feat: repair Go stack traces (#3014)

Author: kolesnikovae

pkg/pprof/fix_go_heap_truncated.go

@@ -0,0 +1,266 @@
+package pprof
+
+import (
+	"bytes"
+	"reflect"
+	"sort"
+	"unsafe"
+
+	"golang.org/x/exp/slices"
+
+	profilev1 "github.com/grafana/pyroscope/api/gen/proto/go/google/v1"
+)
+
+const (
+	minGroupSize = 2
+
+	tokens    = 8
+	tokenLen  = 16
+	suffixLen = tokens + tokenLen
+
+	tokenBytesLen  = tokenLen * 8
+	suffixBytesLen = suffixLen * 8
+)
+
+// MayHaveGoHeapTruncatedStacktraces reports whether there are
+// any chances that the profile may have truncated stack traces.
+func MayHaveGoHeapTruncatedStacktraces(p *profilev1.Profile) bool {
+	if !hasGoHeapSampleTypes(p) {
+		return false
+	}
+	// Some truncated stacks have depth less than the depth limit (32).
+	const minDepth = 28
+	for _, s := range p.Sample {
+		if len(s.LocationId) >= minDepth {
+			return true
+		}
+	}
+	return false
+}
+
+func hasGoHeapSampleTypes(p *profilev1.Profile) bool {
+	for _, st := range p.SampleType {
+		switch p.StringTable[st.Type] {
+		case
+			"alloc_objects",
+			"alloc_space",
+			"inuse_objects",
+			"inuse_space":
+			return true
+		}
+	}
+	return false
+}
+
+// RepairGoHeapTruncatedStacktraces repairs truncated stack traces
+// in Go heap profiles.
+//
+// Go heap profile has a depth limit of 32 frames, which often
+// renders profiles unreadable, and also increases cardinality
+// of stack traces.
+//
+// The function guesses truncated frames based on neighbors and
+// repairs stack traces if there are high chances that this
+// part is present in the profile. The heuristic is as follows:
+//
+// For each stack trace S taller than 24 frames: if there is another
+// stack trace R taller than 24 frames that overlaps with the given
+// one by at least 16 frames in a row from the top, and has frames
+// above its root, stack S considered truncated, and the missing part
+// is copied from R.
+func RepairGoHeapTruncatedStacktraces(p *profilev1.Profile) {
+	// Group stack traces by bottom (closest to the root) locations.
+	// Typically, there are very few groups (a hundred or two).
+	samples, groups := split(p)
+	// Each group's suffix is then tokenized: each part is shifted by one
+	// location from the previous one (like n-grams).
+	// Tokens are written into the token=>group map, Where the value is the
+	// index of the group with the token found at the furthest position from
+	// the root (across all groups).
+	m := make(map[string]group, len(groups)/2)
+	for i := 0; i < len(groups); i++ {
+		g := groups[i]
+		n := len(groups)
+		if i+1 < len(groups) {
+			n = groups[i+1]
+		}
+		if s := n - g; s < minGroupSize {
+			continue
+		}
+		// We take suffix of the first sample in the group.
+		s := suffix(samples[g].LocationId)
+		// Tokenize the suffix: token position is relative to the stack
+		// trace root: 0 means that the token is the closest to the root.
+		//    TODO: unroll?
+		//    0 : 64 : 192 // No need.
+		//    1 : 56 : 184
+		//    2 : 48 : 176
+		//    3 : 40 : 168
+		//    4 : 32 : 160
+		//    5 : 24 : 152
+		//    6 : 16 : 144
+		//    7 :  8 : 136
+		//    8 :  0 : 128
+		//
+		// We skip the top/right-most token, as it is not needed,
+		// because there can be no more complete stack trace.
+		for j := uint32(1); j <= tokens; j++ {
+			hi := suffixBytesLen - j*tokens
+			lo := hi - tokenBytesLen
+			// By taking a string representation of the slice,
+			// we eliminate the need to hash the token explicitly:
+			// Go map will do it this way or another.
+			k := unsafeString(s[lo:hi])
+			// Current candidate: the group where the token is
+			// located at the furthest position from the root.
+			c, ok := m[k]
+			if !ok || j > c.off {
+				// This group has more complete stack traces:
+				m[k] = group{
+					gid: uint32(i),
+					off: j,
+				}
+			}
+		}
+	}
+
+	// Now we handle chaining. Consider the following stacks:
+	//  1   2   3   4
+	//  a   b  [d] (f)
+	//  b   c  [e] (g)
+	//  c  [d] (f)  h
+	//  d  [e] (g)  i
+	//
+	// We can't associate 3-rd stack with the 1-st one because their tokens
+	// do not overlap (given the token size is 2). However, we can associate
+	// it transitively through the 2nd stack.
+	//
+	// Dependencies:
+	//  - group i depends on d[i].
+	//  - d[i] depends on d[d[i].gid].
+	d := make([]group, len(groups))
+	for i := 0; i < len(groups); i++ {
+		g := groups[i]
+		t := topToken(samples[g].LocationId)
+		k := unsafeString(t)
+		c, ok := m[k]
+		if !ok || c.off == 0 || groups[c.gid] == g {
+			// The current group has the most complete stack trace.
+			continue
+		}
+		d[i] = c
+	}
+
+	// Then, for each group, we test, if there is another group with a more
+	// complete suffix, overlapping with the given one by at least one token.
+	// If such stack trace exists, all stack traces of the group are appended
+	// with the missing part.
+	for i := 0; i < len(groups); i++ {
+		g := groups[i]
+		c := d[i]
+		var off uint32
+		for c.off > 0 {
+			off += c.off
+			n := d[c.gid]
+			if n.off == 0 {
+				// Stop early to preserve c.
+				break
+			}
+			c = n
+		}
+		if off == 0 {
+			// The current group has the most complete stack trace.
+			continue
+		}
+		// The reference stack trace.
+		appx := samples[groups[c.gid]].LocationId
+		// It's possible that the reference stack trace does not
+		// include the part we're looking for. In this case, we
+		// simply ignore the group. Although it's possible to infer
+		// this piece from other stacks, this is left for further
+		// improvements.
+		if int(off) >= len(appx) {
+			continue
+		}
+		appx = appx[uint32(len(appx))-off:]
+		// Now we append the missing part to all stack traces of the group.
+		n := len(groups)
+		if i+1 < len(groups) {
+			n = groups[i+1]
+		}
+		for j := g; j < n; j++ {
+			// Locations typically already have some extra capacity,
+			// therefore no major allocations are expected here.
+			samples[j].LocationId = append(samples[j].LocationId, appx...)
+		}
+	}
+}
+
+type group struct {
+	gid uint32
+	off uint32
+}
+
+// suffix returns the last suffixLen locations
+// of the given stack trace represented as bytes.
+// The return slice is always suffixBytesLen long.
+// function panics if s is shorter than suffixLen.
+func suffix(s []uint64) []byte {
+	return locBytes(s[len(s)-suffixLen:])
+}
+
+// topToken returns the last tokenLen locations
+// of the given stack trace represented as bytes.
+// The return slice is always tokenBytesLen long.
+// function panics if s is shorter than tokenLen.
+func topToken(s []uint64) []byte {
+	return locBytes(s[len(s)-tokenLen:])
+}
+
+func locBytes(s []uint64) []byte {
+	size := len(s) * 8
+	h := (*reflect.SliceHeader)(unsafe.Pointer(&s))
+	h.Len = size
+	h.Cap = size
+	return *(*[]byte)(unsafe.Pointer(h))
+}
+
+func unsafeString(b []byte) string {
+	return *(*string)(unsafe.Pointer(&b))
+}
+
+// split into groups of samples by stack trace suffixes.
+// Return slice contains indices of the first sample
+// of each group in the collection of selected samples.
+func split(p *profilev1.Profile) ([]*profilev1.Sample, []int) {
+	slices.SortFunc(p.Sample, func(a, b *profilev1.Sample) int {
+		if len(a.LocationId) < suffixLen {
+			return -1
+		}
+		if len(b.LocationId) < suffixLen {
+			return 1
+		}
+		return bytes.Compare(
+			suffix(a.LocationId),
+			suffix(b.LocationId),
+		)
+	})
+	o := sort.Search(len(p.Sample), func(i int) bool {
+		return len(p.Sample[i].LocationId) >= suffixLen
+	})
+	if o == len(p.Sample) {
+		return nil, nil
+	}
+	samples := p.Sample[o:]
+	const avgGroupSize = 16 // Estimate.
+	groups := make([]int, 0, len(samples)/avgGroupSize)
+	var prev []byte
+	for i := 0; i < len(samples); i++ {
+		cur := suffix(samples[i].LocationId)
+		if !bytes.Equal(cur, prev) {
+			groups = append(groups, i)
+			prev = cur
+		}
+	}
+	return samples, groups
+}

pkg/pprof/fix_go_heap_truncated_test.go

@@ -0,0 +1,43 @@
+package pprof
+
+import (
+	"os"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func Benchmark_RepairGoTruncatedStacktraces(b *testing.B) {
+	p, err := OpenFile("testdata/goheapfix/heap_go_truncated_3.pb.gz")
+	require.NoError(b, err)
+	b.ResetTimer()
+	b.ReportAllocs()
+	for i := 0; i < b.N; i++ {
+		RepairGoHeapTruncatedStacktraces(FixGoProfile(p.CloneVT()))
+	}
+}
+
+func Test_UpdateFixtures_RepairGoTruncatedStacktraces(t *testing.T) {
+	t.Skip()
+	t.Helper()
+	paths := []string{
+		"testdata/goheapfix/heap_go_truncated_1.pb.gz", // Cortex.
+		"testdata/goheapfix/heap_go_truncated_2.pb.gz", // Cortex.
+		"testdata/goheapfix/heap_go_truncated_3.pb.gz", // Loki. Pathological.
+		"testdata/goheapfix/heap_go_truncated_4.pb.gz", // Pyroscope.
+	}
+	for _, path := range paths {
+		func() {
+			p, err := OpenFile(path)
+			require.NoError(t, err, path)
+			defer p.Close()
+			f, err := os.Create(path + ".fixed")
+			require.NoError(t, err, path)
+			defer f.Close()
+			p.Profile = FixGoProfile(p.Profile)
+			RepairGoHeapTruncatedStacktraces(p.Profile)
+			_, err = p.WriteTo(f)
+			require.NoError(t, err, path)
+		}()
+	}
+}

pkg/pprof/fix_go_profile.go

@@ -1,25 +1,47 @@
 package pprof
 
 import (
+	"regexp"
+	"strings"
+
 	profilev1 "github.com/grafana/pyroscope/api/gen/proto/go/google/v1"
 )
 
-// FixGoProfile removes type parameters from Go function names.
-//
-// In go 1.21 and above, function names include type parameters,
-// however, due to the bug, a function name can include any
-// of the type instances regardless of the call site. Thus, e.g.,
-// x[T1].foo and x[T2].foo can't be distinguished in a profile.
-// This leads to incorrect profiles and incorrect flame graphs,
-// and hugely increases cardinality of stack traces.
+// FixGoProfile fixes known issues with profiles collected with
+// the standard Go profiler.
 //
-// FixGoProfile will change x[T1].foo and x[T2].foo to x[...].foo
-// and normalize the profile, if type parameters are present in
-// the profile. Otherwise, the profile returned unchanged.
+// Note that the function presumes that p is a Go profile and does
+// not verify this. It is expected that the function is called
+// very early in the profile processing chain and normalized after,
+// regardless of the function outcome.
 func FixGoProfile(p *profilev1.Profile) *profilev1.Profile {
+	p = DropGoTypeParameters(p)
+	// Now that the profile is normalized, we can try to repair
+	// truncated stack traces, if any. Note that repaired stacks
+	// are not deduplicated, so the caller need to normalize the
+	if MayHaveGoHeapTruncatedStacktraces(p) {
+		RepairGoHeapTruncatedStacktraces(p)
+	}
+	return p
+}
+
+// DropGoTypeParameters removes of type parameters from Go function names.
+//
+// In go 1.21 and above, function names include type parameters, however,
+// due to a bug, a function name could include any of the type instances
+// regardless of the call site. Thus, e.g., x[T1].foo and x[T2].foo can't
+// be distinguished in a profile. This leads to incorrect profiles and
+// incorrect flame graphs, and hugely increases cardinality of stack traces.
+//
+// The function renames x[T1].foo and x[T2].foo to x[...].foo and normalizes
+// the profile, if type parameters are present in the profile. Otherwise, the
+// profile returns unchanged.
+//
+// See https://github.com/golang/go/issues/64528.
+func DropGoTypeParameters(p *profilev1.Profile) *profilev1.Profile {
 	var n int
 	for i, s := range p.StringTable {
-		c := DropGoTypeParameters(s)
+		c := dropGoTypeParameters(s)
 		if c != s {
 			p.StringTable[i] = c
 			n++
@@ -37,3 +59,21 @@ func FixGoProfile(p *profilev1.Profile) *profilev1.Profile {
 	_ = m.MergeNoClone(p)
 	return m.Profile()
 }
+
+var goStructTypeParameterRegex = regexp.MustCompile(`\[go\.shape\..*\]`)
+
+func dropGoTypeParameters(input string) string {
+	matchesIndices := goStructTypeParameterRegex.FindAllStringIndex(input, -1)
+	if len(matchesIndices) == 0 {
+		return input
+	}
+	var modified strings.Builder
+	prevEnd := 0
+	for _, indices := range matchesIndices {
+		start, end := indices[0], indices[1]
+		modified.WriteString(input[prevEnd:start] + "[...]")
+		prevEnd = end
+	}
+	modified.WriteString(input[prevEnd:])
+	return modified.String()
+}