TEST_F(LoggingTests, BasicDeleteTest) {
auto recovery_table = ExecutorTestsUtil::CreateTable(1024);
auto &manager = catalog::Manager::GetInstance();
storage::Database db(DEFAULT_DB_ID);
manager.AddDatabase(&db);
db.AddTable(recovery_table);
EXPECT_EQ(recovery_table->GetNumberOfTuples(), 0);
EXPECT_EQ(recovery_table->GetTileGroupCount(), 1);
logging::WriteAheadFrontendLogger fel(true);
cid_t test_commit_id = 10;
auto curr_rec = new logging::TupleRecord(
LOGRECORD_TYPE_TUPLE_UPDATE, test_commit_id, recovery_table->GetOid(),
INVALID_ITEMPOINTER, ItemPointer(100, 4), nullptr, DEFAULT_DB_ID);
fel.DeleteTuple(curr_rec);
delete curr_rec;
auto tg_header = recovery_table->GetTileGroupById(100)->GetHeader();
EXPECT_EQ(tg_header->GetEndCommitId(4), test_commit_id);
// EXPECT_EQ(recovery_table->GetNumberOfTuples(), 1);
EXPECT_EQ(recovery_table->GetTileGroupCount(), 2);
}
bool PessimisticTxnManager::PerformUpdate(const oid_t &tile_group_id,
const oid_t &tuple_id,
const ItemPointer &new_location) {
LOG_INFO("Performing Write %lu %lu", tile_group_id, tuple_id);
auto transaction_id = current_txn->GetTransactionId();
auto tile_group_header =
catalog::Manager::GetInstance().GetTileGroup(tile_group_id)->GetHeader();
auto new_tile_group_header = catalog::Manager::GetInstance()
.GetTileGroup(new_location.block)->GetHeader();
// if we can perform update, then we must have already locked the older
// version.
assert(tile_group_header->GetTransactionId(tuple_id) == transaction_id);
assert(new_tile_group_header->GetTransactionId(new_location.offset) == INVALID_TXN_ID);
assert(new_tile_group_header->GetBeginCommitId(new_location.offset) == MAX_CID);
assert(new_tile_group_header->GetEndCommitId(new_location.offset) == MAX_CID);
tile_group_header->SetTransactionId(tuple_id, transaction_id);
// The write lock must have been acquired
// Notice: if the executor doesn't call PerformUpdate after AcquireOwnership,
// no
// one will possibly release the write lock acquired by this txn.
// Set double linked list
tile_group_header->SetNextItemPointer(tuple_id, new_location);
new_tile_group_header->SetPrevItemPointer(
new_location.offset, ItemPointer(tile_group_id, tuple_id));
new_tile_group_header->SetTransactionId(new_location.offset, transaction_id);
// Add the old tuple into the update set
current_txn->RecordUpdate(tile_group_id, tuple_id);
return true;
}
const std::string TileGroupHeader::GetInfo() const {
std::ostringstream os;
os << "\t-----------------------------------------------------------\n";
os << "\tTILE GROUP HEADER \n";
oid_t active_tuple_slots = GetCurrentNextTupleSlot();
peloton::ItemPointer item;
for (oid_t header_itr = 0; header_itr < active_tuple_slots; header_itr++) {
txn_id_t txn_id = GetTransactionId(header_itr);
cid_t beg_commit_id = GetBeginCommitId(header_itr);
cid_t end_commit_id = GetEndCommitId(header_itr);
bool insert_commit = GetInsertCommit(header_itr);
bool delete_commit = GetDeleteCommit(header_itr);
int width = 10;
os << "\t txn id : ";
if (txn_id == MAX_TXN_ID)
os << std::setw(width) << "MAX_TXN_ID";
else
os << std::setw(width) << txn_id;
os << " beg cid : ";
if (beg_commit_id == MAX_CID)
os << std::setw(width) << "MAX_CID";
else
os << std::setw(width) << beg_commit_id;
os << " end cid : ";
if (end_commit_id == MAX_CID)
os << std::setw(width) << "MAX_CID";
else
os << std::setw(width) << end_commit_id;
os << " insert commit : ";
if (insert_commit == true)
os << "O";
else
os << "X";
os << " delete commit : ";
if (delete_commit == true)
os << "O";
else
os << "X";
peloton::ItemPointer location = GetNextItemPointer(header_itr);
peloton::ItemPointer location2 = GetPrevItemPointer(header_itr);
os << " next : "
<< "[ " << location.block << " , " << location.offset << " ]\n"
<< " prev : "
<< "[ " << location2.block << " , " << location2.offset << " ]\n";
}
os << "\t-----------------------------------------------------------\n";
return os.str();
}
void PessimisticTxnManager::SetOwnership(const oid_t &tile_group_id,
const oid_t &tuple_id) {
auto &manager = catalog::Manager::GetInstance();
auto tile_group_header = manager.GetTileGroup(tile_group_id)->GetHeader();
auto transaction_id = current_txn->GetTransactionId();
// Set MVCC info
assert(tile_group_header->GetTransactionId(tuple_id) == INVALID_TXN_ID);
assert(tile_group_header->GetBeginCommitId(tuple_id) == MAX_CID);
assert(tile_group_header->GetEndCommitId(tuple_id) == MAX_CID);
tile_group_header->SetTransactionId(tuple_id, transaction_id);
}
void PessimisticTxnManager::PerformUpdate(const oid_t &tile_group_id,
const oid_t &tuple_id) {
auto &manager = catalog::Manager::GetInstance();
auto tile_group_header = manager.GetTileGroup(tile_group_id)->GetHeader();
// Set MVCC info
assert(tile_group_header->GetTransactionId(tuple_id) == current_txn->GetTransactionId());
assert(tile_group_header->GetBeginCommitId(tuple_id) == MAX_CID);
assert(tile_group_header->GetEndCommitId(tuple_id) == MAX_CID);
// Add the old tuple into the update set
auto old_location = tile_group_header->GetPrevItemPointer(tuple_id);
if (old_location.IsNull() == false) {
// update an inserted version
current_txn->RecordUpdate(old_location.block, old_location.offset);
}
}
TEST_F(LoggingTests, BasicUpdateTest) {
auto recovery_table = ExecutorTestsUtil::CreateTable(1024);
auto &manager = catalog::Manager::GetInstance();
storage::Database db(DEFAULT_DB_ID);
manager.AddDatabase(&db);
db.AddTable(recovery_table);
auto tuples = BuildLoggingTuples(recovery_table, 1, false, false);
EXPECT_EQ(recovery_table->GetNumberOfTuples(), 0);
EXPECT_EQ(recovery_table->GetTileGroupCount(), 1);
EXPECT_EQ(tuples.size(), 1);
logging::WriteAheadFrontendLogger fel(true);
// auto bel = logging::WriteAheadBackendLogger::GetInstance();
cid_t test_commit_id = 10;
Value val0 = tuples[0]->GetValue(0);
Value val1 = tuples[0]->GetValue(1);
Value val2 = tuples[0]->GetValue(2);
Value val3 = tuples[0]->GetValue(3);
auto curr_rec = new logging::TupleRecord(
LOGRECORD_TYPE_TUPLE_UPDATE, test_commit_id, recovery_table->GetOid(),
ItemPointer(100, 5), ItemPointer(100, 4), tuples[0], DEFAULT_DB_ID);
curr_rec->SetTuple(tuples[0]);
fel.UpdateTuple(curr_rec);
delete curr_rec;
auto tg_header = recovery_table->GetTileGroupById(100)->GetHeader();
EXPECT_TRUE(tg_header->GetBeginCommitId(5) <= test_commit_id);
EXPECT_EQ(tg_header->GetEndCommitId(5), MAX_CID);
EXPECT_EQ(tg_header->GetEndCommitId(4), test_commit_id);
EXPECT_TRUE(
val0.Compare(recovery_table->GetTileGroupById(100)->GetValue(5, 0)) == 0);
EXPECT_TRUE(
val1.Compare(recovery_table->GetTileGroupById(100)->GetValue(5, 1)) == 0);
EXPECT_TRUE(
val2.Compare(recovery_table->GetTileGroupById(100)->GetValue(5, 2)) == 0);
EXPECT_TRUE(
val3.Compare(recovery_table->GetTileGroupById(100)->GetValue(5, 3)) == 0);
EXPECT_EQ(recovery_table->GetNumberOfTuples(), 0);
EXPECT_EQ(recovery_table->GetTileGroupCount(), 2);
}
std::pair<cid_t, storage::TileGroupHeader *>
WriteBehindFrontendLogger::SetDeleteCommitMark(ItemPointer location) {
auto &manager = catalog::Manager::GetInstance();
auto tile_group = manager.GetTileGroup(location.block);
assert(tile_group != nullptr);
auto tile_group_header = tile_group->GetHeader();
assert(tile_group_header != nullptr);
// Set the commit mark
tile_group_header->SetDeleteCommit(location.offset, true);
LOG_TRACE("<%p, %lu> : slot is delete committed", tile_group.get(),
location.offset);
// Update max oid
if (max_oid < location.block) {
max_oid = location.block;
}
auto end_commit_id = tile_group_header->GetEndCommitId(location.offset);
return std::make_pair(end_commit_id, tile_group_header);
}
void PessimisticTxnManager::PerformDelete(const oid_t &tile_group_id,
const oid_t &tuple_id) {
auto &manager = catalog::Manager::GetInstance();
auto tile_group_header = manager.GetTileGroup(tile_group_id)->GetHeader();
assert(tile_group_header->GetTransactionId(tuple_id) ==
current_txn->GetTransactionId());
assert(tile_group_header->GetBeginCommitId(tuple_id) == MAX_CID);
assert(tile_group_header->GetEndCommitId(tuple_id) == MAX_CID);
tile_group_header->SetEndCommitId(tuple_id, INVALID_CID);
// Add the old tuple into the delete set
auto old_location = tile_group_header->GetPrevItemPointer(tuple_id);
if (old_location.IsNull() == false) {
// delete an inserted version
current_txn->RecordDelete(old_location.block, old_location.offset);
} else {
// if this version is newly inserted.
current_txn->RecordDelete(tile_group_id, tuple_id);
}
}
TEST_F(LoggingTests, OutOfOrderCommitTest) {
auto recovery_table = ExecutorTestsUtil::CreateTable(1024);
auto &manager = catalog::Manager::GetInstance();
storage::Database db(DEFAULT_DB_ID);
manager.AddDatabase(&db);
db.AddTable(recovery_table);
auto tuples = BuildLoggingTuples(recovery_table, 1, false, false);
EXPECT_EQ(recovery_table->GetNumberOfTuples(), 0);
EXPECT_EQ(recovery_table->GetTileGroupCount(), 1);
EXPECT_EQ(tuples.size(), 1);
logging::WriteAheadFrontendLogger fel(true);
// auto bel = logging::WriteAheadBackendLogger::GetInstance();
cid_t test_commit_id = 10;
auto curr_rec = new logging::TupleRecord(
LOGRECORD_TYPE_TUPLE_UPDATE, test_commit_id + 1, recovery_table->GetOid(),
INVALID_ITEMPOINTER, ItemPointer(100, 5), nullptr, DEFAULT_DB_ID);
fel.DeleteTuple(curr_rec);
delete curr_rec;
EXPECT_EQ(recovery_table->GetTileGroupCount(), 2);
curr_rec = new logging::TupleRecord(
LOGRECORD_TYPE_TUPLE_INSERT, test_commit_id, recovery_table->GetOid(),
ItemPointer(100, 5), INVALID_ITEMPOINTER, tuples[0], DEFAULT_DB_ID);
curr_rec->SetTuple(tuples[0]);
fel.InsertTuple(curr_rec);
delete curr_rec;
auto tg_header = recovery_table->GetTileGroupById(100)->GetHeader();
EXPECT_EQ(tg_header->GetEndCommitId(5), test_commit_id + 1);
EXPECT_EQ(recovery_table->GetNumberOfTuples(), 0);
EXPECT_EQ(recovery_table->GetTileGroupCount(), 2);
}
// count number of expired versions.
int GarbageNum(storage::DataTable *table) {
auto table_tile_group_count_ = table->GetTileGroupCount();
auto current_tile_group_offset_ = START_OID;
int old_num = 0;
while (current_tile_group_offset_ < table_tile_group_count_) {
auto tile_group = table->GetTileGroup(current_tile_group_offset_++);
auto tile_group_header = tile_group->GetHeader();
oid_t active_tuple_count = tile_group->GetNextTupleSlot();
for (oid_t tuple_id = 0; tuple_id < active_tuple_count; tuple_id++) {
auto tuple_txn_id = tile_group_header->GetTransactionId(tuple_id);
auto tuple_end_cid = tile_group_header->GetEndCommitId(tuple_id);
if (tuple_txn_id == INITIAL_TXN_ID && tuple_end_cid != MAX_CID) {
old_num++;
}
}
}
LOG_INFO("old version num = %d", old_num);
return old_num;
}
void TimestampOrderingTransactionManager::PerformInsert(
TransactionContext *const current_txn, const ItemPointer &location,
ItemPointer *index_entry_ptr) {
PL_ASSERT(current_txn->GetIsolationLevel() != IsolationLevelType::READ_ONLY);
oid_t tile_group_id = location.block;
oid_t tuple_id = location.offset;
auto &manager = catalog::Manager::GetInstance();
auto tile_group_header = manager.GetTileGroup(tile_group_id)->GetHeader();
auto transaction_id = current_txn->GetTransactionId();
// check MVCC info
// the tuple slot must be empty.
PL_ASSERT(tile_group_header->GetTransactionId(tuple_id) == INVALID_TXN_ID);
PL_ASSERT(tile_group_header->GetBeginCommitId(tuple_id) == MAX_CID);
PL_ASSERT(tile_group_header->GetEndCommitId(tuple_id) == MAX_CID);
tile_group_header->SetTransactionId(tuple_id, transaction_id);
// no need to set next item pointer.
// Add the new tuple into the insert set
current_txn->RecordInsert(location);
InitTupleReserved(tile_group_header, tuple_id);
// Write down the head pointer's address in tile group header
tile_group_header->SetIndirection(tuple_id, index_entry_ptr);
// Increment table insert op stats
if (static_cast<StatsType>(settings::SettingsManager::GetInt(settings::SettingId::stats_mode)) !=
StatsType::INVALID) {
stats::BackendStatsContext::GetInstance()->IncrementTableInserts(
location.block);
}
}
// Validate that MVCC storage is correct, it assumes an old-to-new chain
// Invariants
// 1. Transaction id should either be INVALID_TXNID or INITIAL_TXNID
// 2. Begin commit id should <= end commit id
// 3. Timestamp consistence
// 4. Version doubly linked list consistency
static void ValidateMVCC_OldToNew(storage::DataTable *table) {
auto &catalog_manager = catalog::Manager::GetInstance();
LOG_INFO("Validating MVCC storage");
int tile_group_count = table->GetTileGroupCount();
LOG_INFO("The table has %d tile groups in the table", tile_group_count);
for (int tile_group_offset = 0; tile_group_offset < tile_group_count;
tile_group_offset++) {
LOG_INFO("Validate tile group #%d", tile_group_offset);
auto tile_group = table->GetTileGroup(tile_group_offset);
auto tile_group_header = tile_group->GetHeader();
size_t tuple_count = tile_group->GetAllocatedTupleCount();
LOG_INFO("Tile group #%d has allocated %lu tuples", tile_group_offset,
tuple_count);
// 1. Transaction id should either be INVALID_TXNID or INITIAL_TXNID
for (oid_t tuple_slot = 0; tuple_slot < tuple_count; tuple_slot++) {
txn_id_t txn_id = tile_group_header->GetTransactionId(tuple_slot);
EXPECT_TRUE(txn_id == INVALID_TXN_ID || txn_id == INITIAL_TXN_ID)
<< "Transaction id is not INVALID_TXNID or INITIAL_TXNID";
}
LOG_INFO("[OK] All tuples have valid txn id");
// double avg_version_chain_length = 0.0;
for (oid_t tuple_slot = 0; tuple_slot < tuple_count; tuple_slot++) {
txn_id_t txn_id = tile_group_header->GetTransactionId(tuple_slot);
cid_t begin_cid = tile_group_header->GetBeginCommitId(tuple_slot);
cid_t end_cid = tile_group_header->GetEndCommitId(tuple_slot);
ItemPointer next_location =
tile_group_header->GetNextItemPointer(tuple_slot);
ItemPointer prev_location =
tile_group_header->GetPrevItemPointer(tuple_slot);
// 2. Begin commit id should <= end commit id
EXPECT_TRUE(begin_cid <= end_cid)
<< "Tuple begin commit id is less than or equal to end commit id";
// This test assumes a oldest-to-newest version chain
if (txn_id != INVALID_TXN_ID) {
EXPECT_TRUE(begin_cid != MAX_CID)
<< "Non invalid txn shouldn't have a MAX_CID begin commit id";
// The version is an oldest version
if (prev_location.IsNull()) {
if (next_location.IsNull()) {
EXPECT_EQ(end_cid, MAX_CID)
<< "Single version has a non MAX_CID end commit time";
} else {
cid_t prev_end_cid = end_cid;
ItemPointer prev_location(tile_group->GetTileGroupId(), tuple_slot);
while (!next_location.IsNull()) {
auto next_tile_group =
catalog_manager.GetTileGroup(next_location.block);
auto next_tile_group_header = next_tile_group->GetHeader();
txn_id_t next_txn_id = next_tile_group_header->GetTransactionId(
next_location.offset);
if (next_txn_id == INVALID_TXN_ID) {
// If a version in the version chain has a INVALID_TXN_ID, it
// must be at the tail
// of the chain. It is either because we have deleted a tuple
// (so append a invalid tuple),
// or because this new version is aborted.
EXPECT_TRUE(
next_tile_group_header->GetNextItemPointer(
next_location.offset).IsNull())
<< "Invalid version in a version chain and is not delete";
}
cid_t next_begin_cid = next_tile_group_header->GetBeginCommitId(
next_location.offset);
cid_t next_end_cid =
next_tile_group_header->GetEndCommitId(next_location.offset);
// 3. Timestamp consistence
if (next_begin_cid == MAX_CID) {
// It must be an aborted version, it should be at the end of the
// chain
EXPECT_TRUE(
next_tile_group_header->GetNextItemPointer(
next_location.offset).IsNull())
<< "Version with MAX_CID begin cid is not version tail";
} else {
EXPECT_EQ(prev_end_cid, next_begin_cid)
<< "Prev end commit id should equal net begin commit id";
ItemPointer next_prev_location =
next_tile_group_header->GetPrevItemPointer(
next_location.offset);
// 4. Version doubly linked list consistency
EXPECT_TRUE(next_prev_location.offset == prev_location.offset &&
next_prev_location.block == prev_location.block)
<< "Next version's prev version does not match";
}
prev_location = next_location;
prev_end_cid = next_end_cid;
next_location = next_tile_group_header->GetNextItemPointer(
next_location.offset);
}
// Now prev_location is at the tail of the version chain
ItemPointer last_location = prev_location;
auto last_tile_group =
catalog_manager.GetTileGroup(last_location.block);
auto last_tile_group_header = last_tile_group->GetHeader();
// txn_id_t last_txn_id =
// last_tile_group_header->GetTransactionId(last_location.offset);
cid_t last_end_cid =
last_tile_group_header->GetEndCommitId(last_location.offset);
EXPECT_TRUE(
last_tile_group_header->GetNextItemPointer(last_location.offset)
.IsNull())
<< "Last version has a next pointer";
EXPECT_EQ(last_end_cid, MAX_CID)
<< "Last version doesn't end with MAX_CID";
}
}
} else {
EXPECT_TRUE(tile_group_header->GetNextItemPointer(tuple_slot).IsNull())
<< "Invalid tuple must not have next item pointer";
}
}
LOG_INFO("[OK] oldest-to-newest version chain validated");
}
}
bool IndexScanExecutor::ExecPrimaryIndexLookup() {
PL_ASSERT(!done_);
std::vector<ItemPointer *> tuple_location_ptrs;
PL_ASSERT(index_->GetIndexType() == INDEX_CONSTRAINT_TYPE_PRIMARY_KEY);
if (0 == key_column_ids_.size()) {
index_->ScanAllKeys(tuple_location_ptrs);
} else {
index_->Scan(values_, key_column_ids_, expr_types_,
SCAN_DIRECTION_TYPE_FORWARD, tuple_location_ptrs);
}
if (tuple_location_ptrs.size() == 0) return false;
auto &transaction_manager =
concurrency::TransactionManagerFactory::GetInstance();
std::map<oid_t, std::vector<oid_t>> visible_tuples;
std::vector<ItemPointer> garbage_tuples;
// for every tuple that is found in the index.
for (auto tuple_location_ptr : tuple_location_ptrs) {
ItemPointer tuple_location = *tuple_location_ptr;
auto &manager = catalog::Manager::GetInstance();
auto tile_group = manager.GetTileGroup(tuple_location.block);
auto tile_group_header = tile_group.get()->GetHeader();
size_t chain_length = 0;
while (true) {
++chain_length;
// if the tuple is visible.
if (transaction_manager.IsVisible(tile_group_header,
tuple_location.offset)) {
LOG_TRACE("traverse chain length : %lu", chain_length);
LOG_TRACE("perform read: %u, %u", tuple_location.block,
tuple_location.offset);
// perform predicate evaluation.
if (predicate_ == nullptr) {
visible_tuples[tuple_location.block].push_back(tuple_location.offset);
auto res = transaction_manager.PerformRead(tuple_location);
if (!res) {
transaction_manager.SetTransactionResult(RESULT_FAILURE);
return res;
}
} else {
expression::ContainerTuple<storage::TileGroup> tuple(
tile_group.get(), tuple_location.offset);
auto eval =
predicate_->Evaluate(&tuple, nullptr, executor_context_).IsTrue();
if (eval == true) {
visible_tuples[tuple_location.block]
.push_back(tuple_location.offset);
auto res = transaction_manager.PerformRead(tuple_location);
if (!res) {
transaction_manager.SetTransactionResult(RESULT_FAILURE);
return res;
}
}
}
break;
}
// if the tuple is not visible.
else {
ItemPointer old_item = tuple_location;
cid_t old_end_cid = tile_group_header->GetEndCommitId(old_item.offset);
tuple_location = tile_group_header->GetNextItemPointer(old_item.offset);
// there must exist a visible version.
// FIXME: currently, only speculative read transaction manager **may** see a null version
// it's a potential bug
if(tuple_location.IsNull()) {
transaction_manager.SetTransactionResult(RESULT_FAILURE);
// FIXME: this cause unnecessary abort when we have delete operations
return false;
}
// FIXME: Is this always true? what if we have a deleted tuple? --jiexi
PL_ASSERT(tuple_location.IsNull() == false);
cid_t max_committed_cid = transaction_manager.GetMaxCommittedCid();
// check whether older version is garbage.
if (old_end_cid <= max_committed_cid) {
PL_ASSERT(tile_group_header->GetTransactionId(old_item.offset) == INITIAL_TXN_ID ||
tile_group_header->GetTransactionId(old_item.offset) == INVALID_TXN_ID);
if (tile_group_header->SetAtomicTransactionId(old_item.offset, INVALID_TXN_ID) == true) {
// atomically swap item pointer held in the index bucket.
AtomicUpdateItemPointer(tuple_location_ptr, tuple_location);
// currently, let's assume only primary index exists.
// gc::GCManagerFactory::GetInstance().RecycleTupleSlot(
// table_->GetOid(), old_item.block, old_item.offset,
// transaction_manager.GetNextCommitId());
garbage_tuples.push_back(old_item);
tile_group = manager.GetTileGroup(tuple_location.block);
tile_group_header = tile_group.get()->GetHeader();
tile_group_header->SetPrevItemPointer(tuple_location.offset, INVALID_ITEMPOINTER);
} else {
tile_group = manager.GetTileGroup(tuple_location.block);
tile_group_header = tile_group.get()->GetHeader();
}
} else {
tile_group = manager.GetTileGroup(tuple_location.block);
tile_group_header = tile_group.get()->GetHeader();
}
}
}
}
// Add all garbage tuples to GC manager
if(garbage_tuples.size() != 0) {
cid_t garbage_timestamp = transaction_manager.GetNextCommitId();
for (auto garbage : garbage_tuples) {
gc::GCManagerFactory::GetInstance().RecycleTupleSlot(
table_->GetOid(), garbage.block, garbage.offset, garbage_timestamp);
}
}
// Construct a logical tile for each block
for (auto tuples : visible_tuples) {
auto &manager = catalog::Manager::GetInstance();
auto tile_group = manager.GetTileGroup(tuples.first);
std::unique_ptr<LogicalTile> logical_tile(LogicalTileFactory::GetTile());
// Add relevant columns to logical tile
logical_tile->AddColumns(tile_group, full_column_ids_);
logical_tile->AddPositionList(std::move(tuples.second));
if (column_ids_.size() != 0) {
logical_tile->ProjectColumns(full_column_ids_, column_ids_);
}
result_.push_back(logical_tile.release());
}
done_ = true;
LOG_TRACE("Result tiles : %lu", result_.size());
return true;
}
void TileGroupHeader::PrintVisibility(txn_id_t txn_id, cid_t at_cid) {
oid_t active_tuple_slots = GetNextTupleSlot();
std::stringstream os;
os << "\t-----------------------------------------------------------\n";
for (oid_t header_itr = 0; header_itr < active_tuple_slots; header_itr++) {
bool own = (txn_id == GetTransactionId(header_itr));
bool activated = (at_cid >= GetBeginCommitId(header_itr));
bool invalidated = (at_cid >= GetEndCommitId(header_itr));
txn_id_t txn_id = GetTransactionId(header_itr);
cid_t beg_commit_id = GetBeginCommitId(header_itr);
cid_t end_commit_id = GetEndCommitId(header_itr);
bool insert_commit = GetInsertCommit(header_itr);
bool delete_commit = GetDeleteCommit(header_itr);
int width = 10;
os << "\tslot :: " << std::setw(width) << header_itr;
os << " txn id : ";
if (txn_id == MAX_TXN_ID)
os << std::setw(width) << "MAX_TXN_ID";
else
os << std::setw(width) << txn_id;
os << " beg cid : ";
if (beg_commit_id == MAX_CID)
os << std::setw(width) << "MAX_CID";
else
os << std::setw(width) << beg_commit_id;
os << " end cid : ";
if (end_commit_id == MAX_CID)
os << std::setw(width) << "MAX_CID";
else
os << std::setw(width) << end_commit_id;
os << " insert commit : ";
if (insert_commit == true)
os << "O";
else
os << "X";
os << " delete commit : ";
if (delete_commit == true)
os << "O";
else
os << "X";
peloton::ItemPointer location = GetPrevItemPointer(header_itr);
os << " prev : "
<< "[ " << location.block << " , " << location.offset << " ]"; //<<
os << " own : " << own;
os << " activated : " << activated;
os << " invalidated : " << invalidated << " ";
// Visible iff past Insert || Own Insert
if ((!own && activated && !invalidated) ||
(own && !activated && !invalidated))
os << "\t\t[ true ]\n";
else
os << "\t\t[ false ]\n";
}
os << "\t-----------------------------------------------------------\n";
std::cout << os.str().c_str();
}
void TimestampOrderingTransactionManager::PerformUpdate(
TransactionContext *const current_txn, const ItemPointer &location,
const ItemPointer &new_location) {
PL_ASSERT(current_txn->GetIsolationLevel() != IsolationLevelType::READ_ONLY);
ItemPointer old_location = location;
LOG_TRACE("Performing Update old tuple %u %u", old_location.block,
old_location.offset);
LOG_TRACE("Performing Update new tuple %u %u", new_location.block,
new_location.offset);
auto &manager = catalog::Manager::GetInstance();
auto tile_group_header =
manager.GetTileGroup(old_location.block)->GetHeader();
auto new_tile_group_header =
manager.GetTileGroup(new_location.block)->GetHeader();
auto transaction_id = current_txn->GetTransactionId();
// if we can perform update, then we must have already locked the older
// version.
PL_ASSERT(tile_group_header->GetTransactionId(old_location.offset) ==
transaction_id);
PL_ASSERT(tile_group_header->GetPrevItemPointer(old_location.offset)
.IsNull() == true);
// check whether the new version is empty.
PL_ASSERT(new_tile_group_header->GetTransactionId(new_location.offset) ==
INVALID_TXN_ID);
PL_ASSERT(new_tile_group_header->GetBeginCommitId(new_location.offset) ==
MAX_CID);
PL_ASSERT(new_tile_group_header->GetEndCommitId(new_location.offset) ==
MAX_CID);
// if the executor doesn't call PerformUpdate after AcquireOwnership,
// no one will possibly release the write lock acquired by this txn.
// Set double linked list
tile_group_header->SetPrevItemPointer(old_location.offset, new_location);
new_tile_group_header->SetNextItemPointer(new_location.offset, old_location);
new_tile_group_header->SetTransactionId(new_location.offset, transaction_id);
// we should guarantee that the newer version is all set before linking the
// newer version to older version.
COMPILER_MEMORY_FENCE;
InitTupleReserved(new_tile_group_header, new_location.offset);
// we must be updating the latest version.
// Set the header information for the new version
ItemPointer *index_entry_ptr =
tile_group_header->GetIndirection(old_location.offset);
// if there's no primary index on a table, then index_entry_ptr == nullptr.
if (index_entry_ptr != nullptr) {
new_tile_group_header->SetIndirection(new_location.offset, index_entry_ptr);
// Set the index header in an atomic way.
// We do it atomically because we don't want any one to see a half-done
// pointer.
// In case of contention, no one can update this pointer when we are
// updating it
// because we are holding the write lock. This update should success in
// its first trial.
UNUSED_ATTRIBUTE auto res =
AtomicUpdateItemPointer(index_entry_ptr, new_location);
PL_ASSERT(res == true);
}
// Add the old tuple into the update set
current_txn->RecordUpdate(old_location);
// Increment table update op stats
if (static_cast<StatsType>(settings::SettingsManager::GetInt(settings::SettingId::stats_mode)) !=
StatsType::INVALID) {
stats::BackendStatsContext::GetInstance()->IncrementTableUpdates(
new_location.block);
}
}
bool HybridScanExecutor::ExecPrimaryIndexLookup() {
assert(!index_done_);
std::vector<ItemPointer *> tuple_location_ptrs;
assert(index_->GetIndexType() == INDEX_CONSTRAINT_TYPE_PRIMARY_KEY);
if (0 == key_column_ids_.size()) {
index_->ScanAllKeys(tuple_location_ptrs);
} else {
index_->Scan(values_, key_column_ids_, expr_types_,
SCAN_DIRECTION_TYPE_FORWARD, tuple_location_ptrs);
}
LOG_INFO("Tuple_locations.size(): %lu", tuple_location_ptrs.size());
auto &transaction_manager =
concurrency::TransactionManagerFactory::GetInstance();
if (tuple_location_ptrs.size() == 0) {
index_done_ = true;
return false;
}
//std::set<oid_t> oid_ts;
std::map<oid_t, std::vector<oid_t>> visible_tuples;
// for every tuple that is found in the index.
for (auto tuple_location_ptr : tuple_location_ptrs) {
ItemPointer tuple_location = *tuple_location_ptr;
if (type_ == planner::HYBRID &&
tuple_location.block >= (block_threshold)) {
item_pointers_.insert(tuple_location);
// oid_ts.insert(tuple_location.block);
}
auto &manager = catalog::Manager::GetInstance();
auto tile_group = manager.GetTileGroup(tuple_location.block);
auto tile_group_header = tile_group.get()->GetHeader();
// perform transaction read
size_t chain_length = 0;
while (true) {
++chain_length;
if (transaction_manager.IsVisible(tile_group_header,
tuple_location.offset)) {
visible_tuples[tuple_location.block].push_back(tuple_location.offset);
auto res = transaction_manager.PerformRead(tuple_location);
if (!res) {
transaction_manager.SetTransactionResult(RESULT_FAILURE);
return res;
}
break;
} else {
ItemPointer old_item = tuple_location;
cid_t old_end_cid = tile_group_header->GetEndCommitId(old_item.offset);
tuple_location = tile_group_header->GetNextItemPointer(old_item.offset);
// there must exist a visible version.
assert(tuple_location.IsNull() == false);
cid_t max_committed_cid = transaction_manager.GetMaxCommittedCid();
// check whether older version is garbage.
if (old_end_cid < max_committed_cid) {
assert(tile_group_header->GetTransactionId(old_item.offset) == INITIAL_TXN_ID ||
tile_group_header->GetTransactionId(old_item.offset) == INVALID_TXN_ID);
if (tile_group_header->SetAtomicTransactionId(old_item.offset, INVALID_TXN_ID) == true) {
// atomically swap item pointer held in the index bucket.
AtomicUpdateItemPointer(tuple_location_ptr, tuple_location);
// currently, let's assume only primary index exists.
//gc::GCManagerFactory::GetInstance().RecycleTupleSlot(
// table_->GetOid(), old_item.block, old_item.offset,
// max_committed_cid);
}
}
tile_group = manager.GetTileGroup(tuple_location.block);
tile_group_header = tile_group.get()->GetHeader();
}
}
}
// Construct a logical tile for each block
for (auto tuples : visible_tuples) {
auto &manager = catalog::Manager::GetInstance();
auto tile_group = manager.GetTileGroup(tuples.first);
std::unique_ptr<LogicalTile> logical_tile(LogicalTileFactory::GetTile());
// Add relevant columns to logical tile
logical_tile->AddColumns(tile_group, full_column_ids_);
logical_tile->AddPositionList(std::move(tuples.second));
if (column_ids_.size() != 0) {
logical_tile->ProjectColumns(full_column_ids_, column_ids_);
}
result_.push_back(logical_tile.release());
}
index_done_ = true;
LOG_TRACE("Result tiles : %lu", result_.size());
return true;
}
bool HybridScanExecutor::ExecPrimaryIndexLookup() {
PL_ASSERT(index_done_ == false);
const planner::HybridScanPlan &node = GetPlanNode<planner::HybridScanPlan>();
bool acquire_owner = GetPlanNode<planner::AbstractScan>().IsForUpdate();
auto key_column_ids_ = node.GetKeyColumnIds();
auto expr_type_ = node.GetExprTypes();
std::vector<ItemPointer *> tuple_location_ptrs;
PL_ASSERT(index_->GetIndexType() == INDEX_CONSTRAINT_TYPE_PRIMARY_KEY);
if (0 == key_column_ids_.size()) {
LOG_TRACE("Scan all keys");
index_->ScanAllKeys(tuple_location_ptrs);
} else {
LOG_TRACE("Scan");
index_->Scan(values_,
key_column_ids_,
expr_type_,
SCAN_DIRECTION_TYPE_FORWARD,
tuple_location_ptrs,
&node.GetIndexPredicate().GetConjunctionList()[0]);
}
LOG_TRACE("Result tuple count: %lu", tuple_location_ptrs.size());
auto &transaction_manager =
concurrency::TransactionManagerFactory::GetInstance();
auto current_txn = executor_context_->GetTransaction();
if (tuple_location_ptrs.size() == 0) {
index_done_ = true;
return false;
}
std::map<oid_t, std::vector<oid_t>> visible_tuples;
// for every tuple that is found in the index.
for (auto tuple_location_ptr : tuple_location_ptrs) {
ItemPointer tuple_location = *tuple_location_ptr;
if (type_ == HYBRID_SCAN_TYPE_HYBRID &&
tuple_location.block >= (block_threshold)) {
item_pointers_.insert(tuple_location);
}
auto &manager = catalog::Manager::GetInstance();
auto tile_group = manager.GetTileGroup(tuple_location.block);
auto tile_group_header = tile_group.get()->GetHeader();
// perform transaction read
size_t chain_length = 0;
while (true) {
++chain_length;
auto visibility = transaction_manager.IsVisible(current_txn, tile_group_header, tuple_location.offset);
if (visibility == VISIBILITY_OK) {
visible_tuples[tuple_location.block].push_back(tuple_location.offset);
auto res = transaction_manager.PerformRead(current_txn, tuple_location, acquire_owner);
if (!res) {
transaction_manager.SetTransactionResult(current_txn, RESULT_FAILURE);
return res;
}
break;
} else {
ItemPointer old_item = tuple_location;
cid_t old_end_cid = tile_group_header->GetEndCommitId(old_item.offset);
tuple_location = tile_group_header->GetNextItemPointer(old_item.offset);
// there must exist a visible version.
assert(tuple_location.IsNull() == false);
cid_t max_committed_cid = transaction_manager.GetMaxCommittedCid();
// check whether older version is garbage.
if (old_end_cid < max_committed_cid) {
assert(tile_group_header->GetTransactionId(old_item.offset) ==
INITIAL_TXN_ID ||
tile_group_header->GetTransactionId(old_item.offset) ==
INVALID_TXN_ID);
if (tile_group_header->SetAtomicTransactionId(
old_item.offset, INVALID_TXN_ID) == true) {
// atomically swap item pointer held in the index bucket.
AtomicUpdateItemPointer(tuple_location_ptr, tuple_location);
}
}
tile_group = manager.GetTileGroup(tuple_location.block);
tile_group_header = tile_group.get()->GetHeader();
}
}
}
// Construct a logical tile for each block
for (auto tuples : visible_tuples) {
auto &manager = catalog::Manager::GetInstance();
auto tile_group = manager.GetTileGroup(tuples.first);
std::unique_ptr<LogicalTile> logical_tile(LogicalTileFactory::GetTile());
// Add relevant columns to logical tile
logical_tile->AddColumns(tile_group, full_column_ids_);
logical_tile->AddPositionList(std::move(tuples.second));
if (column_ids_.size() != 0) {
logical_tile->ProjectColumns(full_column_ids_, column_ids_);
}
result_.push_back(logical_tile.release());
}
index_done_ = true;
LOG_TRACE("Result tiles : %lu", result_.size());
return true;
}