// RunWithResponse is a version of Run that returns the BatchResponse.
func (txn *Txn) RunWithResponse(b *Batch) (*roachpb.BatchResponse, *roachpb.Error) {
tracer.AnnotateTrace()
defer tracer.AnnotateTrace()
if pErr := b.prepare(); pErr != nil {
return nil, pErr
}
return sendAndFill(txn.send, b)
}
func (n *scanNode) Next() bool {
tracer.AnnotateTrace()
if n.pErr != nil {
return false
}
if n.kvs == nil {
if !n.initScan() {
return false
}
}
// All of the columns for a particular row will be grouped together. We loop
// over the key/value pairs and decode the key to extract the columns encoded
// within the key and the column ID. We use the column ID to lookup the
// column and decode the value. All of these values go into a map keyed by
// column name. When the index key changes we output a row containing the
// current values.
for {
if n.maybeOutputRow() {
return n.pErr == nil
}
if n.kvIndex == len(n.kvs) {
return false
}
if !n.processKV(n.kvs[n.kvIndex]) {
return false
}
n.kvIndex++
}
}
func (c *v3Conn) executeStatements(stmts string, params []parser.Datum, formatCodes []formatCode, sendDescription bool) error {
tracer.AnnotateTrace()
c.session.Database = c.opts.database
// TODO(dt): this is a clumsy check better left to the actual parser. #3852
if len(strings.TrimSpace(stmts)) == 0 {
// Skip executor and just send EmptyQueryResponse.
c.writeBuf.initMsg(serverMsgEmptyQuery)
return c.writeBuf.finishMsg(c.wr)
}
resp, _, err := c.executor.ExecuteStatements(c.opts.user, c.session, stmts, params)
if err != nil {
return c.sendError(err.Error())
}
c.session.Reset()
if err := c.session.Unmarshal(resp.Session); err != nil {
return err
}
c.opts.database = c.session.Database
return c.sendResponse(resp, formatCodes, sendDescription)
}
// send runs the specified calls synchronously in a single batch and returns
// any errors. Returns a nil response for empty input (no requests).
func (db *DB) send(reqs ...roachpb.Request) (*roachpb.BatchResponse, *roachpb.Error) {
if len(reqs) == 0 {
return nil, nil
}
ba := roachpb.BatchRequest{}
ba.Add(reqs...)
if ba.UserPriority == 0 && db.userPriority != 1 {
ba.UserPriority = db.userPriority
}
tracer.AnnotateTrace()
br, pErr := db.sender.Send(context.TODO(), ba)
if pErr != nil {
if log.V(1) {
log.Infof("failed batch: %s", pErr)
}
return nil, pErr
}
return br, nil
}
// Send implements the batch.Sender interface. It subdivides
// the Batch into batches admissible for sending (preventing certain
// illegal mixtures of requests), executes each individual part
// (which may span multiple ranges), and recombines the response.
// When the request spans ranges, it is split up and the corresponding
// ranges queried serially, in ascending order.
// In particular, the first write in a transaction may not be part of the first
// request sent. This is relevant since the first write is a BeginTransaction
// request, thus opening up a window of time during which there may be intents
// of a transaction, but no entry. Pushing such a transaction will succeed, and
// may lead to the transaction being aborted early.
func (ds *DistSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
tracer.AnnotateTrace()
// In the event that timestamp isn't set and read consistency isn't
// required, set the timestamp using the local clock.
if ba.ReadConsistency == roachpb.INCONSISTENT && ba.Timestamp.Equal(roachpb.ZeroTimestamp) {
ba.Timestamp = ds.clock.Now()
}
if ba.Txn != nil && len(ba.Txn.CertainNodes.Nodes) == 0 {
// Ensure the local NodeID is marked as free from clock offset;
// the transaction's timestamp was taken off the local clock.
if nDesc := ds.getNodeDescriptor(); nDesc != nil {
// TODO(tschottdorf): bad style to assume that ba.Txn is ours.
// No race here, but should have a better way of doing this.
// TODO(tschottdorf): future refactoring should move this to txn
// creation in TxnCoordSender, which is currently unaware of the
// NodeID (and wraps *DistSender through client.Sender since it
// also needs test compatibility with *LocalSender).
ba.Txn.CertainNodes.Add(nDesc.NodeID)
}
}
if len(ba.Requests) < 1 {
panic("empty batch")
}
var rplChunks []*roachpb.BatchResponse
parts := ba.Split(false /* don't split ET */)
for len(parts) > 0 {
part := parts[0]
ba.Requests = part
rpl, pErr, shouldSplitET := ds.sendChunk(ctx, ba)
if shouldSplitET {
// If we tried to send a single round-trip EndTransaction but
// it looks like it's going to hit multiple ranges, split it
// here and try again.
if len(parts) != 1 {
panic("EndTransaction not in last chunk of batch")
}
parts = ba.Split(true /* split ET */)
if len(parts) != 2 {
panic("split of final EndTransaction chunk resulted in != 2 parts")
}
continue
}
if pErr != nil {
return nil, pErr
}
// Propagate transaction from last reply to next request. The final
// update is taken and put into the response's main header.
ba.Txn.Update(rpl.Header().Txn)
rplChunks = append(rplChunks, rpl)
parts = parts[1:]
}
reply := rplChunks[0]
for _, rpl := range rplChunks[1:] {
reply.Responses = append(reply.Responses, rpl.Responses...)
}
*reply.Header() = rplChunks[len(rplChunks)-1].BatchResponse_Header
return reply, nil
}
// sendRPC sends one or more RPCs to replicas from the supplied roachpb.Replica
// slice. First, replicas which have gossiped addresses are corralled (and
// rearranged depending on proximity and whether the request needs to go to a
// leader) and then sent via rpc.Send, with requirement that one RPC to a
// server must succeed. Returns an RPC error if the request could not be sent.
// Note that the reply may contain a higher level error and must be checked in
// addition to the RPC error.
func (ds *DistSender) sendRPC(trace *tracer.Trace, rangeID roachpb.RangeID, replicas replicaSlice, order rpc.OrderingPolicy,
ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if len(replicas) == 0 {
return nil, roachpb.NewError(noNodeAddrsAvailError{})
}
// Build a slice of replica addresses.
addrs := make([]net.Addr, 0, len(replicas))
replicaMap := make(map[string]*roachpb.ReplicaDescriptor, len(replicas))
for i := range replicas {
addr := replicas[i].NodeDesc.Address
addrs = append(addrs, addr)
replicaMap[addr.String()] = &replicas[i].ReplicaDescriptor
}
// TODO(pmattis): This needs to be tested. If it isn't set we'll
// still route the request appropriately by key, but won't receive
// RangeNotFoundErrors.
ba.RangeID = rangeID
// Set RPC opts with stipulation that one of N RPCs must succeed.
rpcOpts := rpc.Options{
Ordering: order,
SendNextTimeout: defaultSendNextTimeout,
Timeout: rpc.DefaultRPCTimeout,
Trace: trace,
}
// getArgs clones the arguments on demand for all but the first replica.
firstArgs := true
getArgs := func(addr net.Addr) proto.Message {
var a *roachpb.BatchRequest
// Use the supplied args proto if this is our first address.
if firstArgs {
firstArgs = false
a = &ba
} else {
// Otherwise, copy the args value and set the replica in the header.
a = proto.Clone(&ba).(*roachpb.BatchRequest)
}
if addr != nil {
// TODO(tschottdorf): see len(replicas) above.
a.Replica = *replicaMap[addr.String()]
}
return a
}
// RPCs are sent asynchronously and there is no synchronized access to
// the reply object, so we don't pass itself to RPCSend.
// Otherwise there maybe a race case:
// If the RPC call times out using our original reply object,
// we must not use it any more; the rpc call might still return
// and just write to it at any time.
// args.CreateReply() should be cheaper than proto.Clone which use reflect.
getReply := func() proto.Message {
return &roachpb.BatchResponse{}
}
tracer.AnnotateTrace()
defer tracer.AnnotateTrace()
const method = "Node.Batch"
reply, err := ds.rpcSend(rpcOpts, method, addrs, getArgs, getReply, ds.rpcContext)
if err != nil {
return nil, roachpb.NewError(err)
}
return reply.(*roachpb.BatchResponse), nil
}
// makePlan creates the query plan for a single SQL statement. The returned
// plan needs to be iterated over using planNode.Next() and planNode.Values()
// in order to retrieve matching rows. If autoCommit is true, the plan is
// allowed (but not required) to commit the transaction along with other KV
// operations.
//
// Note: The autoCommit parameter enables operations to enable the 1PC
// optimization. This is a bit hackish/preliminary at present.
func (p *planner) makePlan(stmt parser.Statement, autoCommit bool) (planNode, *roachpb.Error) {
tracer.AnnotateTrace()
// This will set the system DB trigger for transactions containing
// DDL statements that have no effect, such as
// `BEGIN; INSERT INTO ...; CREATE TABLE IF NOT EXISTS ...; COMMIT;`
// where the table already exists. This will generate some false
// refreshes, but that's expected to be quite rare in practice.
if stmt.StatementType() == parser.DDL {
p.txn.SetSystemConfigTrigger()
}
switch n := stmt.(type) {
case *parser.AlterTable:
return p.AlterTable(n)
case *parser.BeginTransaction:
return p.BeginTransaction(n)
case *parser.CommitTransaction:
return p.CommitTransaction(n)
case *parser.CreateDatabase:
return p.CreateDatabase(n)
case *parser.CreateIndex:
return p.CreateIndex(n)
case *parser.CreateTable:
return p.CreateTable(n)
case *parser.Delete:
return p.Delete(n)
case *parser.DropDatabase:
return p.DropDatabase(n)
case *parser.DropIndex:
return p.DropIndex(n)
case *parser.DropTable:
return p.DropTable(n)
case *parser.Explain:
return p.Explain(n)
case *parser.Grant:
return p.Grant(n)
case *parser.Insert:
return p.Insert(n, autoCommit)
case *parser.ParenSelect:
return p.makePlan(n.Select, autoCommit)
case *parser.RenameColumn:
return p.RenameColumn(n)
case *parser.RenameDatabase:
return p.RenameDatabase(n)
case *parser.RenameIndex:
return p.RenameIndex(n)
case *parser.RenameTable:
return p.RenameTable(n)
case *parser.Revoke:
return p.Revoke(n)
case *parser.RollbackTransaction:
return p.RollbackTransaction(n)
case *parser.Select:
return p.Select(n)
case *parser.Set:
return p.Set(n)
case *parser.SetTimeZone:
return p.SetTimeZone(n)
case *parser.SetTransaction:
return p.SetTransaction(n)
case *parser.Show:
return p.Show(n)
case *parser.ShowColumns:
return p.ShowColumns(n)
case *parser.ShowDatabases:
return p.ShowDatabases(n)
case *parser.ShowGrants:
return p.ShowGrants(n)
case *parser.ShowIndex:
return p.ShowIndex(n)
case *parser.ShowTables:
return p.ShowTables(n)
case *parser.Truncate:
return p.Truncate(n)
case *parser.Update:
return p.Update(n)
case parser.Values:
return p.Values(n)
default:
return nil, roachpb.NewErrorf("unknown statement type: %T", stmt)
}
}