// TableStats is an endpoint that returns columns, indices, and other
// relevant details for the specified table.
func (s *adminServer) TableStats(
ctx context.Context, req *serverpb.TableStatsRequest,
) (*serverpb.TableStatsResponse, error) {
escDBName := parser.Name(req.Database).String()
if err := s.assertNotVirtualSchema(escDBName); err != nil {
return nil, err
}
// Get table span.
var tableSpan roachpb.Span
iexecutor := sql.InternalExecutor{LeaseManager: s.server.leaseMgr}
if err := s.server.db.Txn(ctx, func(txn *client.Txn) error {
var err error
tableSpan, err = iexecutor.GetTableSpan(s.getUser(req), txn, req.Database, req.Table)
return err
}); err != nil {
return nil, s.serverError(err)
}
startKey, err := keys.Addr(tableSpan.Key)
if err != nil {
return nil, s.serverError(err)
}
endKey, err := keys.Addr(tableSpan.EndKey)
if err != nil {
return nil, s.serverError(err)
}
// Get current range descriptors for table. This is done by scanning over
// meta2 keys for the range.
rangeDescKVs, err := s.server.db.Scan(ctx, keys.RangeMetaKey(startKey), keys.RangeMetaKey(endKey), 0)
if err != nil {
return nil, s.serverError(err)
}
// Extract a list of node IDs from the response.
nodeIDs := make(map[roachpb.NodeID]struct{})
for _, kv := range rangeDescKVs {
var rng roachpb.RangeDescriptor
if err := kv.Value.GetProto(&rng); err != nil {
return nil, s.serverError(err)
}
for _, repl := range rng.Replicas {
nodeIDs[repl.NodeID] = struct{}{}
}
}
// Construct TableStatsResponse by sending an RPC to every node involved.
tableStatResponse := serverpb.TableStatsResponse{
NodeCount: int64(len(nodeIDs)),
// TODO(mrtracy): The "RangeCount" returned by TableStats is more
// accurate than the "RangeCount" returned by TableDetails, because this
// method always consistently queries the meta2 key range for the table;
// in contrast, TableDetails uses a method on the DistSender, which
// queries using a range metadata cache and thus may return stale data
// for tables that are rapidly splitting. However, one potential
// *advantage* of using the DistSender is that it will populate the
// DistSender's range metadata cache in the case where meta2 information
// for this table is not already present; the query used by TableStats
// does not populate the DistSender cache. We should consider plumbing
// TableStats' meta2 query through the DistSender so that it will share
// the advantage of populating the cache (without the disadvantage of
// potentially returning stale data).
// See Github #5435 for some discussion.
RangeCount: int64(len(rangeDescKVs)),
}
type nodeResponse struct {
nodeID roachpb.NodeID
resp *serverpb.SpanStatsResponse
err error
}
// Send a SpanStats query to each node. Set a timeout on the context for
// these queries.
responses := make(chan nodeResponse)
nodeCtx, cancel := context.WithTimeout(ctx, base.NetworkTimeout)
defer cancel()
for nodeID := range nodeIDs {
nodeID := nodeID
if err := s.server.stopper.RunAsyncTask(nodeCtx, func(ctx context.Context) {
var spanResponse *serverpb.SpanStatsResponse
client, err := s.server.status.dialNode(nodeID)
if err == nil {
req := serverpb.SpanStatsRequest{
StartKey: startKey,
EndKey: endKey,
NodeID: nodeID.String(),
}
spanResponse, err = client.SpanStats(ctx, &req)
}
response := nodeResponse{
nodeID: nodeID,
resp: spanResponse,
err: err,
}
select {
case responses <- response:
// Response processed.
case <-ctx.Done():
// Context completed, response no longer needed.
}
}); err != nil {
return nil, err
}
}
for remainingResponses := len(nodeIDs); remainingResponses > 0; remainingResponses-- {
select {
case resp := <-responses:
// For nodes which returned an error, note that the node's data
// is missing. For successful calls, aggregate statistics.
if resp.err != nil {
tableStatResponse.MissingNodes = append(
tableStatResponse.MissingNodes,
serverpb.TableStatsResponse_MissingNode{
NodeID: resp.nodeID.String(),
ErrorMessage: resp.err.Error(),
},
)
} else {
tableStatResponse.Stats.Add(resp.resp.TotalStats)
tableStatResponse.ReplicaCount += int64(resp.resp.RangeCount)
}
case <-ctx.Done():
return nil, ctx.Err()
}
}
return &tableStatResponse, nil
}
// TableStats is an endpoint that returns columns, indices, and other
// relevant details for the specified table.
func (s *adminServer) TableStats(
ctx context.Context, req *serverpb.TableStatsRequest,
) (*serverpb.TableStatsResponse, error) {
// Get table span.
var tableSpan roachpb.Span
iexecutor := sql.InternalExecutor{LeaseManager: s.server.leaseMgr}
if err := s.server.db.Txn(ctx, func(txn *client.Txn) error {
var err error
tableSpan, err = iexecutor.GetTableSpan(s.getUser(req), txn, req.Database, req.Table)
return err
}); err != nil {
return nil, s.serverError(err)
}
startKey, err := keys.Addr(tableSpan.Key)
if err != nil {
return nil, s.serverError(err)
}
endKey, err := keys.Addr(tableSpan.EndKey)
if err != nil {
return nil, s.serverError(err)
}
// Get current range descriptors for table. This is done by scanning over
// meta2 keys for the range.
rangeDescKVs, err := s.server.db.Scan(ctx, keys.RangeMetaKey(startKey), keys.RangeMetaKey(endKey), 0)
if err != nil {
return nil, s.serverError(err)
}
// Extract a list of node IDs from the response.
nodeIDs := make(map[roachpb.NodeID]struct{})
for _, kv := range rangeDescKVs {
var rng roachpb.RangeDescriptor
if err := kv.Value.GetProto(&rng); err != nil {
return nil, s.serverError(err)
}
for _, repl := range rng.Replicas {
nodeIDs[repl.NodeID] = struct{}{}
}
}
// Construct TableStatsResponse by sending an RPC to every node involved.
tableStatResponse := serverpb.TableStatsResponse{
NodeCount: int64(len(nodeIDs)),
RangeCount: int64(len(rangeDescKVs)),
}
type nodeResponse struct {
nodeID roachpb.NodeID
resp *serverpb.SpanStatsResponse
err error
}
// Send a SpanStats query to each node. Set a timeout on the context for
// these queries.
responses := make(chan nodeResponse)
nodeCtx, cancel := context.WithTimeout(ctx, base.NetworkTimeout)
defer cancel()
for nodeID := range nodeIDs {
nodeID := nodeID
if err := s.server.stopper.RunAsyncTask(nodeCtx, func(ctx context.Context) {
var spanResponse *serverpb.SpanStatsResponse
client, err := s.server.status.dialNode(nodeID)
if err == nil {
req := serverpb.SpanStatsRequest{
StartKey: startKey,
EndKey: endKey,
NodeID: nodeID.String(),
}
spanResponse, err = client.SpanStats(ctx, &req)
}
response := nodeResponse{
nodeID: nodeID,
resp: spanResponse,
err: err,
}
select {
case responses <- response:
// Response processed.
case <-ctx.Done():
// Context completed, response no longer needed.
}
}); err != nil {
return nil, err
}
}
for remainingResponses := len(nodeIDs); remainingResponses > 0; remainingResponses-- {
select {
case resp := <-responses:
// For nodes which returned an error, note that the node's data
// is missing. For successful calls, aggregate statistics.
if resp.err != nil {
tableStatResponse.MissingNodes = append(
tableStatResponse.MissingNodes,
serverpb.TableStatsResponse_MissingNode{
NodeID: resp.nodeID.String(),
ErrorMessage: resp.err.Error(),
},
)
} else {
tableStatResponse.Stats.Add(resp.resp.TotalStats)
tableStatResponse.ReplicaCount += int64(resp.resp.RangeCount)
}
case <-ctx.Done():
return nil, ctx.Err()
}
}
return &tableStatResponse, nil
}