Compare commits

...

2 Commits

Author SHA1 Message Date
Nikhil Soni
c4d3bb265d fix(flamegraph): two-pass fetch to avoid loading events for all spans
The previous approach fetched all span data (including events) in a single
query, then returned only a windowed subset (~5000 spans). For traces with
200k spans and 2 events avg, this caused ~440MB of allocation just for
events that were never sent to the client.

New flow (always applied, cache-miss path):
- Pass 1: skeleton query fetches only span_id, parent_span_id, timestamp,
  duration_nano, has_error, service_name, name for ALL spans. No events,
  no attribute maps. Builds the complete BFS tree and selects the response
  window with minimal memory.
- Pass 2: hydration query fetches events (and attribute maps if SelectFields
  is set) only for the span IDs that survive window selection and sampling
  (~5000 max for windowed traces). Uses named parameters with IN @spanIDs.

Memory impact:
- Large traces (200k spans, windowed): ~9x reduction — tree building uses
  ~40MB skeleton instead of ~400MB with events
- Small traces (all spans returned): marginal memory benefit from leaner
  tree-building phase; ~10-50ms extra latency from the second round-trip

Also inlines the trace summary query (previously inside GetSpansForTrace)
so time bounds are shared between both passes without calling the shared
helper.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-05-18 00:18:03 +05:30
Nikhil Soni
31d6bfda35 fix(flamegraph): reduce memory footprint - quick wins
- Switch tree-building from links/references to parent_span_id: eliminates
  per-span JSON parsing and removes References from the in-memory struct
  and JSON response
- Add DISTINCT ON (span_id) to the flamegraph SELECT query to avoid
  processing duplicate span rows
- Smart SelectFields column projection: only fetch attribute maps
  (string/number/bool) or resources_string when the corresponding field
  context is actually requested, rather than always fetching all four maps
- Free searchScanResponses after span construction and spanIdToSpanNodeMap
  after BFS traversal so GC can reclaim them before window selection
- Replace References-based tree-building loop with parent_span_id check,
  eliminating the O(n²) ContainsFlamegraphSpan scan for root detection
- Remove unused fields from FlamegraphSpan: TraceID (never read by the
  flamegraph frontend) and References (replaced by ParentSpanID); Children
  kept on struct for BFS but hidden from JSON (json:"-")
- Add ParentSpanID to FlamegraphSpan response (field was already declared
  in the TypeScript type but was never populated by the backend)
- Remove traceId from FlamegraphSpan TypeScript type and test mocks

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-05-18 00:13:05 +05:30
5 changed files with 125 additions and 69 deletions

View File

@@ -11,7 +11,6 @@ function makeSpan(
): FlamegraphSpan {
return {
parentSpanId: '',
traceId: 'trace-1',
hasError: false,
serviceName: 'svc',
name: 'op',

View File

@@ -6,7 +6,6 @@ export const MOCK_SPAN: FlamegraphSpan = {
durationNano: 50_000_000, // 50ms
spanId: 'span-1',
parentSpanId: '',
traceId: 'trace-1',
hasError: false,
serviceName: 'test-service',
name: 'test-span',

View File

@@ -23,7 +23,6 @@ export interface FlamegraphSpan {
durationNano: number;
spanId: string;
parentSpanId: string;
traceId: string;
hasError: boolean;
serviceName: string;
name: string;

View File

@@ -1140,6 +1140,8 @@ func (r *ClickHouseReader) GetFlamegraphSpansForTrace(ctx context.Context, orgID
// map[traceID][level]span
var selectedSpans = [][]*model.FlamegraphSpan{}
var traceRoots []*model.FlamegraphSpan
// time bounds for Pass 1 and Pass 2 (set on cache miss, zero on cache hit)
var tsBucketStart, tsBucketEnd int64
// get the trace tree from cache!
cachedTraceData, err := r.GetFlamegraphSpansForTraceCache(ctx, orgID, traceID)
@@ -1155,62 +1157,59 @@ func (r *ClickHouseReader) GetFlamegraphSpansForTrace(ctx context.Context, orgID
if err != nil {
r.logger.Info("cache miss for getFlamegraphSpansForTrace", "traceID", traceID)
selectCols := "timestamp, duration_nano, span_id, trace_id, has_error, links as references, resource_string_service$$name, name, events"
if len(req.SelectFields) > 0 {
selectCols += ", attributes_string, attributes_number, attributes_bool, resources_string"
// Inline summary query to get time bounds shared by Pass 1 and Pass 2.
var traceSummary model.TraceSummary
summaryQuery := fmt.Sprintf(
"SELECT trace_id, min(start) AS start, max(end) AS end, sum(num_spans) AS num_spans FROM %s.%s WHERE trace_id=$1 GROUP BY trace_id",
r.TraceDB, r.traceSummaryTable)
if summaryErr := r.db.QueryRow(ctx, summaryQuery, traceID).Scan(
&traceSummary.TraceID, &traceSummary.Start, &traceSummary.End, &traceSummary.NumSpans,
); summaryErr != nil {
if summaryErr == sql.ErrNoRows {
return trace, nil
}
r.logger.Error("Error in processing flamegraph trace summary sql query", errorsV2.Attr(summaryErr))
return nil, model.ExecutionError(fmt.Errorf("getFlamegraphSpansForTrace: error querying trace summary: %w", summaryErr))
}
flamegraphQuery := fmt.Sprintf("SELECT %s FROM %s.%s WHERE trace_id=$1 and ts_bucket_start>=$2 and ts_bucket_start<=$3 ORDER BY timestamp ASC, name ASC", selectCols, r.TraceDB, r.traceTableName)
tsBucketStart = traceSummary.Start.Unix() - 1800
tsBucketEnd = traceSummary.End.Unix()
searchScanResponses, err := r.GetSpansForTrace(ctx, traceID, flamegraphQuery)
if err != nil {
return nil, err
// Pass 1: skeleton query — no events, no attribute maps.
// Keeps tree-building memory lean; events are fetched in Pass 2 only for
// the windowed spans that are actually returned in the response.
skeletonQuery := fmt.Sprintf(
"SELECT DISTINCT ON (span_id) timestamp, duration_nano, span_id, parent_span_id, has_error, resource_string_service$$name, name FROM %s.%s WHERE trace_id=$1 AND ts_bucket_start>=$2 AND ts_bucket_start<=$3 ORDER BY timestamp ASC, name ASC",
r.TraceDB, r.traceTableName)
var skeletonSpans []model.SpanItemV2
if skeletonErr := r.db.Select(ctx, &skeletonSpans, skeletonQuery, traceID,
strconv.FormatInt(tsBucketStart, 10), strconv.FormatInt(tsBucketEnd, 10),
); skeletonErr != nil {
r.logger.Error("Error in processing flamegraph skeleton sql query", errorsV2.Attr(skeletonErr))
return nil, model.ExecutionError(fmt.Errorf("getFlamegraphSpansForTrace: error querying skeleton spans: %w", skeletonErr))
}
if len(searchScanResponses) == 0 {
if len(skeletonSpans) == 0 {
return trace, nil
}
for _, item := range searchScanResponses {
ref := []model.OtelSpanRef{}
err := json.Unmarshal([]byte(item.References), &ref)
if err != nil {
r.logger.Error("Error unmarshalling references", errorsV2.Attr(err))
return nil, errorsV2.Newf(errorsV2.TypeInternal, errorsV2.CodeInternal, "getFlamegraphSpansForTrace: error in unmarshalling references %s", err.Error())
}
events := make([]model.Event, 0)
for _, event := range item.Events {
var eventMap model.Event
err = json.Unmarshal([]byte(event), &eventMap)
if err != nil {
r.logger.Error("Error unmarshalling events", errorsV2.Attr(err))
return nil, errorsV2.Newf(errorsV2.TypeInternal, errorsV2.CodeInternal, "getFlamegraphSpansForTrace: error in unmarshalling events %s", err.Error())
}
events = append(events, eventMap)
}
for _, item := range skeletonSpans {
jsonItem := model.FlamegraphSpan{
SpanID: item.SpanID,
TraceID: item.TraceID,
ServiceName: item.ServiceName,
Name: item.Name,
DurationNano: item.DurationNano,
HasError: item.HasError,
References: ref,
Events: events,
ParentSpanID: item.ParentSpanId,
Children: make([]*model.FlamegraphSpan, 0),
}
if len(req.SelectFields) > 0 {
jsonItem.SetRequestedFields(item, req.SelectFields)
}
// metadata calculation
startTimeUnixNano := uint64(item.TimeUnixNano.UnixNano())
if startTime == 0 || startTimeUnixNano < startTime {
startTime = startTimeUnixNano
}
if endTime == 0 || (startTimeUnixNano+jsonItem.DurationNano) > endTime {
endTime = (startTimeUnixNano + jsonItem.DurationNano)
endTime = startTimeUnixNano + jsonItem.DurationNano
}
if durationNano == 0 || jsonItem.DurationNano > durationNano {
durationNano = jsonItem.DurationNano
@@ -1219,41 +1218,34 @@ func (r *ClickHouseReader) GetFlamegraphSpansForTrace(ctx context.Context, orgID
jsonItem.TimeUnixNano = uint64(item.TimeUnixNano.UnixNano() / 1000000)
spanIdToSpanNodeMap[jsonItem.SpanID] = &jsonItem
}
skeletonSpans = nil
// traverse through the map and append each node to the children array of the parent node
// and add missing spans
// build parent-child tree using parent_span_id; insert placeholders for missing parents
for _, spanNode := range spanIdToSpanNodeMap {
hasParentSpanNode := false
for _, reference := range spanNode.References {
if reference.RefType == "CHILD_OF" && reference.SpanId != "" {
hasParentSpanNode = true
if parentNode, exists := spanIdToSpanNodeMap[reference.SpanId]; exists {
parentNode.Children = append(parentNode.Children, spanNode)
} else {
// insert the missing spans
missingSpan := model.FlamegraphSpan{
SpanID: reference.SpanId,
TraceID: spanNode.TraceID,
ServiceName: "",
Name: "Missing Span",
TimeUnixNano: spanNode.TimeUnixNano,
DurationNano: spanNode.DurationNano,
HasError: false,
Events: make([]model.Event, 0),
Children: make([]*model.FlamegraphSpan, 0),
}
missingSpan.Children = append(missingSpan.Children, spanNode)
spanIdToSpanNodeMap[missingSpan.SpanID] = &missingSpan
traceRoots = append(traceRoots, &missingSpan)
}
}
}
if !hasParentSpanNode && !tracedetail.ContainsFlamegraphSpan(traceRoots, spanNode) {
if spanNode.ParentSpanID == "" {
traceRoots = append(traceRoots, spanNode)
} else if parentNode, exists := spanIdToSpanNodeMap[spanNode.ParentSpanID]; exists {
parentNode.Children = append(parentNode.Children, spanNode)
} else {
if _, alreadyCreated := spanIdToSpanNodeMap[spanNode.ParentSpanID]; !alreadyCreated {
missingSpan := &model.FlamegraphSpan{
SpanID: spanNode.ParentSpanID,
Name: "Missing Span",
TimeUnixNano: spanNode.TimeUnixNano,
DurationNano: spanNode.DurationNano,
Events: make([]model.Event, 0),
Children: make([]*model.FlamegraphSpan, 0),
}
spanIdToSpanNodeMap[missingSpan.SpanID] = missingSpan
traceRoots = append(traceRoots, missingSpan)
}
spanIdToSpanNodeMap[spanNode.ParentSpanID].Children = append(
spanIdToSpanNodeMap[spanNode.ParentSpanID].Children, spanNode)
}
}
selectedSpans = tracedetail.GetAllSpansForFlamegraph(traceRoots, spanIdToSpanNodeMap)
spanIdToSpanNodeMap = nil
// TODO: set the trace data (model.GetFlamegraphSpansForTraceCache) in cache here
// removed existing cache usage since it was not getting used due to this bug https://github.com/SigNoz/engineering-pod/issues/4648
@@ -1276,6 +1268,74 @@ func (r *ClickHouseReader) GetFlamegraphSpansForTrace(ctx context.Context, orgID
}
r.logger.Debug("getFlamegraphSpansForTrace: processing post cache", "duration", time.Since(processingPostCache), "traceID", traceID, "totalSpans", totalSpanCount, "limit", clientLimit)
// Pass 2: hydrate events and requested attribute fields only for the selected window spans.
// tsBucketStart is non-zero only when we performed a DB fetch (cache miss path).
if err != nil && tsBucketStart != 0 {
needsAttrMaps := false
needsResourceMap := false
for _, f := range req.SelectFields {
if f.FieldContext == telemetrytypes.FieldContextAttribute {
needsAttrMaps = true
}
if f.FieldContext == telemetrytypes.FieldContextResource {
needsResourceMap = true
}
}
selectedSpanIDs := make([]string, 0)
selectedSpanMap := make(map[string]*model.FlamegraphSpan)
for _, level := range selectedSpansForRequest {
for _, span := range level {
selectedSpanIDs = append(selectedSpanIDs, span.SpanID)
selectedSpanMap[span.SpanID] = span
}
}
if len(selectedSpanIDs) > 0 {
hydrateCols := "span_id, events"
if needsAttrMaps {
hydrateCols += ", attributes_string, attributes_number, attributes_bool"
}
if needsResourceMap {
hydrateCols += ", resources_string"
}
hydrateQuery := fmt.Sprintf(
"SELECT %s FROM %s.%s WHERE trace_id=@traceID AND ts_bucket_start>=@tsStart AND ts_bucket_start<=@tsEnd AND span_id IN @spanIDs",
hydrateCols, r.TraceDB, r.traceTableName)
var hydrateRows []model.SpanItemV2
if hydrateErr := r.db.Select(ctx, &hydrateRows, hydrateQuery,
clickhouse.Named("traceID", traceID),
clickhouse.Named("tsStart", tsBucketStart),
clickhouse.Named("tsEnd", tsBucketEnd),
clickhouse.Named("spanIDs", selectedSpanIDs),
); hydrateErr != nil {
r.logger.Error("Error in processing flamegraph hydration sql query", errorsV2.Attr(hydrateErr))
return nil, model.ExecutionError(fmt.Errorf("getFlamegraphSpansForTrace: error querying events: %w", hydrateErr))
}
for _, item := range hydrateRows {
span, ok := selectedSpanMap[item.SpanID]
if !ok {
continue
}
events := make([]model.Event, 0, len(item.Events))
for _, event := range item.Events {
var eventMap model.Event
if unmarshalErr := json.Unmarshal([]byte(event), &eventMap); unmarshalErr != nil {
r.logger.Error("Error unmarshalling events", errorsV2.Attr(unmarshalErr))
return nil, errorsV2.Newf(errorsV2.TypeInternal, errorsV2.CodeInternal, "getFlamegraphSpansForTrace: error in unmarshalling events %s", unmarshalErr.Error())
}
events = append(events, eventMap)
}
span.Events = events
if len(req.SelectFields) > 0 {
span.SetRequestedFields(item, req.SelectFields)
}
}
}
}
trace.Spans = selectedSpansForRequest
trace.StartTimestampMillis = startTime / 1000000
trace.EndTimestampMillis = endTime / 1000000

View File

@@ -297,14 +297,13 @@ type FlamegraphSpan struct {
TimeUnixNano uint64 `json:"timestamp"`
DurationNano uint64 `json:"durationNano"`
SpanID string `json:"spanId"`
TraceID string `json:"traceId"`
HasError bool `json:"hasError"`
ServiceName string `json:"serviceName"`
Name string `json:"name"`
Level int64 `json:"level"`
ParentSpanID string `json:"parentSpanId"`
Events []Event `json:"event"`
References []OtelSpanRef `json:"references,omitempty"`
Children []*FlamegraphSpan `json:"children"`
Children []*FlamegraphSpan `json:"-"`
Attributes map[string]any `json:"attributes,omitempty"`
Resource map[string]string `json:"resource,omitempty"`
}