/*
* master_copy_shard_placement implements a user-facing UDF to copy data from
* a healthy (source) node to an inactive (target) node. To accomplish this it
* entirely recreates the table structure before copying all data. During this
* time all modifications are paused to the shard. After successful repair, the
* inactive placement is marked healthy and modifications may continue. If the
* repair fails at any point, this function throws an error, leaving the node
* in an unhealthy state.
*/
Datum
master_copy_shard_placement(PG_FUNCTION_ARGS)
{
int64 shardId = PG_GETARG_INT64(0);
text *sourceNodeName = PG_GETARG_TEXT_P(1);
int32 sourceNodePort = PG_GETARG_INT32(2);
text *targetNodeName = PG_GETARG_TEXT_P(3);
int32 targetNodePort = PG_GETARG_INT32(4);
ShardInterval *shardInterval = LoadShardInterval(shardId);
Oid distributedTableId = shardInterval->relationId;
List *shardPlacementList = NIL;
ShardPlacement *sourcePlacement = NULL;
ShardPlacement *targetPlacement = NULL;
List *ddlCommandList = NIL;
bool recreated = false;
bool dataCopied = false;
/*
* By taking an exclusive lock on the shard, we both stop all modifications
* (INSERT, UPDATE, or DELETE) and prevent concurrent repair operations from
* being able to operate on this shard.
*/
LockShard(shardId, ExclusiveLock);
shardPlacementList = LoadShardPlacementList(shardId);
sourcePlacement = SearchShardPlacementInList(shardPlacementList, sourceNodeName,
sourceNodePort);
if (sourcePlacement->shardState != STATE_FINALIZED)
{
ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("source placement must be in finalized state")));
}
targetPlacement = SearchShardPlacementInList(shardPlacementList, targetNodeName,
targetNodePort);
if (targetPlacement->shardState != STATE_INACTIVE)
{
ereport(ERROR, (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("target placement must be in inactive state")));
}
/* retrieve the DDL commands for the table and run them */
ddlCommandList = RecreateTableDDLCommandList(distributedTableId, shardId);
recreated = ExecuteRemoteCommandList(targetPlacement->nodeName,
targetPlacement->nodePort,
ddlCommandList);
if (!recreated)
{
ereport(ERROR, (errmsg("could not recreate shard table"),
errhint("Consult recent messages in the server logs for "
"details.")));
}
HOLD_INTERRUPTS();
dataCopied = CopyDataFromFinalizedPlacement(distributedTableId, shardId,
sourcePlacement, targetPlacement);
if (!dataCopied)
{
ereport(ERROR, (errmsg("could not copy shard data"),
errhint("Consult recent messages in the server logs for "
"details.")));
}
/* the placement is repaired, so return to finalized state */
DeleteShardPlacementRow(targetPlacement->id);
InsertShardPlacementRow(targetPlacement->id, targetPlacement->shardId,
STATE_FINALIZED, targetPlacement->nodeName,
targetPlacement->nodePort);
RESUME_INTERRUPTS();
PG_RETURN_VOID();
}
/*
* DropShards drops all given shards in a relation. The id, name and schema
* for the relation are explicitly provided, since this function may be
* called when the table is already dropped.
*
* We mark shard placements that we couldn't drop as to be deleted later, but
* we do delete the shard metadadata.
*/
static int
DropShards(Oid relationId, char *schemaName, char *relationName,
List *deletableShardIntervalList)
{
ListCell *shardIntervalCell = NULL;
int droppedShardCount = 0;
BeginOrContinueCoordinatedTransaction();
/* At this point we intentionally decided to not use 2PC for reference tables */
if (MultiShardCommitProtocol == COMMIT_PROTOCOL_2PC)
{
CoordinatedTransactionUse2PC();
}
foreach(shardIntervalCell, deletableShardIntervalList)
{
List *shardPlacementList = NIL;
ListCell *shardPlacementCell = NULL;
ShardInterval *shardInterval = (ShardInterval *) lfirst(shardIntervalCell);
uint64 shardId = shardInterval->shardId;
char *quotedShardName = NULL;
char *shardRelationName = pstrdup(relationName);
Assert(shardInterval->relationId == relationId);
/* Build shard relation name. */
AppendShardIdToName(&shardRelationName, shardId);
quotedShardName = quote_qualified_identifier(schemaName, shardRelationName);
shardPlacementList = ShardPlacementList(shardId);
foreach(shardPlacementCell, shardPlacementList)
{
ShardPlacement *shardPlacement =
(ShardPlacement *) lfirst(shardPlacementCell);
char *workerName = shardPlacement->nodeName;
uint32 workerPort = shardPlacement->nodePort;
StringInfo workerDropQuery = makeStringInfo();
MultiConnection *connection = NULL;
uint32 connectionFlags = FOR_DDL;
char storageType = shardInterval->storageType;
if (storageType == SHARD_STORAGE_TABLE)
{
appendStringInfo(workerDropQuery, DROP_REGULAR_TABLE_COMMAND,
quotedShardName);
}
else if (storageType == SHARD_STORAGE_COLUMNAR ||
storageType == SHARD_STORAGE_FOREIGN)
{
appendStringInfo(workerDropQuery, DROP_FOREIGN_TABLE_COMMAND,
quotedShardName);
}
connection = GetPlacementConnection(connectionFlags, shardPlacement, NULL);
RemoteTransactionBeginIfNecessary(connection);
if (PQstatus(connection->pgConn) != CONNECTION_OK)
{
uint64 placementId = shardPlacement->placementId;
ereport(WARNING, (errmsg("could not connect to shard \"%s\" on node "
"\"%s:%u\"", shardRelationName, workerName,
workerPort),
errdetail("Marking this shard placement for "
"deletion")));
UpdateShardPlacementState(placementId, FILE_TO_DELETE);
continue;
}
MarkRemoteTransactionCritical(connection);
ExecuteCriticalRemoteCommand(connection, workerDropQuery->data);
DeleteShardPlacementRow(shardPlacement->placementId);
}
