// sendRPC sends one or more RPCs to replicas from the supplied
// roachpb.Replica slice. 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.
//
// The replicas are assumed to be ordered by preference, with closer
// ones (i.e. expected lowest latency) first.
func (ds *DistSender) sendRPC(
ctx context.Context, rangeID roachpb.RangeID, replicas ReplicaSlice, ba roachpb.BatchRequest,
) (*roachpb.BatchResponse, error) {
if len(replicas) == 0 {
return nil, roachpb.NewSendError(
fmt.Sprintf("no replica node addresses available via gossip for range %d", rangeID))
}
// 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 := SendOptions{
ctx: ctx,
SendNextTimeout: ds.sendNextTimeout,
transportFactory: ds.transportFactory,
}
tracing.AnnotateTrace()
defer tracing.AnnotateTrace()
reply, err := ds.sendToReplicas(rpcOpts, rangeID, replicas, ba, ds.rpcContext)
if err != nil {
return nil, err
}
return reply, nil
}
// sendRPC sends one or more RPCs to replicas from the supplied
// roachpb.Replica slice. 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.
//
// The replicas are assumed to be ordered by preference, with closer
// ones (i.e. expected lowest latency) first.
func (ds *DistSender) sendRPC(
ctx context.Context, rangeID roachpb.RangeID, replicas ReplicaSlice, ba roachpb.BatchRequest,
) (*roachpb.BatchResponse, error) {
if len(replicas) == 0 {
return nil, roachpb.NewSendError(
fmt.Sprintf("no replica node addresses available via gossip for range %d", rangeID))
}
// 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
// A given RPC may generate retries to multiple replicas, but as soon as we
// get a response from one we want to cancel those other RPCs.
ctx, cancel := context.WithCancel(ctx)
defer cancel()
// Set RPC opts with stipulation that one of N RPCs must succeed.
rpcOpts := SendOptions{
SendNextTimeout: ds.sendNextTimeout,
transportFactory: ds.transportFactory,
metrics: &ds.metrics,
}
tracing.AnnotateTrace()
defer tracing.AnnotateTrace()
reply, err := ds.sendToReplicas(ctx, rpcOpts, rangeID, replicas, ba, ds.rpcContext)
if err != nil {
return nil, err
}
return reply, nil
}
// Run executes the operations queued up within a batch. Before executing any
// of the operations the batch is first checked to see if there were any errors
// during its construction (e.g. failure to marshal a proto message).
//
// The operations within a batch are run in parallel and the order is
// non-deterministic. It is an unspecified behavior to modify and retrieve the
// same key within a batch.
//
// Upon completion, Batch.Results will contain the results for each
// operation. The order of the results matches the order the operations were
// added to the batch.
func (txn *Txn) Run(b *Batch) error {
tracing.AnnotateTrace()
defer tracing.AnnotateTrace()
if err := b.prepare(); err != nil {
return err
}
return sendAndFill(txn.send, b)
}
func (n *scanNode) Next() (bool, error) {
tracing.AnnotateTrace()
if !n.scanInitialized {
if err := n.initScan(); err != nil {
return false, err
}
}
if n.explain == explainDebug {
return n.debugNext()
}
// We fetch one row at a time until we find one that passes the filter.
for {
var err error
n.row, err = n.fetcher.NextRowDecoded()
if err != nil || n.row == nil {
return false, err
}
passesFilter, err := sqlbase.RunFilter(n.filter, &n.p.evalCtx)
if err != nil {
return false, err
}
if passesFilter {
return true, nil
}
}
}
func (u *updateNode) Next() (bool, error) {
next, err := u.run.rows.Next()
if !next {
if err == nil {
// We're done. Finish the batch.
err = u.tw.finalize(u.p.ctx())
}
return false, err
}
if u.run.explain == explainDebug {
return true, nil
}
tracing.AnnotateTrace()
oldValues := u.run.rows.Values()
// Our updated value expressions occur immediately after the plain
// columns in the output.
updateValues := oldValues[len(u.tw.ru.fetchCols):]
oldValues = oldValues[:len(u.tw.ru.fetchCols)]
u.checkHelper.loadRow(u.tw.ru.fetchColIDtoRowIndex, oldValues, false)
u.checkHelper.loadRow(u.updateColsIdx, updateValues, true)
if err := u.checkHelper.check(&u.p.evalCtx); err != nil {
return false, err
}
// Ensure that the values honor the specified column widths.
for i := range updateValues {
if err := sqlbase.CheckValueWidth(u.tw.ru.updateCols[i], updateValues[i]); err != nil {
return false, err
}
}
// Update the row values.
for i, col := range u.tw.ru.updateCols {
val := updateValues[i]
if !col.Nullable && val == parser.DNull {
return false, sqlbase.NewNonNullViolationError(col.Name)
}
}
newValues, err := u.tw.row(u.p.ctx(), append(oldValues, updateValues...))
if err != nil {
return false, err
}
resultRow, err := u.rh.cookResultRow(newValues)
if err != nil {
return false, err
}
u.run.resultRow = resultRow
return true, nil
}
func (c *v3Conn) executeStatements(
ctx context.Context,
stmts string,
pinfo *parser.PlaceholderInfo,
formatCodes []formatCode,
sendDescription bool,
limit int,
) error {
tracing.AnnotateTrace()
// Note: sql.Executor gets its Context from c.session.context, which
// has been bound by v3Conn.setupSession().
results := c.executor.ExecuteStatements(c.session, stmts, pinfo)
defer results.Close()
tracing.AnnotateTrace()
if results.Empty {
// Skip executor and just send EmptyQueryResponse.
c.writeBuf.initMsg(serverMsgEmptyQuery)
return c.writeBuf.finishMsg(c.wr)
}
return c.sendResponse(ctx, results.ResultList, formatCodes, sendDescription, limit)
}
func (db *DB) prepareToSend(ba *roachpb.BatchRequest) *roachpb.Error {
if ba.ReadConsistency == roachpb.INCONSISTENT {
for _, ru := range ba.Requests {
req := ru.GetInner()
if req.Method() != roachpb.Get && req.Method() != roachpb.Scan &&
req.Method() != roachpb.ReverseScan {
return roachpb.NewErrorf("method %s not allowed with INCONSISTENT batch", req.Method)
}
}
}
if db.ctx.UserPriority != 1 {
ba.UserPriority = db.ctx.UserPriority
}
tracing.AnnotateTrace()
return nil
}
// Update updates columns for a selection of rows from a table.
// Privileges: UPDATE and SELECT on table. We currently always use a select statement.
// Notes: postgres requires UPDATE. Requires SELECT with WHERE clause with table.
// mysql requires UPDATE. Also requires SELECT with WHERE clause with table.
// TODO(guanqun): need to support CHECK in UPDATE
func (p *planner) Update(
n *parser.Update, desiredTypes []parser.Type, autoCommit bool,
) (planNode, error) {
tracing.AnnotateTrace()
tn, err := p.getAliasedTableName(n.Table)
if err != nil {
return nil, err
}
en, err := p.makeEditNode(tn, autoCommit, privilege.UPDATE)
if err != nil {
return nil, err
}
exprs := make([]*parser.UpdateExpr, len(n.Exprs))
for i, expr := range n.Exprs {
// Replace the sub-query nodes.
newExpr, err := p.replaceSubqueries(expr.Expr, len(expr.Names))
if err != nil {
return nil, err
}
exprs[i] = &parser.UpdateExpr{Tuple: expr.Tuple, Expr: newExpr, Names: expr.Names}
}
// Determine which columns we're inserting into.
names, err := p.namesForExprs(exprs)
if err != nil {
return nil, err
}
updateCols, err := p.processColumns(en.tableDesc, names)
if err != nil {
return nil, err
}
defaultExprs, err := makeDefaultExprs(updateCols, &p.parser, &p.evalCtx)
if err != nil {
return nil, err
}
var requestedCols []sqlbase.ColumnDescriptor
if len(n.Returning) > 0 || len(en.tableDesc.Checks) > 0 {
// TODO(dan): This could be made tighter, just the rows needed for RETURNING
// exprs.
requestedCols = en.tableDesc.Columns
}
fkTables := tablesNeededForFKs(*en.tableDesc, CheckUpdates)
if err := p.fillFKTableMap(fkTables); err != nil {
return nil, err
}
ru, err := makeRowUpdater(p.txn, en.tableDesc, fkTables, updateCols, requestedCols, rowUpdaterDefault)
if err != nil {
return nil, err
}
tw := tableUpdater{ru: ru, autoCommit: autoCommit}
tracing.AnnotateTrace()
// Generate the list of select targets. We need to select all of the columns
// plus we select all of the update expressions in case those expressions
// reference columns (e.g. "UPDATE t SET v = v + 1"). Note that we flatten
// expressions for tuple assignments just as we flattened the column names
// above. So "UPDATE t SET (a, b) = (1, 2)" translates into select targets of
// "*, 1, 2", not "*, (1, 2)".
targets := sqlbase.ColumnsSelectors(ru.fetchCols)
i := 0
// Remember the index where the targets for exprs start.
exprTargetIdx := len(targets)
desiredTypesFromSelect := make([]parser.Type, len(targets), len(targets)+len(exprs))
for i := range targets {
desiredTypesFromSelect[i] = parser.TypeAny
}
for _, expr := range exprs {
if expr.Tuple {
switch t := expr.Expr.(type) {
case (*parser.Tuple):
for _, e := range t.Exprs {
typ := updateCols[i].Type.ToDatumType()
e := fillDefault(e, typ, i, defaultExprs)
targets = append(targets, parser.SelectExpr{Expr: e})
desiredTypesFromSelect = append(desiredTypesFromSelect, typ)
i++
}
default:
return nil, fmt.Errorf("cannot use this expression to assign multiple columns: %s", expr.Expr)
}
} else {
typ := updateCols[i].Type.ToDatumType()
e := fillDefault(expr.Expr, typ, i, defaultExprs)
targets = append(targets, parser.SelectExpr{Expr: e})
desiredTypesFromSelect = append(desiredTypesFromSelect, typ)
i++
}
}
rows, err := p.SelectClause(&parser.SelectClause{
Exprs: targets,
From: &parser.From{Tables: []parser.TableExpr{n.Table}},
Where: n.Where,
}, nil, nil, desiredTypesFromSelect, publicAndNonPublicColumns)
if err != nil {
return nil, err
}
// Placeholders have their types populated in the above Select if they are part
// of an expression ("SET a = 2 + $1") in the type check step where those
// types are inferred. For the simpler case ("SET a = $1"), populate them
// using checkColumnType. This step also verifies that the expression
// types match the column types.
sel := rows.(*selectTopNode).source.(*selectNode)
for i, target := range sel.render[exprTargetIdx:] {
// DefaultVal doesn't implement TypeCheck
if _, ok := target.(parser.DefaultVal); ok {
continue
}
// TODO(nvanbenschoten) isn't this TypeCheck redundant with the call to SelectClause?
typedTarget, err := parser.TypeCheck(target, &p.semaCtx, updateCols[i].Type.ToDatumType())
if err != nil {
return nil, err
}
err = sqlbase.CheckColumnType(updateCols[i], typedTarget.ResolvedType(), p.semaCtx.Placeholders)
if err != nil {
return nil, err
}
}
updateColsIdx := make(map[sqlbase.ColumnID]int, len(ru.updateCols))
for i, col := range ru.updateCols {
updateColsIdx[col.ID] = i
}
un := &updateNode{
n: n,
editNodeBase: en,
updateCols: ru.updateCols,
updateColsIdx: updateColsIdx,
tw: tw,
}
if err := un.checkHelper.init(p, tn, en.tableDesc); err != nil {
return nil, err
}
if err := un.run.initEditNode(&un.editNodeBase, rows, n.Returning, desiredTypes); err != nil {
return nil, err
}
return un, 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 by range and a partial
// subset of the batch request is sent to affected ranges in parallel.
//
// The first write in a transaction may not arrive before writes to
// other ranges. This is relevant in the case of a BeginTransaction
// request. Intents written to other ranges before the transaction
// record is created will cause the transaction to abort early.
func (ds *DistSender) Send(
ctx context.Context, ba roachpb.BatchRequest,
) (*roachpb.BatchResponse, *roachpb.Error) {
tracing.AnnotateTrace()
if pErr := ds.initAndVerifyBatch(ctx, &ba); pErr != nil {
return nil, pErr
}
ctx = ds.AnnotateCtx(ctx)
ctx, cleanup := tracing.EnsureContext(ctx, ds.AmbientContext.Tracer)
defer cleanup()
var rplChunks []*roachpb.BatchResponse
parts := ba.Split(false /* don't split ET */)
if len(parts) > 1 && ba.MaxSpanRequestKeys != 0 {
// We already verified above that the batch contains only scan requests of the same type.
// Such a batch should never need splitting.
panic("batch with MaxSpanRequestKeys needs splitting")
}
for len(parts) > 0 {
part := parts[0]
ba.Requests = part
// The minimal key range encompassing all requests contained within.
// Local addressing has already been resolved.
// TODO(tschottdorf): consider rudimentary validation of the batch here
// (for example, non-range requests with EndKey, or empty key ranges).
rs, err := keys.Range(ba)
if err != nil {
return nil, roachpb.NewError(err)
}
rpl, pErr := ds.divideAndSendBatchToRanges(ctx, ba, rs, true /* isFirst */)
if pErr == errNo1PCTxn {
// 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.UpdateTxn(rpl.Txn)
rplChunks = append(rplChunks, rpl)
parts = parts[1:]
}
reply := rplChunks[0]
for _, rpl := range rplChunks[1:] {
reply.Responses = append(reply.Responses, rpl.Responses...)
reply.CollectedSpans = append(reply.CollectedSpans, rpl.CollectedSpans...)
}
reply.BatchResponse_Header = rplChunks[len(rplChunks)-1].BatchResponse_Header
return reply, nil
}
// newPlan constructs a planNode from a statement. This is used
// recursively by the various node constructors.
func (p *planner) newPlan(
stmt parser.Statement, desiredTypes []parser.Type, autoCommit bool,
) (planNode, error) {
tracing.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 CopyDataBlock:
return p.CopyData(n, autoCommit)
case *parser.CopyFrom:
return p.CopyFrom(n, autoCommit)
case *parser.CreateDatabase:
return p.CreateDatabase(n)
case *parser.CreateIndex:
return p.CreateIndex(n)
case *parser.CreateTable:
return p.CreateTable(n)
case *parser.CreateUser:
return p.CreateUser(n)
case *parser.CreateView:
return p.CreateView(n)
case *parser.Delete:
return p.Delete(n, desiredTypes, autoCommit)
case *parser.DropDatabase:
return p.DropDatabase(n)
case *parser.DropIndex:
return p.DropIndex(n)
case *parser.DropTable:
return p.DropTable(n)
case *parser.DropView:
return p.DropView(n)
case *parser.Explain:
return p.Explain(n, autoCommit)
case *parser.Grant:
return p.Grant(n)
case *parser.Help:
return p.Help(n)
case *parser.Insert:
return p.Insert(n, desiredTypes, autoCommit)
case *parser.ParenSelect:
return p.newPlan(n.Select, desiredTypes, 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.Select:
return p.Select(n, desiredTypes, autoCommit)
case *parser.SelectClause:
return p.SelectClause(n, nil, nil, desiredTypes, publicColumns)
case *parser.Set:
return p.Set(n)
case *parser.SetTimeZone:
return p.SetTimeZone(n)
case *parser.SetTransaction:
return p.SetTransaction(n)
case *parser.SetDefaultIsolation:
return p.SetDefaultIsolation(n)
case *parser.Show:
return p.Show(n)
case *parser.ShowColumns:
return p.ShowColumns(n)
case *parser.ShowConstraints:
return p.ShowConstraints(n)
case *parser.ShowCreateTable:
return p.ShowCreateTable(n)
case *parser.ShowCreateView:
return p.ShowCreateView(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.ShowUsers:
return p.ShowUsers(n)
case *parser.Split:
return p.Split(n)
case *parser.Truncate:
return p.Truncate(n)
case *parser.UnionClause:
return p.UnionClause(n, desiredTypes, autoCommit)
case *parser.Update:
return p.Update(n, desiredTypes, autoCommit)
case *parser.ValuesClause:
return p.ValuesClause(n, desiredTypes)
default:
return nil, errors.Errorf("unknown statement type: %T", stmt)
}
}
// newPlan constructs a planNode from a statement. This is used
// recursively by the various node constructors.
func (p *planner) newPlan(
stmt parser.Statement, desiredTypes []parser.Type, autoCommit bool,
) (planNode, error) {
tracing.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()
}
// TODO(dan): This iteration makes the plan dispatch no longer constant
// time. We could fix that with a map of `reflect.Type` but including
// reflection in such a primary codepath is unfortunate. Instead, the
// upcoming IR work will provide unique numeric type tags, which will
// elegantly solve this.
for _, planHook := range planHooks {
if fn, header, err := planHook(p.ctx(), stmt, p.execCfg); err != nil {
return nil, err
} else if fn != nil {
return &hookFnNode{f: fn, header: header}, nil
}
}
switch n := stmt.(type) {
case *parser.AlterTable:
return p.AlterTable(n)
case *parser.BeginTransaction:
return p.BeginTransaction(n)
case CopyDataBlock:
return p.CopyData(n, autoCommit)
case *parser.CopyFrom:
return p.CopyFrom(n, autoCommit)
case *parser.CreateDatabase:
return p.CreateDatabase(n)
case *parser.CreateIndex:
return p.CreateIndex(n)
case *parser.CreateTable:
return p.CreateTable(n)
case *parser.CreateUser:
return p.CreateUser(n)
case *parser.CreateView:
return p.CreateView(n)
case *parser.Delete:
return p.Delete(n, desiredTypes, autoCommit)
case *parser.DropDatabase:
return p.DropDatabase(n)
case *parser.DropIndex:
return p.DropIndex(n)
case *parser.DropTable:
return p.DropTable(n)
case *parser.DropView:
return p.DropView(n)
case *parser.Explain:
return p.Explain(n, autoCommit)
case *parser.Grant:
return p.Grant(n)
case *parser.Help:
return p.Help(n)
case *parser.Insert:
return p.Insert(n, desiredTypes, autoCommit)
case *parser.ParenSelect:
return p.newPlan(n.Select, desiredTypes, 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.Select:
return p.Select(n, desiredTypes, autoCommit)
case *parser.SelectClause:
return p.SelectClause(n, nil, nil, desiredTypes, publicColumns)
case *parser.Set:
return p.Set(n)
case *parser.SetTimeZone:
return p.SetTimeZone(n)
case *parser.SetTransaction:
return p.SetTransaction(n)
case *parser.SetDefaultIsolation:
return p.SetDefaultIsolation(n)
case *parser.Show:
return p.Show(n)
case *parser.ShowColumns:
return p.ShowColumns(n)
case *parser.ShowConstraints:
return p.ShowConstraints(n)
case *parser.ShowCreateTable:
return p.ShowCreateTable(n)
case *parser.ShowCreateView:
return p.ShowCreateView(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.ShowUsers:
return p.ShowUsers(n)
case *parser.Split:
return p.Split(n)
case *parser.Truncate:
return p.Truncate(n)
case *parser.UnionClause:
return p.UnionClause(n, desiredTypes, autoCommit)
case *parser.Update:
return p.Update(n, desiredTypes, autoCommit)
case *parser.ValuesClause:
return p.ValuesClause(n, desiredTypes)
default:
return nil, errors.Errorf("unknown statement type: %T", stmt)
}
}