private List<long> DoPersist(TreeWrite sequence, int tryCount)
{
// NOTE: nodes are written in order of branches and then leaf nodes. All
// branch nodes and leafs are grouped together.
// The list of nodes to be allocated,
IList<ITreeNode> allBranches = sequence.BranchNodes;
IList<ITreeNode> allLeafs = sequence.LeafNodes;
List<ITreeNode> nodes = new List<ITreeNode>(allBranches.Count + allLeafs.Count);
nodes.AddRange(allBranches);
nodes.AddRange(allLeafs);
int sz = nodes.Count;
// The list of allocated referenced for the nodes,
DataAddress[] refs = new DataAddress[sz];
long[] outRefs = new long[sz];
MessageStream allocateMessage = new MessageStream(MessageType.Request);
// Make a connection with the manager server,
IMessageProcessor manager = connector.Connect(managerAddress, ServiceType.Manager);
// Allocate the space first,
for (int i = 0; i < sz; ++i) {
ITreeNode node = nodes[i];
RequestMessage request = new RequestMessage("allocateNode");
// Is it a branch node?
if (node is TreeBranch) {
// Branch nodes are 1K in size,
request.Arguments.Add(1024);
} else {
// Leaf nodes are 4k in size,
request.Arguments.Add(4096);
}
allocateMessage.AddMessage(request);
}
// The result of the set of allocations,
MessageStream resultStream = (MessageStream) manager.Process(allocateMessage);
//DEBUG: ++network_comm_count;
// The unique list of blocks,
List<long> uniqueBlocks = new List<long>();
// Parse the result stream one message at a time, the order will be the
// order of the allocation messages,
int n = 0;
foreach (ResponseMessage m in resultStream) {
if (m.HasError)
throw m.Error.AsException();
DataAddress addr = (DataAddress) m.Arguments[0].Value;
refs[n] = addr;
// Make a list of unique block identifiers,
if (!uniqueBlocks.Contains(addr.BlockId)) {
uniqueBlocks.Add(addr.BlockId);
}
++n;
}
// Get the block to server map for each of the blocks,
IDictionary<long, IList<BlockServerElement>> blockToServerMap = GetServersForBlock(uniqueBlocks);
// Make message streams for each unique block
int ubid_count = uniqueBlocks.Count;
MessageStream[] ubidStream = new MessageStream[ubid_count];
for (int i = 0; i < ubidStream.Length; ++i) {
ubidStream[i] = new MessageStream(MessageType.Request);
}
// Scan all the blocks and create the message streams,
for (int i = 0; i < sz; ++i) {
byte[] nodeBuf;
ITreeNode node = nodes[i];
// Is it a branch node?
if (node is TreeBranch) {
TreeBranch branch = (TreeBranch)node;
// Make a copy of the branch (NOTE; we Clone() the array here).
long[] curNodeData = (long[])branch.ChildPointers.Clone();
int curNdsz = branch.DataSize;
branch = new TreeBranch(refs[i].Value, curNodeData, curNdsz);
// The number of children
int chsz = branch.ChildCount;
// For each child, if it's a heap node, look up the child id and
// reference map in the sequence and set the reference accordingly,
for (int o = 0; o < chsz; ++o) {
long childRef = branch.GetChild(o);
if (childRef < 0) {
// The ref is currently on the heap, so adjust accordingly
int ref_id = sequence.LookupRef(i, o);
branch.SetChildOverride(o, refs[ref_id].Value);
}
}
// Turn the branch into a 'node_buf' byte[] array object for
// serialization.
long[] nodeData = branch.ChildPointers;
int ndsz = branch.DataSize;
MemoryStream bout = new MemoryStream(1024);
BinaryWriter dout = new BinaryWriter(bout, Encoding.Unicode);
dout.Write(BranchType);
dout.Write(ndsz);
for (int o = 0; o < ndsz; ++o) {
dout.Write(nodeData[o]);
}
dout.Flush();
// Turn it into a byte array,
nodeBuf = bout.ToArray();
// Put this branch into the local cache,
networkCache.SetNode(refs[i], branch);
} else {
// If it's a leaf node,
TreeLeaf leaf = (TreeLeaf)node;
int lfsz = leaf.Length;
nodeBuf = new byte[lfsz + 6];
// Technically, we could comment these next two lines out.
ByteBuffer.WriteInt2(LeafType, nodeBuf, 0);
ByteBuffer.WriteInt4(lfsz, nodeBuf, 2);
leaf.Read(0, nodeBuf, 6, lfsz);
// Put this leaf into the local cache,
leaf = new ByteArrayTreeLeaf(refs[i].Value, nodeBuf);
networkCache.SetNode(refs[i], leaf);
}
// The DataAddress this node is being written to,
DataAddress address = refs[i];
// Get the block id,
long blockId = address.BlockId;
int bid = uniqueBlocks.IndexOf(blockId);
RequestMessage request = new RequestMessage("writeToBlock");
request.Arguments.Add(address);
request.Arguments.Add(nodeBuf);
request.Arguments.Add(0);
request.Arguments.Add(nodeBuf.Length);
ubidStream[bid].AddMessage(request);
// Update 'outRefs' array,
outRefs[i] = refs[i].Value;
}
// A log of successfully processed operations,
List<object> successProcess = new List<object>(64);
// Now process the streams on the servers,
for (int i = 0; i < ubidStream.Length; ++i) {
// The output message,
MessageStream requestMessageStream = ubidStream[i];
// Get the servers this message needs to be sent to,
long block_id = uniqueBlocks[i];
IList<BlockServerElement> blockServers = blockToServerMap[block_id];
// Format a message for writing this node out,
int bssz = blockServers.Count;
IMessageProcessor[] blockServerProcs = new IMessageProcessor[bssz];
// Make the block server connections,
for (int o = 0; o < bssz; ++o) {
IServiceAddress address = blockServers[o].Address;
blockServerProcs[o] = connector.Connect(address, ServiceType.Block);
MessageStream responseMessageStream = (MessageStream) blockServerProcs[o].Process(requestMessageStream);
//DEBUG: ++network_comm_count;
if (responseMessageStream.HasError) {
// If this is an error, we need to report the failure to the
// manager server,
ReportBlockServerFailure(address);
// Remove the block id from the server list cache,
networkCache.RemoveServers(block_id);
// Rollback any server writes already successfully made,
for (int p = 0; p < successProcess.Count; p += 2) {
IServiceAddress blockAddress = (IServiceAddress) successProcess[p];
MessageStream toRollback = (MessageStream) successProcess[p + 1];
List<DataAddress> rollbackNodes = new List<DataAddress>(128);
foreach(Message rm in toRollback) {
DataAddress raddr = (DataAddress) rm.Arguments[0].Value;
rollbackNodes.Add(raddr);
}
// Create the rollback message,
RequestMessage rollbackRequest = new RequestMessage("rollbackNodes");
rollbackRequest.Arguments.Add(rollbackNodes.ToArray());
// Send it to the block server,
Message responseMessage = connector.Connect(blockAddress, ServiceType.Block).Process(rollbackRequest);
//DEBUG: ++network_comm_count;
// If rollback generated an error we throw the error now
// because this likely is a serious network error.
if (responseMessage.HasError)
throw new NetworkException("Rollback wrote failed: " + responseMessage.ErrorMessage);
}
// Retry,
if (tryCount > 0)
return DoPersist(sequence, tryCount - 1);
// Otherwise we fail the write
throw new NetworkException(responseMessageStream.ErrorMessage);
}
// If we succeeded without an error, add to the log
successProcess.Add(address);
successProcess.Add(requestMessageStream);
}
}
// Return the references,
return new List<long>(outRefs);
}
public IList<ITreeNode> FetchNodes(long[] nids)
{
// The number of nodes,
int node_count = nids.Length;
// The array of read nodes,
ITreeNode[] result_nodes = new ITreeNode[node_count];
// Resolve special nodes first,
{
int i = 0;
foreach (long nodeId in nids) {
if ((nodeId & 0x01000000000000000L) != 0)
result_nodes[i] = SparseLeafNode.Create(nodeId);
++i;
}
}
// Group all the nodes to the same block,
List<long> uniqueBlocks = new List<long>();
List<List<long>> uniqueBlockList = new List<List<long>>();
{
int i = 0;
foreach (long node_ref in nids) {
// If it's not a special node,
if ((node_ref & 0x01000000000000000L) == 0) {
// Get the block id and add it to the list of unique blocks,
DataAddress address = new DataAddress(node_ref);
// Check if the node is in the local cache,
ITreeNode node = networkCache.GetNode(address);
if (node != null) {
result_nodes[i] = node;
} else {
// Not in the local cache so we need to bundle this up in a node
// request on the block servers,
// Group this node request by the block identifier
long blockId = address.BlockId;
int ind = uniqueBlocks.IndexOf(blockId);
if (ind == -1) {
ind = uniqueBlocks.Count;
uniqueBlocks.Add(blockId);
uniqueBlockList.Add(new List<long>());
}
List<long> blist = uniqueBlockList[ind];
blist.Add(node_ref);
}
}
++i;
}
}
// Exit early if no blocks,
if (uniqueBlocks.Count == 0)
return result_nodes;
// Resolve server records for the given block identifiers,
IDictionary<long, IList<BlockServerElement>> servers_map = GetServersForBlock(uniqueBlocks);
// The result nodes list,
List<ITreeNode> nodes = new List<ITreeNode>();
// For each unique block list,
foreach (List<long> blist in uniqueBlockList) {
// Make a block server request for each node in the block,
MessageStream block_server_msg = new MessageStream(MessageType.Request);
long block_id = -1;
foreach (long node_ref in blist) {
DataAddress address = new DataAddress(node_ref);
RequestMessage request = new RequestMessage("readFromBlock");
request.Arguments.Add(address);
block_server_msg.AddMessage(request);
block_id = address.BlockId;
}
if (block_id == -1)
throw new ApplicationException("block_id == -1");
// Get the shuffled list of servers the block is stored on,
IList<BlockServerElement> servers = servers_map[block_id];
// Go through the servers one at a time to fetch the block,
bool success = false;
for (int z = 0; z < servers.Count && !success; ++z) {
BlockServerElement server = servers[z];
// If the server is up,
if (server.IsStatusUp) {
// Open a connection with the block server,
IMessageProcessor block_server_proc = connector.Connect(server.Address, ServiceType.Block);
MessageStream message_in = (MessageStream) block_server_proc.Process(block_server_msg);
// DEBUG: ++networkCommCount;
// DEBUG: ++networkFetchCommCount;
bool is_error = false;
bool severe_error = false;
// Turn each none-error message into a node
foreach (ResponseMessage m in message_in) {
if (m.HasError) {
// See if this error is a block read error. If it is, we don't
// tell the manager server to lock this server out completely.
bool is_block_read_error = m.Error.Source.Equals("Deveel.Data.Net.BlockReadException");
if (!is_block_read_error) {
// If it's something other than a block read error, we mark
// this error as severe,
severe_error = true;
}
is_error = true;
} else if (!is_error) {
// The reply contains the block of data read.
NodeSet node_set = (NodeSet)m.Arguments[0].Value;
// Decode the node items into node objects,
IEnumerator<Node> item_iterator = node_set.GetEnumerator();
while (item_iterator.MoveNext()) {
// Get the node item,
Node node_item = item_iterator.Current;
long node_ref = node_item.Id;
DataAddress address = new DataAddress(node_ref);
// Wrap around a buffered DataInputStream for reading values
// from the store.
BinaryReader input = new BinaryReader(node_item.Input, Encoding.Unicode);
short node_type = input.ReadInt16();
ITreeNode read_node;
// Is the node type a leaf node?
if (node_type == LeafType) {
// Read the key
int leaf_size = input.ReadInt32();
byte[] buf = ReadNodeAsBuffer(node_item);
if (buf == null) {
buf = new byte[leaf_size + 6];
input.Read(buf, 6, leaf_size);
// Technically, we could comment these next two lines out.
ByteBuffer.WriteInt2(node_type, buf, 0);
ByteBuffer.WriteInt4(leaf_size, buf, 2);
}
// Create a leaf that's mapped to this data
read_node = new ByteArrayTreeLeaf(node_ref, buf); ;
}
// Is the node type a branch node?
else if (node_type == BranchType) {
// Note that the entire branch is loaded into memory,
int child_data_size = input.ReadInt32();
long[] data_arr = new long[child_data_size];
for (int n = 0; n < child_data_size; ++n) {
data_arr[n] = input.ReadInt64();
}
// Create the branch node,
read_node = new TreeBranch(node_ref, data_arr, child_data_size);
} else {
throw new InvalidDataState("Unknown node type: " + node_type, address);
}
// Is the node already in the list? If so we don't add it.
if (!IsInNodeList(node_ref, nodes)) {
// Put the read node in the cache and add it to the 'nodes'
// list.
networkCache.SetNode(address, read_node);
nodes.Add(read_node);
}
}
}
}
// If there was no error while reading the result, we assume the node
// requests were successfully read.
if (is_error == false) {
success = true;
} else {
if (severe_error) {
// If this is an error, we need to report the failure to the
// manager server,
ReportBlockServerFailure(server.Address);
// Remove the block id from the server list cache,
networkCache.RemoveServers(block_id);
} else {
// Otherwise, not a severe error (probably a corrupt block on a
// server), so shuffle the server list for this block_id so next
// time there's less chance of hitting this bad block.
IList<BlockServerElement> srvs = networkCache.GetServers(block_id);
List<BlockServerElement> server_list = new List<BlockServerElement>();
server_list.AddRange(srvs);
CollectionsUtil.Shuffle(server_list);
networkCache.SetServers(block_id, server_list, 15 * 60 * 1000);
}
}
}
}
// If the nodes were not successfully read, we generate an exception,
if (!success) {
// Remove from the cache,
networkCache.RemoveServers(block_id);
throw new ApplicationException("Unable to fetch node from block server");
}
}
int sz = nodes.Count;
if (sz == 0)
throw new ApplicationException("Empty nodes list");
for (int i = 0; i < sz; ++i) {
ITreeNode node = nodes[i];
long node_ref = node.Id;
for (int n = 0; n < nids.Length; ++n) {
if (nids[n] == node_ref)
result_nodes[n] = node;
}
}
// Check the result_nodes list is completely populated,
for (int n = 0; n < result_nodes.Length; ++n) {
if (result_nodes[n] == null)
throw new ApplicationException("Assertion failed: result_nodes not completely populated.");
}
return result_nodes;
}
