/*
* lock_shard_metadata allows the shard distribution metadata to be locked
* remotely to block concurrent writes from workers in MX tables.
*
* This function does not sort the array to avoid deadlock, callers
* must ensure a consistent order.
*/
Datum
lock_shard_metadata(PG_FUNCTION_ARGS)
{
LOCKMODE lockMode = IntToLockMode(PG_GETARG_INT32(0));
ArrayType *shardIdArrayObject = PG_GETARG_ARRAYTYPE_P(1);
Datum *shardIdArrayDatum = NULL;
int shardIdCount = 0;
int shardIdIndex = 0;
if (ARR_NDIM(shardIdArrayObject) == 0)
{
ereport(ERROR, (errmsg("no locks specified")));
}
/* we don't want random users to block writes */
EnsureSuperUser();
shardIdCount = ArrayObjectCount(shardIdArrayObject);
shardIdArrayDatum = DeconstructArrayObject(shardIdArrayObject);
for (shardIdIndex = 0; shardIdIndex < shardIdCount; shardIdIndex++)
{
int64 shardId = DatumGetInt64(shardIdArrayDatum[shardIdIndex]);
LockShardDistributionMetadata(shardId, lockMode);
}
PG_RETURN_VOID();
}
/*
* LockShardListMetadata takes shared locks on the metadata of all shards in
* shardIntervalList to prevents concurrent placement changes.
*/
void
LockShardListMetadata(List *shardIntervalList, LOCKMODE lockMode)
{
ListCell *shardIntervalCell = NULL;
/* lock shards in order of shard id to prevent deadlock */
shardIntervalList = SortList(shardIntervalList, CompareShardIntervalsById);
foreach(shardIntervalCell, shardIntervalList)
{
ShardInterval *shardInterval = (ShardInterval *) lfirst(shardIntervalCell);
int64 shardId = shardInterval->shardId;
LockShardDistributionMetadata(shardId, lockMode);
}
/*
* LockShards takes shared locks on the metadata and the data of all shards in
* shardIntervalList. This prevents concurrent placement changes and concurrent
* DML statements that require an exclusive lock.
*/
void
LockShards(List *shardIntervalList, LOCKMODE lockMode)
{
ListCell *shardIntervalCell = NULL;
/* lock shards in order of shard id to prevent deadlock */
shardIntervalList = SortList(shardIntervalList, CompareShardIntervalsById);
foreach(shardIntervalCell, shardIntervalList)
{
ShardInterval *shardInterval = (ShardInterval *) lfirst(shardIntervalCell);
int64 shardId = shardInterval->shardId;
/* prevent concurrent changes to number of placements */
LockShardDistributionMetadata(shardId, lockMode);
/* prevent concurrent update/delete statements */
LockShardResource(shardId, lockMode);
}
/*
* CreateShardsWithRoundRobinPolicy creates empty shards for the given table
* based on the specified number of initial shards. The function first updates
* metadata on the coordinator node to make this shard (and its placements)
* visible. Note that the function assumes the table is hash partitioned and
* calculates the min/max hash token ranges for each shard, giving them an equal
* split of the hash space. Finally, function creates empty shard placements on
* worker nodes.
*/
void
CreateShardsWithRoundRobinPolicy(Oid distributedTableId, int32 shardCount,
int32 replicationFactor, bool useExclusiveConnections)
{
char shardStorageType = 0;
List *workerNodeList = NIL;
int32 workerNodeCount = 0;
uint32 placementAttemptCount = 0;
uint64 hashTokenIncrement = 0;
List *existingShardList = NIL;
int64 shardIndex = 0;
DistTableCacheEntry *cacheEntry = DistributedTableCacheEntry(distributedTableId);
bool colocatedShard = false;
List *insertedShardPlacements = NIL;
/* make sure table is hash partitioned */
CheckHashPartitionedTable(distributedTableId);
/*
* In contrast to append/range partitioned tables it makes more sense to
* require ownership privileges - shards for hash-partitioned tables are
* only created once, not continually during ingest as for the other
* partitioning types.
*/
EnsureTableOwner(distributedTableId);
/* we plan to add shards: get an exclusive lock on relation oid */
LockRelationOid(distributedTableId, ExclusiveLock);
/* validate that shards haven't already been created for this table */
existingShardList = LoadShardList(distributedTableId);
if (existingShardList != NIL)
{
char *tableName = get_rel_name(distributedTableId);
ereport(ERROR, (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
errmsg("table \"%s\" has already had shards created for it",
tableName)));
}
/* make sure that at least one shard is specified */
if (shardCount <= 0)
{
ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("shard_count must be positive")));
}
/* make sure that at least one replica is specified */
if (replicationFactor <= 0)
{
ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("replication_factor must be positive")));
}
/* make sure that RF=1 if the table is streaming replicated */
if (cacheEntry->replicationModel == REPLICATION_MODEL_STREAMING &&
replicationFactor > 1)
{
char *relationName = get_rel_name(cacheEntry->relationId);
ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("using replication factor %d with the streaming "
"replication model is not supported",
replicationFactor),
errdetail("The table %s is marked as streaming replicated and "
"the shard replication factor of streaming replicated "
"tables must be 1.", relationName),
errhint("Use replication factor 1.")));
}
/* calculate the split of the hash space */
hashTokenIncrement = HASH_TOKEN_COUNT / shardCount;
/* don't allow concurrent node list changes that require an exclusive lock */
LockRelationOid(DistNodeRelationId(), RowShareLock);
/* load and sort the worker node list for deterministic placement */
workerNodeList = ActivePrimaryNodeList();
workerNodeList = SortList(workerNodeList, CompareWorkerNodes);
/* make sure we don't process cancel signals until all shards are created */
HOLD_INTERRUPTS();
workerNodeCount = list_length(workerNodeList);
if (replicationFactor > workerNodeCount)
{
ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("replication_factor (%d) exceeds number of worker nodes "
"(%d)", replicationFactor, workerNodeCount),
errhint("Add more worker nodes or try again with a lower "
"replication factor.")));
}
/* if we have enough nodes, add an extra placement attempt for backup */
placementAttemptCount = (uint32) replicationFactor;
if (workerNodeCount > replicationFactor)
{
placementAttemptCount++;
}
/* set shard storage type according to relation type */
shardStorageType = ShardStorageType(distributedTableId);
for (shardIndex = 0; shardIndex < shardCount; shardIndex++)
{
uint32 roundRobinNodeIndex = shardIndex % workerNodeCount;
/* initialize the hash token space for this shard */
text *minHashTokenText = NULL;
text *maxHashTokenText = NULL;
int32 shardMinHashToken = INT32_MIN + (shardIndex * hashTokenIncrement);
int32 shardMaxHashToken = shardMinHashToken + (hashTokenIncrement - 1);
uint64 shardId = GetNextShardId();
List *currentInsertedShardPlacements = NIL;
/* if we are at the last shard, make sure the max token value is INT_MAX */
if (shardIndex == (shardCount - 1))
{
shardMaxHashToken = INT32_MAX;
}
/* insert the shard metadata row along with its min/max values */
minHashTokenText = IntegerToText(shardMinHashToken);
maxHashTokenText = IntegerToText(shardMaxHashToken);
/*
* Grabbing the shard metadata lock isn't technically necessary since
* we already hold an exclusive lock on the partition table, but we'll
* acquire it for the sake of completeness. As we're adding new active
* placements, the mode must be exclusive.
*/
LockShardDistributionMetadata(shardId, ExclusiveLock);
InsertShardRow(distributedTableId, shardId, shardStorageType,
minHashTokenText, maxHashTokenText);
currentInsertedShardPlacements = InsertShardPlacementRows(distributedTableId,
shardId,
workerNodeList,
roundRobinNodeIndex,
replicationFactor);
insertedShardPlacements = list_concat(insertedShardPlacements,
currentInsertedShardPlacements);
}
CreateShardsOnWorkers(distributedTableId, insertedShardPlacements,
useExclusiveConnections, colocatedShard);
if (QueryCancelPending)
{
ereport(WARNING, (errmsg("cancel requests are ignored during shard creation")));
QueryCancelPending = false;
}
RESUME_INTERRUPTS();
}
/*
* master_create_worker_shards creates empty shards for the given table based
* on the specified number of initial shards. The function first gets a list of
* candidate nodes and issues DDL commands on the nodes to create empty shard
* placements on those nodes. The function then updates metadata on the master
* node to make this shard (and its placements) visible. Note that the function
* assumes the table is hash partitioned and calculates the min/max hash token
* ranges for each shard, giving them an equal split of the hash space.
*/
Datum
master_create_worker_shards(PG_FUNCTION_ARGS)
{
text *tableNameText = PG_GETARG_TEXT_P(0);
int32 shardCount = PG_GETARG_INT32(1);
int32 replicationFactor = PG_GETARG_INT32(2);
Oid distributedTableId = ResolveRelationId(tableNameText);
char relationKind = get_rel_relkind(distributedTableId);
char *tableName = text_to_cstring(tableNameText);
char *relationOwner = NULL;
char shardStorageType = '\0';
List *workerNodeList = NIL;
List *ddlCommandList = NIL;
int32 workerNodeCount = 0;
uint32 placementAttemptCount = 0;
uint64 hashTokenIncrement = 0;
List *existingShardList = NIL;
int64 shardIndex = 0;
/* make sure table is hash partitioned */
CheckHashPartitionedTable(distributedTableId);
/*
* In contrast to append/range partitioned tables it makes more sense to
* require ownership privileges - shards for hash-partitioned tables are
* only created once, not continually during ingest as for the other
* partitioning types.
*/
EnsureTableOwner(distributedTableId);
/* we plan to add shards: get an exclusive metadata lock */
LockRelationDistributionMetadata(distributedTableId, ExclusiveLock);
relationOwner = TableOwner(distributedTableId);
/* validate that shards haven't already been created for this table */
existingShardList = LoadShardList(distributedTableId);
if (existingShardList != NIL)
{
ereport(ERROR, (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
errmsg("table \"%s\" has already had shards created for it",
tableName)));
}
/* make sure that at least one shard is specified */
if (shardCount <= 0)
{
ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("shard_count must be positive")));
}
/* make sure that at least one replica is specified */
if (replicationFactor <= 0)
{
ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("replication_factor must be positive")));
}
/* calculate the split of the hash space */
hashTokenIncrement = HASH_TOKEN_COUNT / shardCount;
/* load and sort the worker node list for deterministic placement */
workerNodeList = WorkerNodeList();
workerNodeList = SortList(workerNodeList, CompareWorkerNodes);
/* make sure we don't process cancel signals until all shards are created */
HOLD_INTERRUPTS();
/* retrieve the DDL commands for the table */
ddlCommandList = GetTableDDLEvents(distributedTableId);
workerNodeCount = list_length(workerNodeList);
if (replicationFactor > workerNodeCount)
{
ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("replication_factor (%d) exceeds number of worker nodes "
"(%d)", replicationFactor, workerNodeCount),
errhint("Add more worker nodes or try again with a lower "
"replication factor.")));
}
/* if we have enough nodes, add an extra placement attempt for backup */
placementAttemptCount = (uint32) replicationFactor;
if (workerNodeCount > replicationFactor)
{
placementAttemptCount++;
}
/* set shard storage type according to relation type */
if (relationKind == RELKIND_FOREIGN_TABLE)
{
bool cstoreTable = CStoreTable(distributedTableId);
if (cstoreTable)
{
shardStorageType = SHARD_STORAGE_COLUMNAR;
}
else
{
shardStorageType = SHARD_STORAGE_FOREIGN;
}
}
else
{
shardStorageType = SHARD_STORAGE_TABLE;
}
for (shardIndex = 0; shardIndex < shardCount; shardIndex++)
{
uint32 roundRobinNodeIndex = shardIndex % workerNodeCount;
/* initialize the hash token space for this shard */
text *minHashTokenText = NULL;
text *maxHashTokenText = NULL;
int32 shardMinHashToken = INT32_MIN + (shardIndex * hashTokenIncrement);
int32 shardMaxHashToken = shardMinHashToken + (hashTokenIncrement - 1);
Datum shardIdDatum = master_get_new_shardid(NULL);
int64 shardId = DatumGetInt64(shardIdDatum);
/* if we are at the last shard, make sure the max token value is INT_MAX */
if (shardIndex == (shardCount - 1))
{
shardMaxHashToken = INT32_MAX;
}
/* insert the shard metadata row along with its min/max values */
minHashTokenText = IntegerToText(shardMinHashToken);
maxHashTokenText = IntegerToText(shardMaxHashToken);
/*
* Grabbing the shard metadata lock isn't technically necessary since
* we already hold an exclusive lock on the partition table, but we'll
* acquire it for the sake of completeness. As we're adding new active
* placements, the mode must be exclusive.
*/
LockShardDistributionMetadata(shardId, ExclusiveLock);
CreateShardPlacements(shardId, ddlCommandList, relationOwner, workerNodeList,
roundRobinNodeIndex, replicationFactor);
InsertShardRow(distributedTableId, shardId, shardStorageType,
minHashTokenText, maxHashTokenText);
}
if (QueryCancelPending)
{
ereport(WARNING, (errmsg("cancel requests are ignored during shard creation")));
QueryCancelPending = false;
}
RESUME_INTERRUPTS();
PG_RETURN_VOID();
}