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

1083 lines
40 KiB
C++
Raw Blame History

////////////////////////////////////////////////////////////////////////////////
/// DISCLAIMER
///
/// Copyright 2014-2016 ArangoDB GmbH, Cologne, Germany
/// Copyright 2004-2014 triAGENS 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 Dr. Frank Celler
////////////////////////////////////////////////////////////////////////////////
#include "MMFilesCompactionFeature.h"
#include "MMFilesCompactorThread.h"
#include "Basics/ConditionLocker.h"
#include "Basics/ReadLocker.h"
#include "Basics/WriteLocker.h"
#include "Basics/conversions.h"
#include "Basics/files.h"
#include "Basics/FileUtils.h"
#include "Basics/memory-map.h"
#include "Logger/Logger.h"
#include "MMFiles/MMFilesCollection.h"
#include "MMFiles/MMFilesCompactionLocker.h"
#include "MMFiles/MMFilesDatafileHelper.h"
#include "MMFiles/MMFilesDatafileStatisticsContainer.h"
#include "MMFiles/MMFilesDocumentPosition.h"
#include "MMFiles/MMFilesEngine.h"
#include "MMFiles/MMFilesIndexElement.h"
#include "MMFiles/MMFilesPrimaryIndex.h"
#include "StorageEngine/EngineSelectorFeature.h"
#include "Utils/SingleCollectionTransaction.h"
#include "Transaction/Helpers.h"
#include "Transaction/Hints.h"
#include "Transaction/StandaloneContext.h"
#include "VocBase/LogicalCollection.h"
#include "VocBase/vocbase.h"
using namespace arangodb;
static char const* ReasonCorrupted =
"skipped compaction because collection has corrupted datafile(s)";
static char const* ReasonNoDatafiles =
"skipped compaction because collection has no datafiles";
static char const* ReasonCompactionBlocked =
"skipped compaction because existing compactor file is in the way and "
"waits to be processed";
static char const* ReasonDatafileSmall =
"compacting datafile because it's small and will be merged with next";
static char const* ReasonEmpty =
"compacting datafile because collection is empty";
static char const* ReasonOnlyDeletions =
"compacting datafile because it contains only deletion markers";
static char const* ReasonDeadSize =
"compacting datafile because it contains much dead object space";
static char const* ReasonDeadSizeShare =
"compacting datafile because it contains high share of dead objects";
static char const* ReasonDeadCount =
"compacting datafile because it contains many dead objects";
static char const* ReasonNothingToCompact =
"checked datafiles, but no compaction opportunity found";
/// @brief compaction state
namespace arangodb {
struct CompactionContext {
transaction::Methods* _trx;
LogicalCollection* _collection;
MMFilesDatafile* _compactor;
MMFilesDatafileStatisticsContainer _dfi;
bool _keepDeletions;
CompactionContext(CompactionContext const&) = delete;
CompactionContext() : _trx(nullptr), _collection(nullptr), _compactor(nullptr), _dfi(), _keepDeletions(true) {}
};
}
/// @brief callback to drop a datafile
void MMFilesCompactorThread::DropDatafileCallback(MMFilesDatafile* df, LogicalCollection* collection) {
auto physical = static_cast<MMFilesCollection*>(collection->getPhysical());
TRI_ASSERT(physical != nullptr);
TRI_ASSERT(df != nullptr);
std::unique_ptr<MMFilesDatafile> datafile(df);
TRI_voc_fid_t fid = datafile->fid();
std::string copy;
std::string name("deleted-" + std::to_string(fid) + ".db");
std::string filename = arangodb::basics::FileUtils::buildFilename(physical->path(), name);
if (datafile->isPhysical()) {
// copy the current filename
copy = datafile->getName();
int res = datafile->rename(filename);
if (res != TRI_ERROR_NO_ERROR) {
LOG_TOPIC(ERR, Logger::COMPACTOR)
<< "cannot rename obsolete datafile '" << copy << "' to '"
<< filename << "': " << TRI_errno_string(res);
} else {
LOG_TOPIC(DEBUG, Logger::COMPACTOR)
<< "renamed obsolete datafile '" << copy << "' to '"
<< filename << "': " << TRI_errno_string(res);
}
}
LOG_TOPIC(DEBUG, Logger::COMPACTOR)
<< "finished compacting datafile '" << datafile->getName() << "'";
int res = datafile->close();
if (res != TRI_ERROR_NO_ERROR) {
LOG_TOPIC(ERR, Logger::COMPACTOR)
<< "cannot close obsolete datafile '" << datafile->getName()
<< "': " << TRI_errno_string(res);
} else if (datafile->isPhysical()) {
LOG_TOPIC(DEBUG, Logger::COMPACTOR)
<< "wiping compacted datafile '" << datafile->getName()
<< "' from disk";
res = TRI_UnlinkFile(filename.c_str());
if (res != TRI_ERROR_NO_ERROR) {
LOG_TOPIC(ERR, Logger::COMPACTOR)
<< "cannot wipe obsolete datafile '" << datafile->getName()
<< "': " << TRI_errno_string(res);
}
// check for .dead files
if (!copy.empty()) {
// remove .dead file for datafile
std::string deadfile = copy + ".dead";
// check if .dead file exists, then remove it
if (TRI_ExistsFile(deadfile.c_str())) {
TRI_UnlinkFile(deadfile.c_str());
}
}
}
}
////////////////////////////////////////////////////////////////////////////////
/// @brief callback to rename a datafile
///
/// The datafile will be renamed to "temp-abc.db" (where "abc" is the fid of
/// the datafile) first. If this rename operation fails, there will be a
/// compactor file and a datafile. On startup, the datafile will be preferred
/// in this case.
/// If renaming succeeds, the compactor will be named to the original datafile.
/// If that does not succeed, there is a compactor file and a renamed datafile.
/// On startup, the compactor file will be used, and the renamed datafile
/// will be treated as a temporary file and dropped.
////////////////////////////////////////////////////////////////////////////////
void MMFilesCompactorThread::RenameDatafileCallback(MMFilesDatafile* datafile,
MMFilesDatafile* compactor,
LogicalCollection* collection) {
TRI_ASSERT(datafile != nullptr);
TRI_ASSERT(compactor != nullptr);
TRI_ASSERT(collection != nullptr);
auto physical = static_cast<MMFilesCollection*>(collection->getPhysical());
TRI_ASSERT(physical != nullptr);
std::string compactorName = compactor->getName();
bool ok = false;
TRI_ASSERT(datafile->fid() == compactor->fid());
if (datafile->isPhysical()) {
// construct a suitable tempname
std::string jname("temp-" + std::to_string(datafile->fid()) + ".db");
std::string tempFilename = arangodb::basics::FileUtils::buildFilename(physical->path(), jname);
std::string realName = datafile->getName();
int res = datafile->rename(tempFilename);
if (res != TRI_ERROR_NO_ERROR) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "unable to rename datafile '" << datafile->getName() << "' to '" << tempFilename << "': " << TRI_errno_string(res);
} else {
LOG_TOPIC(DEBUG, arangodb::Logger::COMPACTOR)
<< "renamed datafile from '" << realName << "' to '"
<< tempFilename << "'";
res = compactor->rename(realName);
if (res != TRI_ERROR_NO_ERROR) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "unable to rename compaction file '" << compactor->getName() << "' to '" << realName << "': " << TRI_errno_string(res);
} else {
LOG_TOPIC(DEBUG, arangodb::Logger::COMPACTOR)
<< "renamed datafile from '" << compactorName << "' to '"
<< tempFilename << "'";
}
}
ok = (res == TRI_ERROR_NO_ERROR);
} else {
ok = true;
}
if (ok) {
int res = static_cast<MMFilesCollection*>(collection->getPhysical())->replaceDatafileWithCompactor(datafile, compactor);
if (res != TRI_ERROR_NO_ERROR) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "logic error: could not swap datafile and compactor files";
return;
}
DropDatafileCallback(datafile, collection);
}
}
/// @brief remove an empty compactor file
int MMFilesCompactorThread::removeCompactor(LogicalCollection* collection,
MMFilesDatafile* compactor) {
LOG_TOPIC(DEBUG, Logger::COMPACTOR) << "removing empty compaction file '" << compactor->getName() << "'";
// remove the compactor from the list of compactors
bool ok = static_cast<MMFilesCollection*>(collection->getPhysical())->removeCompactor(compactor);
if (!ok) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "logic error: could not locate compactor";
return TRI_ERROR_INTERNAL;
}
// close the file & remove it
if (compactor->isPhysical()) {
std::string filename = compactor->getName();
delete compactor;
TRI_UnlinkFile(filename.c_str());
} else {
delete compactor;
}
return TRI_ERROR_NO_ERROR;
}
/// @brief remove an empty datafile
int MMFilesCompactorThread::removeDatafile(LogicalCollection* collection,
MMFilesDatafile* df) {
LOG_TOPIC(DEBUG, Logger::COMPACTOR) << "removing empty datafile '" << df->getName() << "'";
bool ok = static_cast<MMFilesCollection*>(collection->getPhysical())->removeDatafile(df);
if (!ok) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "logic error: could not locate datafile";
return TRI_ERROR_INTERNAL;
}
// update dfi
static_cast<MMFilesCollection*>(collection->getPhysical())->_datafileStatistics.remove(df->fid());
return TRI_ERROR_NO_ERROR;
}
/// @brief calculate the target size for the compactor to be created
MMFilesCompactorThread::CompactionInitialContext MMFilesCompactorThread::getCompactionContext(
transaction::Methods* trx, LogicalCollection* collection,
std::vector<CompactionInfo> const& toCompact) {
CompactionInitialContext context(trx, collection);
// this is the minimum required size
context._targetSize =
sizeof(MMFilesDatafileHeaderMarker) + sizeof(MMFilesCollectionHeaderMarker) +
sizeof(MMFilesDatafileFooterMarker) + 256; // allow for some overhead
size_t const n = toCompact.size();
for (size_t i = 0; i < n; ++i) {
auto compaction = toCompact[i];
MMFilesDatafile* df = compaction._datafile;
// We will sequentially scan the logfile for collection:
if (df->isPhysical()) {
df->sequentialAccess();
df->willNeed();
}
if (i == 0) {
// extract and store fid
context._fid = compaction._datafile->fid();
}
context._keepDeletions = compaction._keepDeletions;
/// @brief datafile iterator, calculates necessary total size
auto calculateSize = [&context](MMFilesMarker const* marker, MMFilesDatafile* datafile) -> bool {
LogicalCollection* collection = context._collection;
TRI_ASSERT(collection != nullptr);
auto physical = static_cast<MMFilesCollection*>(collection->getPhysical());
TRI_ASSERT(physical != nullptr);
MMFilesMarkerType const type = marker->getType();
// new or updated document
if (type == TRI_DF_MARKER_VPACK_DOCUMENT) {
VPackSlice const slice(reinterpret_cast<char const*>(marker) + MMFilesDatafileHelper::VPackOffset(type));
TRI_ASSERT(slice.isObject());
VPackSlice keySlice = transaction::helpers::extractKeyFromDocument(slice);
// check if the document is still active
auto primaryIndex = physical->primaryIndex();
MMFilesMarker const* markerPtr = nullptr;
MMFilesSimpleIndexElement element = primaryIndex->lookupKey(context._trx, keySlice);
if (element) {
MMFilesDocumentPosition const old =
physical->lookupDocument(element.localDocumentId());
markerPtr = reinterpret_cast<MMFilesMarker const*>(
static_cast<uint8_t const*>(old.dataptr()) -
MMFilesDatafileHelper::VPackOffset(TRI_DF_MARKER_VPACK_DOCUMENT));
}
bool deleted = (markerPtr == nullptr || marker != markerPtr);
if (deleted) {
return true;
}
context._keepDeletions = true;
context._targetSize += MMFilesDatafileHelper::AlignedMarkerSize<int64_t>(marker);
}
// deletions
else if (type == TRI_DF_MARKER_VPACK_REMOVE) {
if (context._keepDeletions) {
context._targetSize += MMFilesDatafileHelper::AlignedMarkerSize<int64_t>(marker);
}
}
return true;
};
bool ok;
{
auto physical = static_cast<MMFilesCollection*>(context._collection->getPhysical());
TRI_ASSERT(physical != nullptr);
bool const useDeadlockDetector = false;
int res = physical->lockRead(useDeadlockDetector, trx->state(), 86400.0);
if (res != TRI_ERROR_NO_ERROR) {
ok = false;
} else {
// got read lock
try {
ok = TRI_IterateDatafile(df, calculateSize);
} catch (...) {
ok = false;
}
physical->unlockRead(useDeadlockDetector, trx->state());
}
}
if (df->isPhysical()) {
df->randomAccess();
}
if (!ok) {
context._failed = true;
break;
}
}
return context;
}
/// @brief compact the specified datafiles
void MMFilesCompactorThread::compactDatafiles(LogicalCollection* collection,
std::vector<CompactionInfo> const& toCompact) {
TRI_ASSERT(collection != nullptr);
auto physical = static_cast<MMFilesCollection*>(collection->getPhysical());
TRI_ASSERT(physical != nullptr);
size_t const n = toCompact.size();
TRI_ASSERT(n > 0);
auto context = std::make_unique<CompactionContext>();
/// @brief datafile iterator, copies "live" data from datafile into compactor
/// this function is called for all markers in the collected datafiles. Its
/// purpose is to find the still-alive markers and copy them into the compactor
/// file.
/// IMPORTANT: if the logic inside this function is adjusted, the total size
/// calculated by function CalculateSize might need adjustment, too!!
auto compactifier = [&context, &physical, this](MMFilesMarker const* marker, MMFilesDatafile* datafile) -> bool {
TRI_voc_fid_t const targetFid = context->_compactor->fid();
MMFilesMarkerType const type = marker->getType();
// new or updated document
if (type == TRI_DF_MARKER_VPACK_DOCUMENT) {
VPackSlice const slice(reinterpret_cast<char const*>(marker) + MMFilesDatafileHelper::VPackOffset(type));
TRI_ASSERT(slice.isObject());
VPackSlice keySlice = transaction::helpers::extractKeyFromDocument(slice);
// check if the document is still active
auto primaryIndex = physical->primaryIndex();
MMFilesMarker const* markerPtr = nullptr;
MMFilesSimpleIndexElement element = primaryIndex->lookupKey(context->_trx, keySlice);
if (element) {
MMFilesDocumentPosition const old = physical->lookupDocument(element.localDocumentId());
markerPtr = reinterpret_cast<MMFilesMarker const*>(static_cast<uint8_t const*>(old.dataptr()) - MMFilesDatafileHelper::VPackOffset(TRI_DF_MARKER_VPACK_DOCUMENT));
}
bool deleted = (markerPtr == nullptr || marker != markerPtr);
if (deleted) {
// found a dead document
return true;
}
context->_keepDeletions = true;
// write to compactor files
MMFilesMarker* result;
int res = copyMarker(context->_compactor, marker, &result);
if (res != TRI_ERROR_NO_ERROR) {
THROW_ARANGO_EXCEPTION_MESSAGE(res, std::string("cannot write document marker into compactor file: ") + TRI_errno_string(res));
}
// let marker point to the new position
uint8_t const* dataptr = reinterpret_cast<uint8_t const*>(result) + MMFilesDatafileHelper::VPackOffset(TRI_DF_MARKER_VPACK_DOCUMENT);
physical->updateLocalDocumentId(element.localDocumentId(), dataptr, targetFid, false);
context->_dfi.numberAlive++;
context->_dfi.sizeAlive += MMFilesDatafileHelper::AlignedMarkerSize<int64_t>(marker);
}
// deletions
else if (type == TRI_DF_MARKER_VPACK_REMOVE) {
if (context->_keepDeletions) {
// write to compactor files
MMFilesMarker* result;
int res = copyMarker(context->_compactor, marker, &result);
if (res != TRI_ERROR_NO_ERROR) {
THROW_ARANGO_EXCEPTION_MESSAGE(res, std::string("cannot write remove marker into compactor file: ") + TRI_errno_string(res));
}
// update datafile info
context->_dfi.numberDeletions++;
}
}
return true;
};
arangodb::SingleCollectionTransaction trx(
arangodb::transaction::StandaloneContext::Create(collection->vocbase()),
*collection,
AccessMode::Type::WRITE
);
trx.addHint(transaction::Hints::Hint::NO_BEGIN_MARKER);
trx.addHint(transaction::Hints::Hint::NO_ABORT_MARKER);
trx.addHint(transaction::Hints::Hint::NO_COMPACTION_LOCK);
trx.addHint(transaction::Hints::Hint::NO_THROTTLING);
// when we get into this function, the caller has already acquired the
// collection's status lock - so we better do not lock it again
trx.addHint(transaction::Hints::Hint::NO_USAGE_LOCK);
CompactionInitialContext initial = getCompactionContext(&trx, collection, toCompact);
if (initial._failed) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "could not create initialize compaction";
return;
}
LOG_TOPIC(DEBUG, Logger::COMPACTOR) << "compaction writes to be executed for collection '" << collection->id() << "', number of source datafiles: " << n << ", target datafile size: " << initial._targetSize;
// now create a new compactor file
// we are re-using the _fid of the first original datafile!
MMFilesDatafile* compactor = nullptr;
try {
compactor = physical->createCompactor(initial._fid, static_cast<uint32_t>(initial._targetSize));
} catch (std::exception const& ex) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "could not create compactor file: " << ex.what();
return;
} catch (...) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "could not create compactor file: unknown exception";
return;
}
TRI_ASSERT(compactor != nullptr);
LOG_TOPIC(DEBUG, Logger::COMPACTOR) << "created new compactor file '" << compactor->getName() << "', size: " << compactor->maximalSize();
// these attributes remain the same for all datafiles we collect
context->_collection = collection;
context->_compactor = compactor;
context->_trx = &trx;
Result res = trx.begin();
if (!res.ok()) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "error during compaction: " << res.errorMessage();
return;
}
// now compact all datafiles
uint64_t nrCombined = 0;
uint64_t compactionBytesRead = 0;
for (size_t i = 0; i < n; ++i) {
auto compaction = toCompact[i];
MMFilesDatafile* df = compaction._datafile;
compactionBytesRead += df->currentSize();
LOG_TOPIC(DEBUG, Logger::COMPACTOR) << "compacting datafile '" << df->getName() << "' into '" << compactor->getName() << "', number: " << i << ", keep deletions: " << compaction._keepDeletions;
// if this is the first datafile in the list of datafiles, we can also
// collect deletion markers
context->_keepDeletions = compaction._keepDeletions;
// run the actual compaction of a single datafile
bool ok;
try {
ok = TRI_IterateDatafile(df, compactifier);
} catch (std::exception const& ex) {
LOG_TOPIC(WARN, Logger::COMPACTOR) << "failed to compact datafile '" << df->getName() << "': " << ex.what();
throw;
}
if (!ok) {
LOG_TOPIC(WARN, Logger::COMPACTOR) << "failed to compact datafile '" << df->getName() << "'";
// compactor file does not need to be removed now. will be removed on next
// startup
return;
}
++nrCombined;
} // next file
TRI_ASSERT(context->_dfi.numberDead == 0);
TRI_ASSERT(context->_dfi.sizeDead == 0);
physical->_datafileStatistics.compactionRun(nrCombined, compactionBytesRead, context->_dfi.sizeAlive);
try {
physical->_datafileStatistics.replace(compactor->fid(), context->_dfi, true);
} catch (...) {
}
trx.commit();
// remove all datafile statistics that we don't need anymore
for (size_t i = 1; i < n; ++i) {
auto compaction = toCompact[i];
physical->_datafileStatistics.remove(compaction._datafile->fid());
}
if (physical->closeCompactor(compactor) != TRI_ERROR_NO_ERROR) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "could not close compactor file";
// TODO: how do we recover from this state?
return;
}
if (context->_dfi.numberAlive == 0 && context->_dfi.numberDead == 0 &&
context->_dfi.numberDeletions == 0) {
// everything is empty after compaction
if (n > 1) {
// create .dead files for all collected files
for (size_t i = 0; i < n; ++i) {
auto compaction = toCompact[i];
MMFilesDatafile* datafile = compaction._datafile;
if (datafile->isPhysical()) {
std::string filename(datafile->getName());
filename.append(".dead");
TRI_WriteFile(filename.c_str(), "", 0);
}
}
}
// compactor is fully empty. remove it
removeCompactor(collection, compactor);
for (size_t i = 0; i < n; ++i) {
auto compaction = toCompact[i];
// datafile is also empty after compaction and thus useless
removeDatafile(collection, compaction._datafile);
// add a deletion ditch to the collection
auto b = arangodb::MMFilesCollection::toMMFilesCollection(collection)
->ditches()
->createMMFilesDropDatafileDitch(compaction._datafile, collection,
DropDatafileCallback, __FILE__,
__LINE__);
if (b == nullptr) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "out of memory when creating datafile-drop ditch";
}
}
} else {
if (n > 1) {
// create .dead files for all collected files but the first
for (size_t i = 1; i < n; ++i) {
auto compaction = toCompact[i];
MMFilesDatafile* datafile = compaction._datafile;
if (datafile->isPhysical()) {
std::string filename(datafile->getName());
filename.append(".dead");
TRI_WriteFile(filename.c_str(), "", 0);
}
}
}
for (size_t i = 0; i < n; ++i) {
auto compaction = toCompact[i];
if (i == 0) {
// add a rename marker
auto b = arangodb::MMFilesCollection::toMMFilesCollection(collection)
->ditches()
->createMMFilesRenameDatafileDitch(
compaction._datafile, context->_compactor,
context->_collection, RenameDatafileCallback, __FILE__,
__LINE__);
if (b == nullptr) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "out of memory when creating datafile-rename ditch";
} else {
_vocbase.signalCleanup();
}
} else {
// datafile is empty after compaction and thus useless
removeDatafile(collection, compaction._datafile);
// add a drop datafile marker
auto b = arangodb::MMFilesCollection::toMMFilesCollection(collection)
->ditches()
->createMMFilesDropDatafileDitch(compaction._datafile, collection,
DropDatafileCallback, __FILE__,
__LINE__);
if (b == nullptr) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "out of memory when creating datafile-drop ditch";
} else {
_vocbase.signalCleanup();
}
}
}
}
}
/// @brief checks all datafiles of a collection
bool MMFilesCompactorThread::compactCollection(LogicalCollection* collection, bool& wasBlocked) {
// we can hopefully get away without the lock here...
// if (! document->isFullyCollected()) {
// return false;
// }
wasBlocked = false;
// if we cannot acquire the read lock instantly, we will exit directly.
// otherwise we'll risk a multi-thread deadlock between synchronizer,
// compactor and data-modification threads (e.g. POST /_api/document)
MMFilesCollection* physical = static_cast<MMFilesCollection*>(collection->getPhysical());
TRI_ASSERT(physical != nullptr);
TRY_READ_LOCKER(readLocker, physical->_filesLock);
if (!readLocker.isLocked()) {
// unable to acquire the lock at the moment
wasBlocked = true;
return false;
}
// check if there is already a compactor file
if (!physical->_compactors.empty()) {
// we already have created a compactor file in progress.
// if this happens, then a previous compaction attempt for this collection
// failed or is not finished yet
physical->setCompactionStatus(ReasonCompactionBlocked);
wasBlocked = true;
return false;
}
// copy datafiles vector
std::vector<MMFilesDatafile*> datafiles = physical->_datafiles;
if (datafiles.empty()) {
// collection has no datafiles
physical->setCompactionStatus(ReasonNoDatafiles);
return false;
}
std::vector<CompactionInfo> toCompact;
toCompact.reserve(MMFilesCompactionFeature::COMPACTOR->maxFiles());
// now we have datafiles that we can process
size_t const n = datafiles.size();
LOG_TOPIC(DEBUG, Logger::COMPACTOR) << "inspecting datafiles of collection '" << collection->name() << "' for compaction opportunities";
size_t start = physical->getNextCompactionStartIndex();
// get number of documents from collection
uint64_t const numDocuments = getNumberOfDocuments(*collection);
// get maximum size of result file
uint64_t maxSize = MMFilesCompactionFeature::COMPACTOR->maxSizeFactor() * static_cast<MMFilesCollection*>(collection->getPhysical())->journalSize();
if (maxSize < 8 * 1024 * 1024) {
maxSize = 8 * 1024 * 1024;
}
if (maxSize >= MMFilesCompactionFeature::COMPACTOR->maxResultFilesize()) {
maxSize = MMFilesCompactionFeature::COMPACTOR->maxResultFilesize();
}
if (start >= n || numDocuments == 0) {
start = 0;
}
int64_t numAlive = 0;
if (start > 0) {
// we don't know for sure if there are alive documents in the first
// datafile,
// so let's assume there are some
numAlive = 16384;
}
bool doCompact = false;
uint64_t totalSize = 0;
char const* reason = nullptr;
char const* firstReason = nullptr;
for (size_t i = start; i < n; ++i) {
MMFilesDatafile* df = datafiles[i];
if (df->state() == TRI_DF_STATE_OPEN_ERROR || df->state() == TRI_DF_STATE_WRITE_ERROR) {
LOG_TOPIC(WARN, Logger::COMPACTOR) << "cannot compact datafile " << df->fid() << " of collection '" << collection->name() << "' because it has errors";
physical->setCompactionStatus(ReasonCorrupted);
return false;
}
}
for (size_t i = start; i < n; ++i) {
MMFilesDatafile* df = datafiles[i];
TRI_ASSERT(df != nullptr);
MMFilesDatafileStatisticsContainer dfi = static_cast<MMFilesCollection*>(collection->getPhysical())->_datafileStatistics.get(df->fid());
if (dfi.numberUncollected > 0) {
LOG_TOPIC(DEBUG, Logger::COMPACTOR) << "cannot compact datafile " << df->fid() << " of collection '" << collection->name() << "' because it still has uncollected entries";
start = i + 1;
break;
}
if (!doCompact &&
df->maximalSize() < MMFilesCompactionFeature::COMPACTOR->smallDatafileSize() &&
(i < n - 1)) {
// very small datafile and not the last one. let's compact it so it's
// merged with others
doCompact = true;
reason = ReasonDatafileSmall;
} else if (numDocuments == 0 &&
(dfi.numberAlive > 0 || dfi.numberDead > 0 ||
dfi.numberDeletions > 0)) {
// collection is empty, but datafile statistics indicate there is
// something in this datafile
doCompact = true;
reason = ReasonEmpty;
} else if (numAlive == 0 && dfi.numberAlive == 0 &&
dfi.numberDeletions > 0) {
// compact first datafile(s) if they contain only deletions
doCompact = true;
reason = ReasonOnlyDeletions;
} else if (dfi.sizeDead >= MMFilesCompactionFeature::COMPACTOR->deadSizeThreshold()) {
// the size of dead objects is above some threshold
doCompact = true;
reason = ReasonDeadSize;
} else if (dfi.sizeDead > 0 &&
(((double)dfi.sizeDead /
((double)dfi.sizeDead + (double)dfi.sizeAlive) >= MMFilesCompactionFeature::COMPACTOR->deadShare()) ||
((double)dfi.sizeDead / (double)df->maximalSize() >= MMFilesCompactionFeature::COMPACTOR->deadShare()))) {
// the size of dead objects is above some share
doCompact = true;
reason = ReasonDeadSizeShare;
} else if (dfi.numberDead >= MMFilesCompactionFeature::COMPACTOR->deadNumberThreshold()) {
// the number of dead objects is above some threshold
doCompact = true;
reason = ReasonDeadCount;
}
if (!doCompact) {
numAlive += static_cast<int64_t>(dfi.numberAlive);
continue;
}
TRI_ASSERT(doCompact);
if (firstReason == nullptr) {
firstReason = reason;
}
// remember for next compaction
start = i + 1;
// if we got only deletions then it's safe to continue compaction, regardless of
// the size of the resulting file. this is because deletions will reduce the
// size of the resulting file
if (reason != ReasonOnlyDeletions) {
if (!toCompact.empty() &&
totalSize + (uint64_t)df->maximalSize() >= maxSize &&
(toCompact.size() != 1 || reason != ReasonDatafileSmall)) {
// found enough files to compact (in terms of cumulated size)
// there's one exception to this: if we're merging multiple datafiles,
// then we don't stop at the first one even if the merge of file #1 and #2
// would be too big. if we wouldn't stop in this case, then file #1 would
// be selected for compaction over and over
// normally this case won't happen at all, it can occur however if one
// decreases the journalSize configuration for the collection afterwards, and
// there are already datafiles which are more than 3 times bigger than the
// new (smaller) journalSize value
break;
}
}
TRI_ASSERT(reason != nullptr);
LOG_TOPIC(DEBUG, Logger::COMPACTOR) << "found datafile #" << i << " eligible for compaction. fid: " << df->fid() << ", size: " << df->maximalSize() << ", reason: " << reason << ", numberDead: " << dfi.numberDead << ", numberAlive: " << dfi.numberAlive << ", numberDeletions: " << dfi.numberDeletions << ", numberUncollected: " << dfi.numberUncollected << ", sizeDead: " << dfi.sizeDead << ", sizeAlive: " << dfi.sizeAlive;
totalSize += static_cast<uint64_t>(df->maximalSize());
CompactionInfo compaction;
compaction._datafile = df;
compaction._keepDeletions = (numAlive > 0 && i > 0);
// TODO: verify that keepDeletions actually works with wrong numAlive stats
try {
toCompact.push_back(compaction);
} catch (...) {
// silently fail. either we had found something to compact or not
// if not, then we can try again next time. if yes, then we'll simply
// forget
// about it and also try again next time
break;
}
// we stop at the first few datafiles.
// this is better than going over all datafiles in a collection in one go
// because the compactor is single-threaded, and collecting all datafiles
// might take a long time (it might even be that there is a request to
// delete the collection in the middle of compaction, but the compactor
// will not pick this up as it is read-locking the collection status)
if (totalSize >= maxSize) {
// result file will be big enough
break;
}
if (totalSize >= MMFilesCompactionFeature::COMPACTOR->smallDatafileSize() &&
toCompact.size() >= MMFilesCompactionFeature::COMPACTOR->maxFiles()) {
// found enough files to compact
break;
}
numAlive += static_cast<int64_t>(dfi.numberAlive);
}
// we can now continue without the lock
readLocker.unlock();
if (toCompact.empty()) {
// nothing to compact. now reset start index
physical->setNextCompactionStartIndex(0);
// cleanup local variables
physical->setCompactionStatus(ReasonNothingToCompact);
LOG_TOPIC(DEBUG, Logger::COMPACTOR) << "inspecting datafiles of collection yielded: " << ReasonNothingToCompact;
return false;
}
// handle datafiles with dead objects
TRI_ASSERT(toCompact.size() >= 1);
TRI_ASSERT(reason != nullptr);
physical->setCompactionStatus(reason);
physical->setNextCompactionStartIndex(start);
compactDatafiles(collection, toCompact);
return true;
}
MMFilesCompactorThread::MMFilesCompactorThread(TRI_vocbase_t& vocbase)
: Thread("MMFilesCompactor"), _vocbase(vocbase) {}
MMFilesCompactorThread::~MMFilesCompactorThread() { shutdown(); }
void MMFilesCompactorThread::signal() {
CONDITION_LOCKER(locker, _condition);
locker.signal();
}
void MMFilesCompactorThread::run() {
MMFilesEngine* engine = static_cast<MMFilesEngine*>(EngineSelectorFeature::ENGINE);
std::vector<std::shared_ptr<arangodb::LogicalCollection>> collections;
int numCompacted = 0;
while (true) {
// keep initial _state value as vocbase->_state might change during
// compaction loop
TRI_vocbase_t::State state = _vocbase.state();
try {
engine->tryPreventCompaction(
&_vocbase,
[this, &numCompacted, &collections, &engine](TRI_vocbase_t* vocbase) {
// compaction is currently allowed
numCompacted = 0;
try {
// copy all collections
collections = _vocbase.collections(false);
} catch (...) {
collections.clear();
}
for (auto& collection : collections) {
bool worked = false;
if (engine->isCompactionDisabled()) {
continue;
}
auto callback = [this, &collection, &worked, &engine]() -> void {
if (collection->status() != TRI_VOC_COL_STATUS_LOADED &&
collection->status() != TRI_VOC_COL_STATUS_UNLOADING) {
return;
}
bool doCompact = static_cast<MMFilesCollection*>(collection->getPhysical())->doCompact();
if (engine->isCompactionDisabled()) {
doCompact = false;
}
// for document collection, compactify datafiles
if (collection->status() == TRI_VOC_COL_STATUS_LOADED && doCompact) {
// check whether someone else holds a read-lock on the compaction
// lock
auto physical = static_cast<MMFilesCollection*>(collection->getPhysical());
TRI_ASSERT(physical != nullptr);
MMFilesTryCompactionLocker compactionLocker(physical);
if (!compactionLocker.isLocked()) {
// someone else is holding the compactor lock, we'll not compact
return;
}
try {
double const now = TRI_microtime();
if (physical->lastCompactionStamp() + MMFilesCompactionFeature::COMPACTOR->compactionCollectionInterval() <= now) {
auto ce = arangodb::MMFilesCollection::toMMFilesCollection(collection.get())
->ditches()
->createMMFilesCompactionDitch(__FILE__, __LINE__);
if (ce == nullptr) {
// out of memory
LOG_TOPIC(WARN, Logger::COMPACTOR) << "out of memory when trying to create compaction ditch";
} else {
try {
bool wasBlocked = false;
worked = compactCollection(collection.get(), wasBlocked);
if (!worked && !wasBlocked) {
// set compaction stamp
physical->lastCompactionStamp(now);
}
// if we worked or were blocked, then we don't set the compaction stamp to
// force another round of compaction
} catch (std::exception const& ex) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "caught exception during compaction: " << ex.what();
} catch (...) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "an unknown exception occurred during compaction";
// in case an error occurs, we must still free this ditch
}
arangodb::MMFilesCollection::toMMFilesCollection(collection.get())
->ditches()
->freeDitch(ce);
}
}
} catch (std::exception const& ex) {
LOG_TOPIC(ERR, Logger::COMPACTOR) << "caught exception during compaction: " << ex.what();
} catch (...) {
// in case an error occurs, we must still relase the lock
LOG_TOPIC(ERR, Logger::COMPACTOR) << "an unknown exception occurred during compaction";
}
}
};
if (!collection->tryExecuteWhileStatusLocked(callback)) {
continue;
}
if (worked) {
++numCompacted;
// signal the cleanup thread that we worked and that it can now wake
// up
CONDITION_LOCKER(locker, _condition);
locker.signal();
}
}
}, true);
if (numCompacted > 0) {
// no need to sleep long or go into wait state if we worked.
// maybe there's still work left
std::this_thread::sleep_for(std::chrono::microseconds(1000));
} else if (state != TRI_vocbase_t::State::SHUTDOWN_COMPACTOR
&& _vocbase.state() == TRI_vocbase_t::State::NORMAL) {
// only sleep while server is still running
CONDITION_LOCKER(locker, _condition);
_condition.wait(MMFilesCompactionFeature::COMPACTOR->compactionSleepTime());
}
if (state == TRI_vocbase_t::State::SHUTDOWN_COMPACTOR || isStopping()) {
// server shutdown or database has been removed
break;
}
} catch (...) {
// caught an error during compaction. simply ignore it and go on
}
}
LOG_TOPIC(TRACE, Logger::COMPACTOR) << "shutting down compactor thread";
}
/// @brief determine the number of documents in the collection
uint64_t MMFilesCompactorThread::getNumberOfDocuments(
LogicalCollection& collection
) {
SingleCollectionTransaction trx(
transaction::StandaloneContext::Create(_vocbase),
collection,
AccessMode::Type::READ
);
// only try to acquire the lock here
// if lock acquisition fails, we go on and report an (arbitrary) positive number
trx.addHint(transaction::Hints::Hint::TRY_LOCK);
trx.addHint(transaction::Hints::Hint::NO_THROTTLING);
// when we get into this function, the caller has already acquired the
// collection's status lock - so we better do not lock it again
trx.addHint(transaction::Hints::Hint::NO_USAGE_LOCK);
Result res = trx.begin();
if (!res.ok()) {
return 16384; // assume some positive value
}
return collection.numberDocuments(&trx, transaction::CountType::Normal);
}
/// @brief write a copy of the marker into the datafile
int MMFilesCompactorThread::copyMarker(MMFilesDatafile* compactor, MMFilesMarker const* marker,
MMFilesMarker** result) {
int res = compactor->reserveElement(marker->getSize(), result, 0);
if (res != TRI_ERROR_NO_ERROR) {
return TRI_ERROR_ARANGO_NO_JOURNAL;
}
return compactor->writeElement(*result, marker);
}
Reference in New Issue View Git Blame Copy Permalink