait state responses without budget limit as they only do in memory modifications
_shared.writerShardRespQueue.pull(_waitResponses, WRITER_QUEUE_SIZE);
// we prioritize LogsDB requests and responses as they make us progress state
remainingPullBudget -= _shared.logsDBResponseQueue.pull(_logsDBResponses, remainingPullBudget);
remainingPullBudget -= _shared.logsDBRequestQueue.pull(_logsDBRequests, remainingPullBudget);
// then ProxyLogsDB requests/responses as they are needed to make progress on secondary locations
remainingPullBudget -= _shared.proxyLogsDBResponseQueue.pull(_proxyLogsDBResponses, remainingPullBudget);
remainingPullBudget -= _shared.proxyLogsDBRequestQueue.pull(_proxyLogsDBRequests, remainingPullBudget);
// then shard reponses as these are responses for requests we sent to primary location
remainingPullBudget -= _shared.writerProxyShardRespQueue.pull(_shardResponses, remainingPullBudget);
// last are new requests from local clients and then proxy locations
remainingPullBudget -= _shared.writerShardReqQueue.pull(_shardRequests, remainingPullBudget);
logsDBStep();
auto loopTime = ternNow() - start;
}
};
struct ShardReader : Loop {
private:
ShardShared& _shared;
SPSC<ShardReq>& _queue;
AES128Key _expandedShardKey;
AES128Key _expandedCDCKey;
ShardRespMsg _respContainer;
CdcToShardRespMsg _checkPointedrespContainer;
std::vector<ShardReq> _requests;
UDPSender _sender;
RandomGenerator _packetDropRand;
uint64_t _outgoingPacketDropProbability; // probability * 10,000
virtual void sendStop() override {
_queue.close();
}
public:
ShardReader(Logger& logger, std::shared_ptr<XmonAgent>& xmon, ShardShared& shared, SPSC<ShardReq>& queue) :
Loop(logger, xmon, "reader"),
_shared(shared),
_queue(queue),
_sender(UDPSenderConfig{.maxMsgSize = MAX_UDP_MTU}),
_packetDropRand(ternNow().ns),
_outgoingPacketDropProbability(0)
{
expandKey(ShardKey, _expandedShardKey);
expandKey(CDCKey, _expandedCDCKey);
auto convertProb = [this](const std::string& what, double prob, uint64_t& iprob) {
if (prob != 0.0) {
LOG_INFO(_env, "Will drop %s%% of %s packets", prob*100.0, what);
iprob = prob * 10'000.0;
ALWAYS_ASSERT(iprob > 0 && iprob < 10'000);
}
};
convertProb("outgoing", _shared.options.serverOptions.simulateOutgoingPacketDrop, _outgoingPacketDropProbability);
_requests.reserve(MAX_RECV_MSGS * 2);
}
virtual ~ShardReader() = default;
virtual void step() override {
_requests.clear();
uint32_t pulled = _queue.pull(_requests, MAX_RECV_MSGS * 2);
auto start = ternNow();
if (unlikely(_queue.isClosed())) {
// queue is closed, stop
stop();
return;
}
LOG_DEBUG(_env, "pulled %s requests from read queue", pulled);
for(auto& req : _requests) {
ALWAYS_ASSERT(readOnlyShardReq(req.msg.body.kind()));
bool dropArtificially = _packetDropRand.generate64() % 10'000 < _outgoingPacketDropProbability;
switch(req.protocol){
case SHARD_REQ_PROTOCOL_VERSION:
{
ShardRespMsg resp;
resp.id = req.msg.id;
_shared.shardDB.read(req.msg.body, resp.body);
packShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, req, resp);
break;
}
case CDC_TO_SHARD_REQ_PROTOCOL_VERSION:
{
CdcToShardRespMsg resp;
resp.id = req.msg.id;
resp.body.checkPointIdx = _shared.shardDB.read(req.msg.body, resp.body.resp);
packCheckPointedShardResponse(_env, _shared, _shared.sock().addr(), _sender, dropArtificially, req, resp, _expandedCDCKey);
break;
}
default:
LOG_ERROR(_env, "Unknown protocol version %s", req.protocol);
}
}
_sender.sendMessages(_env, _shared.sock());
}
};
struct ShardRegisterer : PeriodicLoop {
private:
ShardShared& _shared;
const ShardReplicaId _shrid;
const uint8_t _location;
const bool _noReplication;
const std::string _registryHost;
const uint16_t _registryPort;
XmonNCAlert _alert;
public:
ShardRegisterer(Logger& logger, std::shared_ptr<XmonAgent>& xmon, ShardShared& shared) :
PeriodicLoop(logger, xmon, "registerer", {1_sec, 1, 2_mins, 1}),
_shared(shared),
_shrid(_shared.options.shrid()),
_location(_shared.options.logsDBOptions.location),
_noReplication(_shared.options.logsDBOptions.noReplication),
_registryHost(_shared.options.registryClientOptions.host),
_registryPort(_shared.options.registryClientOptions.port)
{}
virtual ~ShardRegisterer() = default;
void init() {
_env.updateAlert(_alert, "Waiting to register ourselves for the first time");
}
virtual bool periodicStep() {
{
LOG_INFO(_env, "Registering ourselves (shard %s, location %s, %s) with registry", _shrid, (int)_location, _shared.sock().addr());
// ToDO: once leader election is fully enabled report or leader status instead of value of flag passed on startup
const auto [err, errStr] = registerShard(_registryHost, _registryPort, 10_sec, _shrid, _location, _shared.options.isLeader(), _shared.sock().addr());
if (err == EINTR) { return false; }
if (err) {
_env.updateAlert(_alert, "Couldn't register ourselves with registry: %s", errStr);
return false;
}
}
{
std::vector<FullShardInfo> allReplicas;
LOG_INFO(_env, "Fetching replicas for shardId %s from registry", _shrid.shardId());
const auto [err, errStr] = fetchShardReplicas(_registryHost, _registryPort, 10_sec, _shrid.shardId(), allReplicas);
if (err == EINTR) { return false; }
if (err) {
_env.updateAlert(_alert, "Failed getting shard replicas from registry: %s", errStr);
return false;
}
std::vector<FullShardInfo> leadersAtOtherLocations;
std::array<AddrsInfo, LogsDB::REPLICA_COUNT> localReplicas;
for (auto& replica : allReplicas) {
if (replica.id.shardId() != _shrid.shardId()) {
continue;
}
if (replica.locationId == _location) {
localReplicas[replica.id.replicaId().u8] = replica.addrs;
continue;
}
if (replica.isLeader && replica.addrs.addrs[0].port != 0) {
leadersAtOtherLocations.emplace_back(std::move(replica));
}
}
if (_shared.sock().addr() != localReplicas[_shrid.replicaId().u8]) {
_env.updateAlert(_alert, "AddrsInfo in registry: %s , not matching local AddrsInfo: %s", localReplicas[_shrid.replicaId().u8], _shared.sock().addr());
return false;
}
auto oldReplicas = _shared.getReplicas();
if (unlikely(!oldReplicas)) {
size_t emptyReplicas{0};
for (auto& replica : localReplicas) {
if (replica.addrs[0].port == 0) {
++emptyReplicas;
}
}
if (emptyReplicas > 0 && !_noReplication) {
_env.updateAlert(_alert, "Didn't get enough replicas with known addresses from registry");
return false;
}
}
if (unlikely(!oldReplicas || *oldReplicas != localReplicas)) {
LOG_DEBUG(_env, "Updating replicas to %s %s %s %s %s", localReplicas[0], localReplicas[1], localReplicas[2], localReplicas[3], localReplicas[4]);
_shared.setReplicas(std::move(localReplicas));
}
LOG_DEBUG(_env, "Updating leaders at other locations to %s", leadersAtOtherLocations);
_shared.setLeadersAtOtherLocations(std::move(leadersAtOtherLocations));
}
_shared.isInitiated.store(true, std::memory_order_release);
_env.clearAlert(_alert);
return true;
}
};
struct ShardBlockServiceUpdater : PeriodicLoop {
private:
ShardShared& _shared;
ShardReplicaId _shrid;
std::string _registryHost;
uint16_t _registryPort;
XmonNCAlert _alert;
std::vector<BlockServiceDeprecatedInfo> _blockServices;
std::vector<BlockServiceInfoShort> _currentBlockServices;
bool _updatedOnce;
public:
ShardBlockServiceUpdater(Logger& logger, std::shared_ptr<XmonAgent>& xmon, ShardShared& shared):
PeriodicLoop(logger, xmon, "bs_updater", {1_sec, shared.options.isLeader() ? 30_sec : 2_mins}),
_shared(shared),
_shrid(_shared.options.shrid()),
_registryHost(_shared.options.registryClientOptions.host),
_registryPort(_shared.options.registryClientOptions.port),
_updatedOnce(false)
{
_env.updateAlert(_alert, "Waiting to fetch block services for the first time");
}
virtual bool periodicStep() override {
if (!_blockServices.empty()) {
// We delayed applying cache update most likely we were leader. We should apply it now
_shared.blockServicesCache.updateCache(_blockServices, _currentBlockServices);
_blockServices.clear();
_currentBlockServices.clear();
}
LOG_INFO(_env, "about to fetch block services from %s:%s", _registryHost, _registryPort);
const auto [err, errStr] = fetchBlockServices(_registryHost, _registryPort, 10_sec, _shrid.shardId(), _blockServices, _currentBlockServices);
if (err == EINTR) { return false; }
if (err) {
_env.updateAlert(_alert, "could not reach registry: %s", errStr);
return false;
}
if (_blockServices.empty()) {
_env.updateAlert(_alert, "got no block services");
return false;
}
// We immediately update cache if we are leader and delay until next iteration on leader unless this is first update which we apply immediately
if (!_shared.options.isLeader() || !_updatedOnce) {
_updatedOnce = true;
_shared.blockServicesCache.updateCache(_blockServices, _currentBlockServices);
_blockServices.clear();
_currentBlockServices.clear();
_shared.isBlockServiceCacheInitiated.store(true, std::memory_order_release);
LOG_DEBUG(_env, "updated block services");
}
_env.clearAlert(_alert);
return true;
}
};
static void logsDBstatsToMetrics(struct MetricsBuilder& metricsBuilder, const LogsDBStats& stats, ShardReplicaId shrid, uint8_t location, TernTime now) {
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldU64( "idle_time", stats.idleTime.load(std::memory_order_relaxed).ns);
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldU64( "processing_time", stats.processingTime.load(std::memory_order_relaxed).ns);
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldU64( "leader_last_active", stats.leaderLastActive.load(std::memory_order_relaxed).ns);
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "append_window", stats.appendWindow.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "entries_released", stats.entriesReleased.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "follower_lag", stats.followerLag.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "reader_lag", stats.readerLag.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "catchup_window", stats.catchupWindow.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "entries_read", stats.entriesRead.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "requests_received", stats.requestsReceived.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "responses_received", stats.requestsReceived.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "requests_sent", stats.requestsSent.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "responses_sent", stats.responsesSent.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldFloat( "requests_timedout", stats.requestsTimedOut.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
{
metricsBuilder.measurement("eggsfs_shard_logsdb");
metricsBuilder.tag("shard", shrid);
metricsBuilder.tag("location", int(location));
metricsBuilder.tag("leader", stats.isLeader.load(std::memory_order_relaxed));
metricsBuilder.fieldU64( "current_epoch", stats.currentEpoch.load(std::memory_order_relaxed));
metricsBuilder.timestamp(now);
}
}
struct Metric {
std::string name;
std::atomic<uint64_t>* value;
};
struct ShardMetricsInserter : PeriodicLoop {
private:
InfluxDB _influxDB;
ShardShared& _shared;
ShardReplicaId _shrid;
uint8_t _location;
XmonNCAlert _sendMetricsAlert;
MetricsBuilder _metricsBuilder;
std::unordered_map<std::string, uint64_t> _rocksDBStats;
std::vector<Metric> _stats;
public:
ShardMetricsInserter(Logger& logger, std::shared_ptr<XmonAgent>& xmon, const InfluxDB& influxDB, ShardShared& shared):
PeriodicLoop(logger, xmon, "metrics", {1_sec, 1.0, 1_mins, 0.1}),
_influxDB(influxDB),
_shared(shared),
_shrid(_shared.options.shrid()),
_location(_shared.options.logsDBOptions.location),
_sendMetricsAlert(XmonAppType::DAYTIME, 5_mins)
{
_stats = {
// received counts
{"eggsfs_shard_logsdb_requests_received", &_shared.receivedLogsDBReqs},
{"eggsfs_shard_logsdb_responses_received", &_shared.receivedLogsDBResps},
{"eggsfs_shard_proxy_logsdb_requests_received", &_shared.receivedProxyLogsDBReqs},
{"eggsfs_shard_proxy_logsdb_responses_received",&_shared.receivedProxyLogsDBResps},
{"eggsfs_shard_write_requests_received", &_shared.receivedWriterReqs},
{"eggsfs_shard_write_proxy_responses_received", &_shared.receivedWriterProxyResps},
{"eggsfs_shard_read_requests_received", &_shared.receivedReaderReqs},
// dropped counts
{"eggsfs_shard_logsdb_requests_dropped", &_shared.droppedLogsDBReqs},
{"eggsfs_shard_logsdb_responses_dropped", &_shared.droppedLogsDBResps},
{"eggsfs_shard_proxy_logsdb_requests_dropped", &_shared.droppedProxyLogsDBReqs},
{"eggsfs_shard_proxy_logsdb_responses_dropped",&_shared.droppedProxyLogsDBResps},
{"eggsfs_shard_write_requests_dropped", &_shared.droppedWriteReqs},
{"eggsfs_shard_write_proxy_responses_droppedd", &_shared.droppedProxyWriteResps},
{"eggsfs_shard_read_requests_dropped", &_shared.droppedReadReqs},
};
}
virtual ~ShardMetricsInserter() = default;
virtual bool periodicStep() {
_shared.sharedDB.dumpRocksDBStatistics();
auto newMeassurement = [this](const std::string& name) {
_metricsBuilder.measurement(name);
_metricsBuilder.tag("shard", _shrid);
_metricsBuilder.tag("location", int(_location));
_metricsBuilder.tag("leader", _shared.logsDB.isLeader());
};
auto now = ternNow();
for (ShardMessageKind kind : allShardMessageKind) {
const ErrorCount& errs = _shared.errors[(int)kind];
for (int i = 0; i < errs.count.size(); i++) {
uint64_t count = errs.count[i].load();
if (count == 0) { continue; }
newMeassurement("eggsfs_shard_requests");
_metricsBuilder.tag("kind", kind);
_metricsBuilder.tag("write", !readOnlyShardReq(kind));
if (i == 0) {
_metricsBuilder.tag("error", "NO_ERROR");
} else {
_metricsBuilder.tag("error", (TernError)i);
}
_metricsBuilder.fieldU64("count", count);
_metricsBuilder.timestamp(now);
}
}
for (int i = 0; i < _shared.receivedRequests.size(); i++) {
_metricsBuilder.measurement("eggsfs_shard_received_requests");
_metricsBuilder.tag("shard", _shrid);
_metricsBuilder.tag("location", int(_location));
_metricsBuilder.tag("socket", i);
_metricsBuilder.fieldFloat("count", _shared.receivedRequests[i]);
_metricsBuilder.timestamp(now);
}
for (const auto& stat : _stats) {
newMeassurement(stat.name);
_metricsBuilder.fieldU64("count", *stat.value);
_metricsBuilder.timestamp(now);
}
{
_rocksDBStats.clear();
_shared.sharedDB.rocksDBMetrics(_rocksDBStats);
for (const auto& [name, value]: _rocksDBStats) {
newMeassurement("eggsfs_shard_rocksdb");
_metricsBuilder.fieldU64(name, value);
_metricsBuilder.timestamp(now);
}
}
logsDBstatsToMetrics(_metricsBuilder, _shared.logsDB.getStats(), _shrid, _location, now);
std::string err = sendMetrics(_influxDB, 10_sec, _metricsBuilder.payload());
_metricsBuilder.reset();
if (err.empty()) {
LOG_INFO(_env, "Sent metrics to influxdb");
_env.clearAlert(_sendMetricsAlert);
return true;
} else {
_env.updateAlert(_sendMetricsAlert, "Could not insert metrics: %s", err);
return false;
}
}
};
void runShard(ShardOptions& options) {
int logOutFd = STDOUT_FILENO;
if (!options.logOptions.logFile.empty()) {
logOutFd = open(options.logOptions.logFile.c_str(), O_WRONLY|O_CREAT|O_APPEND, 0644);
if (logOutFd < 0) {
throw SYSCALL_EXCEPTION("open");
}
}
Logger logger(options.logOptions.logLevel, logOutFd, options.logOptions.syslog, true);
std::shared_ptr<XmonAgent> xmon;
if (!options.xmonOptions.addr.empty()) {
xmon = std::make_shared<XmonAgent>();
}
Env env(logger, xmon, "startup");
{
LOG_INFO(env, "Running shard %s at location %s, with options:", options.shrid(), (int)options.logsDBOptions.location);
LOG_INFO(env, " level = %s", options.logOptions.logLevel);
LOG_INFO(env, " logFile = '%s'", options.logOptions.logFile);
LOG_INFO(env, " registryHost = '%s'", options.registryClientOptions.host);
LOG_INFO(env, " registryPort = %s", options.registryClientOptions.port);
LOG_INFO(env, " ownAddres = %s", options.serverOptions.addrs);
LOG_INFO(env, " simulateOutgoingPacketDrop = %s", options.serverOptions.simulateOutgoingPacketDrop);
LOG_INFO(env, " syslog = %s", (int)options.logOptions.syslog);
LOG_INFO(env, "Using LogsDB with options:");
LOG_INFO(env, " noReplication = '%s'", (int)options.logsDBOptions.noReplication);
LOG_INFO(env, " avoidBeingLeader = '%s'", (int)options.logsDBOptions.avoidBeingLeader);
}
// Immediately start xmon: we want the database initializing update to
// be there.
LoopThreads threads;
if (xmon) {
{
std::ostringstream ss;
ss << "eggsshard_" << std::setfill('0') << std::setw(3) << options.shrid().shardId() << "_" << options.shrid().replicaId();
options.xmonOptions.appInstance = ss.str();
}
options.xmonOptions.appType = XmonAppType::CRITICAL;
threads.emplace_back(LoopThread::Spawn(std::make_unique<Xmon>(logger, xmon, options.xmonOptions)));
}
// then everything else
XmonNCAlert dbInitAlert;
env.updateAlert(dbInitAlert, "initializing database");
SharedRocksDB sharedDB(logger, xmon, options.logsDBOptions.dbDir + "/db", options.logsDBOptions.dbDir + "/db-statistics.txt");
sharedDB.registerCFDescriptors(ShardDB::getColumnFamilyDescriptors());
sharedDB.registerCFDescriptors(LogsDB::getColumnFamilyDescriptors());
sharedDB.registerCFDescriptors(BlockServicesCacheDB::getColumnFamilyDescriptors());
rocksdb::Options rocksDBOptions;
rocksDBOptions.create_if_missing = true;
rocksDBOptions.create_missing_column_families = true;
rocksDBOptions.compression = rocksdb::kLZ4Compression;
rocksDBOptions.bottommost_compression = rocksdb::kZSTD;
// 1000*256 = 256k open files at once, given that we currently run on a
// single machine this is appropriate.
rocksDBOptions.max_open_files = 1000;
// We batch writes and flush manually.
rocksDBOptions.manual_wal_flush = true;
sharedDB.open(rocksDBOptions);
BlockServicesCacheDB blockServicesCache(logger, xmon, sharedDB);
ShardDB shardDB(logger, xmon, options.shardId, options.logsDBOptions.location, options.transientDeadlineInterval, sharedDB, blockServicesCache);
LogsDB logsDB(logger, xmon, sharedDB, options.logsDBOptions.replicaId, shardDB.lastAppliedLogEntry(), options.logsDBOptions.noReplication, options.logsDBOptions.avoidBeingLeader);
env.clearAlert(dbInitAlert);
ShardShared shared(options, sharedDB, blockServicesCache, shardDB, logsDB, UDPSocketPair(env, options.serverOptions.addrs));
threads.emplace_back(LoopThread::Spawn(std::make_unique<ShardServer>(logger, xmon, shared)));
for (uint16_t i = 0; i < options.numReaders; ++i) {
threads.emplace_back(LoopThread::Spawn(std::make_unique<ShardReader>(logger, xmon, shared, *shared.readerRequestsQueues[i])));
}
threads.emplace_back(LoopThread::Spawn(std::make_unique<ShardWriter>(logger, xmon, shared)));
threads.emplace_back(LoopThread::Spawn(std::make_unique<ShardRegisterer>(logger, xmon, shared)));
threads.emplace_back(LoopThread::Spawn(std::make_unique<ShardBlockServiceUpdater>(logger, xmon, shared)));
if (!options.metricsOptions.origin.empty()) {
threads.emplace_back(LoopThread::Spawn(std::make_unique<ShardMetricsInserter>(logger, xmon, options.metricsOptions, shared)));
}
// from this point on termination on SIGINT/SIGTERM will be graceful
LoopThread::waitUntilStopped(threads);
threads.clear();
shardDB.close();
logsDB.close();
sharedDB.close();
LOG_INFO(env, "shard terminating gracefully, bye.");
}