Hawkin
/
arangodb
mirror of https://gitee.com/bigwinds/arangodb
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
arangodb/arangod/Aql/QuerySnippet.cpp

433 lines
18 KiB
C++
Raw Blame History

////////////////////////////////////////////////////////////////////////////////
/// DISCLAIMER
///
/// Copyright 2019-2019 ArangoDB GmbH, Cologne, Germany
///
/// 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.
///
/// Copyright holder is ArangoDB GmbH, Cologne, Germany
///
/// @author Michael Hackstein
////////////////////////////////////////////////////////////////////////////////
#include "QuerySnippet.h"
#include "Aql/ClusterNodes.h"
#include "Aql/Collection.h"
#include "Aql/CollectionAccessingNode.h"
#include "Aql/DistributeConsumerNode.h"
#include "Aql/ExecutionNode.h"
#include "Aql/ExecutionPlan.h"
#include "Aql/GraphNode.h"
#include "Aql/IResearchViewNode.h"
#include "Aql/ShardLocking.h"
#include "Basics/StringUtils.h"
#include "Cluster/ServerState.h"
#include <velocypack/Builder.h>
#include <velocypack/velocypack-aliases.h>
using namespace arangodb;
using namespace arangodb::aql;
using namespace arangodb::basics;
void QuerySnippet::addNode(ExecutionNode* node) {
TRI_ASSERT(node != nullptr);
_nodes.push_back(node);
switch (node->getType()) {
case ExecutionNode::ENUMERATE_COLLECTION:
case ExecutionNode::INDEX:
case ExecutionNode::INSERT:
case ExecutionNode::UPDATE:
case ExecutionNode::REMOVE:
case ExecutionNode::REPLACE:
case ExecutionNode::UPSERT: {
// We do not actually need to know the details here.
// We just wanna know the shards!
auto collectionAccessingNode = dynamic_cast<CollectionAccessingNode*>(node);
TRI_ASSERT(collectionAccessingNode != nullptr);
_expansions.emplace_back(node, !collectionAccessingNode->isUsedAsSatellite(),
collectionAccessingNode->isUsedAsSatellite());
break;
}
case ExecutionNode::TRAVERSAL:
case ExecutionNode::SHORTEST_PATH:
case ExecutionNode::K_SHORTEST_PATHS: {
_expansions.emplace_back(node, false, false);
break;
}
case ExecutionNode::ENUMERATE_IRESEARCH_VIEW: {
auto viewNode = ExecutionNode::castTo<iresearch::IResearchViewNode*>(node);
// evaluate node volatility before the distribution
// can't do it on DB servers since only parts of the plan will be sent
viewNode->volatility(true);
_expansions.emplace_back(node, false, false);
break;
}
case ExecutionNode::MATERIALIZE: {
auto collectionAccessingNode = dynamic_cast<CollectionAccessingNode*>(node);
// Materialize index node - true
// Materialize view node - false
if (collectionAccessingNode != nullptr) {
_expansions.emplace_back(node, true, false);
}
break;
}
default:
// do nothing
break;
}
}
void QuerySnippet::serializeIntoBuilder(ServerID const& server, ShardLocking& shardLocking,
std::unordered_map<size_t, size_t>& nodeAliases,
VPackBuilder& infoBuilder) {
TRI_ASSERT(!_nodes.empty());
TRI_ASSERT(!_expansions.empty());
auto firstBranchRes = prepareFirstBranch(server, shardLocking);
if (!firstBranchRes.ok()) {
// We have at least one expansion that has no shard on this server.
// we do not need to write this snippet here.
TRI_ASSERT(firstBranchRes.is(TRI_ERROR_CLUSTER_NOT_LEADER));
return;
}
std::unordered_map<ExecutionNode*, std::set<ShardID>>& localExpansions =
firstBranchRes.get();
// We clone every Node* and maintain a list of ReportingGroups for profiler
ExecutionNode* lastNode = _nodes.back();
bool lastIsRemote = lastNode->getType() == ExecutionNode::REMOTE;
// Query can only ed with a REMOTE or SINGLETON
TRI_ASSERT(lastIsRemote || lastNode->getType() == ExecutionNode::SINGLETON);
// Singleton => noDependency
TRI_ASSERT(lastNode->getType() != ExecutionNode::SINGLETON || !lastNode->hasDependency());
if (lastIsRemote) {
auto rem = ExecutionNode::castTo<RemoteNode*>(lastNode);
if (!_madeResponsibleForShutdown) {
// Enough to do this step once
// We need to connect this Node to the sink
// update the remote node with the information about the query
rem->server("server:" + arangodb::ServerState::instance()->getId());
rem->queryId(_inputSnippet);
// A Remote can only contact a global SCATTER or GATHER node.
TRI_ASSERT(rem->getFirstDependency() != nullptr);
TRI_ASSERT(rem->getFirstDependency()->getType() == ExecutionNode::SCATTER ||
rem->getFirstDependency()->getType() == ExecutionNode::DISTRIBUTE);
TRI_ASSERT(rem->hasDependency());
// Need to wire up the internal scatter and distribute consumer in between the last two nodes
// Note that we need to clean up this after we produced the snippets.
_globalScatter = ExecutionNode::castTo<ScatterNode*>(rem->getFirstDependency());
// Let the globalScatter node distribute data by server
_globalScatter->setScatterType(ScatterNode::ScatterType::SERVER);
// All of the above could be repeated as it is constant information.
// this one line is the important one
rem->isResponsibleForInitializeCursor(true);
_madeResponsibleForShutdown = true;
} else {
rem->isResponsibleForInitializeCursor(false);
}
// We have it all serverbased now
rem->setDistributeId(server);
// Wire up this server to the global scatter
TRI_ASSERT(_globalScatter != nullptr);
_globalScatter->addClient(rem);
// For serialization remove the dependency of Remote
#ifdef ARANGODB_ENABLE_MAINTAINER_MODE
std::vector<ExecutionNode*> deps;
lastNode->dependencies(deps);
TRI_ASSERT(deps.size() == 1);
TRI_ASSERT(deps[0] == _globalScatter);
#endif
lastNode->removeDependencies();
}
// The Key is required to build up the queryId mapping later
infoBuilder.add(VPackValue(
arangodb::basics::StringUtils::itoa(_idOfSinkRemoteNode) + ":" + server));
if (!localExpansions.empty()) {
// We have Expansions to permutate, guaranteed they have
// all identical lengths.
size_t numberOfShardsToPermutate = localExpansions.begin()->second.size();
std::vector<std::string> distIds{};
// Reserve the amount of localExpansions,
distIds.reserve(numberOfShardsToPermutate);
// Create an internal GatherNode, that will connect to all execution
// steams of the query
auto plan = lastNode->plan();
TRI_ASSERT(plan == _sinkNode->plan());
// Clone the sink node, we do not need dependencies (second bool)
// And we do not need variables
GatherNode* internalGather =
ExecutionNode::castTo<GatherNode*>(_sinkNode->clone(plan, false, false));
// Use the same elements for sorting
internalGather->elements(_sinkNode->elements());
// We need to modify the registerPlanning.
// The internalGather is NOT allowed to reduce the number of registers,
// it needs to expose it's input register by all means
internalGather->setVarsUsedLater(_nodes.front()->getVarsUsedLater());
internalGather->setRegsToClear({});
nodeAliases.try_emplace(internalGather->id(), std::numeric_limits<size_t>::max());
ScatterNode* internalScatter = nullptr;
if (lastIsRemote) {
TRI_ASSERT(_globalScatter != nullptr);
TRI_ASSERT(plan == _globalScatter->plan());
internalScatter =
ExecutionNode::castTo<ScatterNode*>(_globalScatter->clone(plan, false, false));
internalScatter->clearClients();
internalScatter->addDependency(lastNode);
// Let the local Scatter node distribute data by SHARD
internalScatter->setScatterType(ScatterNode::ScatterType::SHARD);
nodeAliases.try_emplace(internalScatter->id(), std::numeric_limits<size_t>::max());
if (_globalScatter->getType() == ExecutionNode::DISTRIBUTE) {
{
// We are not allowed to generate new keys on the DBServer.
// We need to take the ones generated by the node above.
DistributeNode* internalDist =
ExecutionNode::castTo<DistributeNode*>(internalScatter);
internalDist->setCreateKeys(false);
}
DistributeNode const* dist =
ExecutionNode::castTo<DistributeNode const*>(_globalScatter);
TRI_ASSERT(dist != nullptr);
auto distCollection = dist->collection();
TRI_ASSERT(distCollection != nullptr);
// Now find the node that provides the distribute information
for (auto const& exp : localExpansions) {
auto colAcc = dynamic_cast<CollectionAccessingNode*>(exp.first);
TRI_ASSERT(colAcc != nullptr);
if (colAcc->collection() == distCollection) {
// Found one, use all shards of it
for (auto const& s : exp.second) {
distIds.emplace_back(s);
}
break;
}
}
} else {
// In this case we actually do not care for the real value, we just need
// to ensure that every client get's exactly one copy.
for (size_t i = 0; i < numberOfShardsToPermutate; i++) {
distIds.emplace_back(StringUtils::itoa(i));
}
}
DistributeConsumerNode* consumer =
createConsumerNode(plan, internalScatter, distIds[0]);
nodeAliases.try_emplace(consumer->id(), std::numeric_limits<size_t>::max());
// now wire up the temporary nodes
TRI_ASSERT(_nodes.size() > 1);
ExecutionNode* secondToLast = _nodes[_nodes.size() - 2];
TRI_ASSERT(secondToLast->hasDependency());
secondToLast->swapFirstDependency(consumer);
}
// We do not need to copy the first stream, we can use the one we have.
// We only need copies for the other streams.
internalGather->addDependency(_nodes.front());
// NOTE: We will copy over the entire snippet stream here.
// We will inject the permuted shards on the way.
// Also note: the local plan will take memory responsibility
// of the ExecutionNodes created during this procedure.
for (size_t i = 1; i < numberOfShardsToPermutate; ++i) {
ExecutionNode* previous = nullptr;
for (auto enIt = _nodes.rbegin(), end = _nodes.rend(); enIt != end; ++enIt) {
ExecutionNode* current = *enIt;
if (lastIsRemote && current == lastNode) {
// Do never clone the remote, link following node
// to Consumer and Scatter instead
TRI_ASSERT(internalScatter != nullptr);
DistributeConsumerNode* consumer =
createConsumerNode(plan, internalScatter, distIds[i]);
consumer->isResponsibleForInitializeCursor(false);
nodeAliases.try_emplace(consumer->id(), std::numeric_limits<size_t>::max());
previous = consumer;
continue;
}
ExecutionNode* clone = current->clone(plan, false, false);
auto permuter = localExpansions.find(current);
if (permuter != localExpansions.end()) {
auto collectionAccessingNode = dynamic_cast<CollectionAccessingNode*>(clone);
TRI_ASSERT(collectionAccessingNode != nullptr);
// Get the `i` th shard
collectionAccessingNode->setUsedShard(*std::next(permuter->second.begin(), i));
}
if (previous != nullptr) {
clone->addDependency(previous);
}
TRI_ASSERT(clone->id() != current->id());
nodeAliases.try_emplace(clone->id(), current->id());
previous = clone;
}
TRI_ASSERT(previous != nullptr);
// Previous is now the last node, where our internal GATHER needs to be connected to
internalGather->addDependency(previous);
}
const unsigned flags = ExecutionNode::SERIALIZE_DETAILS;
internalGather->toVelocyPack(infoBuilder, flags, false);
// We need to clean up ONLY if we have injected the local scatter
if (lastIsRemote) {
TRI_ASSERT(internalScatter != nullptr);
TRI_ASSERT(_nodes.size() > 1);
ExecutionNode* secondToLast = _nodes[_nodes.size() - 2];
TRI_ASSERT(secondToLast->hasDependency());
secondToLast->swapFirstDependency(lastNode);
}
} else {
const unsigned flags = ExecutionNode::SERIALIZE_DETAILS;
_nodes.front()->toVelocyPack(infoBuilder, flags, false);
}
if (lastIsRemote) {
// For the local copy read the dependency of Remote
lastNode->addDependency(_globalScatter);
}
}
ResultT<std::unordered_map<ExecutionNode*, std::set<ShardID>>> QuerySnippet::prepareFirstBranch(
ServerID const& server, ShardLocking& shardLocking) {
size_t numberOfShardsToPermutate = 0;
std::unordered_map<ExecutionNode*, std::set<ShardID>> localExpansions;
std::unordered_map<ShardID, ServerID> const& shardMapping =
shardLocking.getShardMapping();
for (auto const& exp : _expansions) {
if (exp.node->getType() == ExecutionNode::ENUMERATE_IRESEARCH_VIEW) {
// Special case, VIEWs can serve more than 1 shard per Node.
// We need to inject them all at once.
auto* viewNode = ExecutionNode::castTo<iresearch::IResearchViewNode*>(exp.node);
auto& viewShardList = viewNode->shards();
viewShardList.clear();
auto collections = viewNode->collections();
for (aql::Collection const& c : collections) {
auto const& shards = shardLocking.shardsForSnippet(id(), &c);
for (auto const& s : shards) {
auto check = shardMapping.find(s);
// If we find a shard here that is not in this mapping,
// we have 1) a problem with locking before that should have thrown
// 2) a problem with shardMapping lookup that should have thrown before
TRI_ASSERT(check != shardMapping.end());
if (check->second == server) {
viewShardList.emplace_back(s);
}
}
}
continue;
} else if ( exp.node->getType() == ExecutionNode::TRAVERSAL
|| exp.node->getType() == ExecutionNode::SHORTEST_PATH
|| exp.node->getType() == ExecutionNode::K_SHORTEST_PATHS) {
// the same translation is copied to all servers
// there are no local expansions
GraphNode* graphNode = ExecutionNode::castTo<GraphNode*>(exp.node);
graphNode->setCollectionToShard({}); //clear previous information
for(auto* aqlCollection : graphNode->collections()) {
auto const& shards = aqlCollection->shardIds();
TRI_ASSERT(!shards->empty());
for (std::string const& shard : *shards) {
auto found = shardMapping.find(shard);
if (found != shardMapping.end() && found->second == server) {
// provide a correct translation from collection to shard
// to be used in toVelocyPack methods of classes derived
// from GraphNode
graphNode->addCollectionToShard(aqlCollection->name(), shard);
}
}
}
continue; // skip rest - there are no local expansions
}
// It is of utmost importance that this is an ordered set of Shards.
// We can only join identical indexes of shards for each collection
// locally.
std::set<ShardID> myExp;
auto modNode = dynamic_cast<CollectionAccessingNode const*>(exp.node);
// Only accessing nodes can endup here.
TRI_ASSERT(modNode != nullptr);
auto col = modNode->collection();
// Should be hit earlier, a modification node here is required to have a collection
TRI_ASSERT(col != nullptr);
auto const& shards = shardLocking.shardsForSnippet(id(), col);
for (auto const& s : shards) {
auto check = shardMapping.find(s);
// If we find a shard here that is not in this mapping,
// we have 1) a problem with locking before that should have thrown
// 2) a problem with shardMapping lookup that should have thrown before
TRI_ASSERT(check != shardMapping.end());
if (check->second == server || exp.isSatellite) {
// add all shards on satellites.
// and all shards where this server is the leader
myExp.emplace(s);
}
}
if (myExp.empty()) {
return {TRI_ERROR_CLUSTER_NOT_LEADER};
}
// For all other Nodes we can inject a single shard at a time.
// Always use the list of nodes we maintain to hold the first
// of all shards.
// We later use a cone mechanism to inject other shards of permutation
auto collectionAccessingNode = dynamic_cast<CollectionAccessingNode*>(exp.node);
TRI_ASSERT(collectionAccessingNode != nullptr);
collectionAccessingNode->setUsedShard(*myExp.begin());
if (exp.doExpand) {
TRI_ASSERT(!collectionAccessingNode->isUsedAsSatellite());
// All parts need to have exact same size, they need to be permutated pairwise!
TRI_ASSERT(numberOfShardsToPermutate == 0 || myExp.size() == numberOfShardsToPermutate);
// set the max loop index (note this will essentially be done only once)
numberOfShardsToPermutate = myExp.size();
if (numberOfShardsToPermutate > 1) {
// Only in this case we really need to do an expansion
// Otherwise we get away with only using the main stream for
// this server
// NOTE: This might differ between servers.
// One server might require an expansion (many shards) while another does not (only one shard).
localExpansions.emplace(exp.node, std::move(myExp));
}
} else {
TRI_ASSERT(myExp.size() == 1);
}
} // for _expansions - end;
return {localExpansions};
}
DistributeConsumerNode* QuerySnippet::createConsumerNode(ExecutionPlan* plan,
ScatterNode* internalScatter,
std::string const& distributeId) {
auto uniq_consumer =
std::make_unique<DistributeConsumerNode>(plan, plan->nextId(), distributeId);
auto consumer = uniq_consumer.get();
TRI_ASSERT(consumer != nullptr);
// Hand over responsibility to plan, s.t. it can clean up if one of the below fails
plan->registerNode(uniq_consumer.release());
consumer->addDependency(internalScatter);
consumer->cloneRegisterPlan(internalScatter);
internalScatter->addClient(consumer);
return consumer;
}
Reference in New Issue View Git Blame Copy Permalink