From c6860d32ed05a9c934c8f9e8811c251ffb993d7d Mon Sep 17 00:00:00 2001
From: Radu Berinde <radu@cockroachlabs.com>
Date: Tue, 26 May 2026 08:59:09 -0700
Subject: [PATCH] rowblk: fix IsLowerBound to account for synthetic suffix

`Iter.firstUserKey` has the synthetic prefix prepended but the on-disk
suffix is preserved; suffix replacement is only applied per-returned key
via `maybeReplaceSuffix`. Comparing `firstUserKey` directly against the
probe could return true even though the iterator's actual first key
(after synthetic suffix replacement) sorts strictly less than the probe,
violating the `blockiter.Data.IsLowerBound` contract.

When a synthetic suffix is in effect, only compare the prefix portion of
`firstUserKey` and require it to be strictly greater than the probe. The
contract permits false negatives, so this conservative comparison is
sound and avoids reasoning about the unknown suffix.

Adds `TestIsLowerBoundRandomized` which cross-checks `IsLowerBound`
against `First()` across random blocks and transform combinations, and
documents the transform state of `firstUserKey`.
---
 sstable/rowblk/rowblk_iter.go      | 13 +++++
 sstable/rowblk/rowblk_iter_test.go | 76 ++++++++++++++++++++++++++++++
 2 files changed, 89 insertions(+)

diff --git a/sstable/rowblk/rowblk_iter.go b/sstable/rowblk/rowblk_iter.go
index ff81a21a4ef..93c728cf16f 100644
--- a/sstable/rowblk/rowblk_iter.go
+++ b/sstable/rowblk/rowblk_iter.go
@@ -187,6 +187,10 @@ type Iter struct {
 	cachedBuf []byte
 	handle    block.BufferHandle
 	// for block iteration for already loaded blocks.
+	//
+	// firstUserKey is the user key of the first KV in the block. It has the
+	// synthetic prefix applied (if any), but does NOT have the synthetic suffix
+	// applied — the on-disk suffix is preserved.
 	firstUserKey      []byte
 	lazyValueHandling struct {
 		getValue       block.GetInternalValueForPrefixAndValueHandler
@@ -541,6 +545,15 @@ func (i *Iter) cacheEntry() {
 // IsLowerBound implements the block.DataBlockIterator interface.
 func (i *Iter) IsLowerBound(k []byte) bool {
 	// Note: we ignore HideObsoletePoints, but false negatives are allowed.
+	if i.transforms.HasSyntheticSuffix() {
+		// firstUserKey has the synthetic prefix applied, but its suffix is the
+		// original on-disk suffix — not the synthetic suffix that the iterator will
+		// actually return. Instead of applying the synthetic suffix, we
+		// conservatively require firstUserKey's prefix portion to be strictly
+		// greater than k (false negatives are allowed).
+		firstPrefix := i.firstUserKey[:i.split(i.firstUserKey)]
+		return i.cmp(firstPrefix, k) > 0
+	}
 	return i.cmp(i.firstUserKey, k) >= 0
 }
 
diff --git a/sstable/rowblk/rowblk_iter_test.go b/sstable/rowblk/rowblk_iter_test.go
index 348b6039797..4e66271ceca 100644
--- a/sstable/rowblk/rowblk_iter_test.go
+++ b/sstable/rowblk/rowblk_iter_test.go
@@ -7,8 +7,11 @@ package rowblk
 import (
 	"bytes"
 	"fmt"
+	"math/rand/v2"
+	"slices"
 	"strings"
 	"testing"
+	"time"
 	"unsafe"
 
 	"github.com/cockroachdb/datadriven"
@@ -465,3 +468,76 @@ func TestBlockSyntheticSuffix(t *testing.T) {
 func ikey(s string) base.InternalKey {
 	return base.InternalKey{UserKey: []byte(s)}
 }
+
+func TestIsLowerBoundRand(t *testing.T) {
+	cmp := testkeys.Comparer.Compare
+	suffixCmp := testkeys.Comparer.ComparePointSuffixes
+	split := testkeys.Comparer.Split
+
+	prefixes := []string{"a", "b", "c", "d", "e", "p/", "zzz/"}
+	suffixes := []string{"", "@1", "@5", "@10", "@20", "@100"}
+
+	seed := uint64(time.Now().UnixNano())
+	t.Logf("seed: %d", seed)
+	rng := rand.New(rand.NewPCG(seed, seed))
+	randKey := func() string {
+		return prefixes[rng.IntN(len(prefixes))] + suffixes[rng.IntN(len(suffixes))]
+	}
+
+	for iter := 0; iter < 1000; iter++ {
+		// Construct a block with a random combination of prefixes and suffixes.
+		n := 1 + rng.IntN(20)
+		keySet := make(map[string]struct{}, n)
+		for len(keySet) < n {
+			keySet[randKey()] = struct{}{}
+		}
+		keys := make([]string, 0, n)
+		for k := range keySet {
+			keys = append(keys, k)
+		}
+		slices.SortFunc(keys, func(a, b string) int { return cmp([]byte(a), []byte(b)) })
+
+		w := &Writer{RestartInterval: 1 + rng.IntN(4)}
+		for _, k := range keys {
+			require.NoError(t, w.Add(ikey(k), nil))
+		}
+		blk := w.Finish()
+
+		// Pick a random prefix and suffix for the transform.
+		var synthPrefix, synthSuffix []byte
+		if rand.IntN(2) == 0 {
+			synthPrefix = []byte(prefixes[rng.IntN(len(prefixes))])
+		}
+		if rand.IntN(2) == 0 {
+			synthSuffix = []byte(suffixes[rng.IntN(len(suffixes))])
+		}
+		transforms := block.IterTransforms{
+			SyntheticPrefixAndSuffix: block.MakeSyntheticPrefixAndSuffix(synthPrefix, synthSuffix),
+		}
+
+		it, err := NewIter(cmp, suffixCmp, split, blk, transforms)
+		require.NoError(t, err)
+
+		kv := it.First()
+		require.NotNil(t, kv)
+		// firstKey is the key after the transform.
+		firstKey := slices.Clone(kv.K.UserKey)
+
+		for probe := 0; probe < 50; probe++ {
+			probeKey := []byte(randKey())
+			if rand.IntN(2) == 0 {
+				// The transformed block keys are random keys with synthPrefix
+				// prepended. To generate more interesting probe keys, sometimes prepend
+				// this prefix to the random key.
+				probeKey = append(slices.Clip(synthPrefix), probeKey...)
+			}
+			if it.IsLowerBound(probeKey) {
+				if cmp(firstKey, probeKey) < 0 {
+					t.Fatalf("IsLowerBound(%q)=true but First()=%q < probe (transforms=%+v, keys=%v)",
+						probeKey, firstKey, transforms, keys)
+				}
+			}
+		}
+		require.NoError(t, it.Close())
+	}
+}