sstable, db: document SeekPrefixGE for sstable and level iters

This is an attempt to document the current semantics as they are
currently implemented.

Also see the discussions in #3845

I hope to be looking at a way to resolve some of these complexities,
perhaps by adding more tombstone-related logic in the merging iterator
(postponing seeks instead of seeking beyond the prefix just to enact a
levelIter positioning suitable for a subsequent try-seek-using-next).

Author: RaduBerinde

level_iter.go

@@ -703,6 +703,70 @@ func (l *levelIter) internalSeekGE(
 	return l.verify(kv), kvMeta
 }
 
+// SeekPrefixGE implements InternalIterator.SeekPrefixGE. It positions the
+// iterator at the first key greater than or equal to key across all files in
+// the level. It returns the key-value pair at that position, or nil if no such
+// key exists.
+//
+// The prefix argument is passed to each file's iterator for bloom filter
+// checking. If a file's bloom filter indicates that prefix is not present, that
+// file is skipped. The key argument is used for the actual seek positioning.
+//
+// # Prefix vs key
+//
+// The prefix is typically the prefix of key (i.e. Split.Prefix(key) == prefix),
+// but this is not required. The only requirement is that prefix be less than or
+// equal to Split.Prefix(key). This flexibility is used when a higher-level
+// range deletion invalidates keys between prefix and Split.Prefix(key).
+//
+// For example, consider a SeekPrefixGE for prefix P1. A range deletion from a
+// higher level may cover all keys with prefix P1, requiring the merging
+// iterator to seek past the tombstone to a key with prefix P3 > P1. In this
+// case, the merging iterator calls SeekPrefixGE(P1, keyWithPrefixP3, flags).
+// The bloom filter is checked against P1 (the original prefix), while the
+// actual seek targets keyWithPrefixP3.
+//
+// # File positioning and TrySeekUsingNext
+//
+// The prefix is stored and controls file advancement when the current file is
+// exhausted. If the current file's largest key has a prefix greater than the
+// seek prefix, the iterator stops rather than advancing to the next file.
+//
+// This stopping condition is critical for TrySeekUsingNext correctness. Without
+// it, the following scenario would produce incorrect results:
+//
+//  1. SeekPrefixGE(P1, key1) - bloom filter misses on file F, no key found
+//  2. Iterator advances to file G (incorrectly, if G's smallest prefix > P1)
+//  3. SeekPrefixGE(P2, key2, TrySeekUsingNext) where P1 < P2 < G's smallest prefix
+//  4. TrySeekUsingNext starts from G, completely skipping file F which may
+//     contain keys with prefix P2
+//
+// The stopping condition prevents step 2: if F's largest prefix > P1, the
+// iterator remains at F, allowing the subsequent seek for P2 to correctly
+// examine F.
+//
+// # Return values and iterator state
+//
+//   - If a key is found: returns the key-value pair and positions the iterator
+//     at that key. The iterator may be positioned in any file that contains
+//     matching keys.
+//
+//   - If no key is found: returns nil. The iterator is exhausted in the forward
+//     direction. Subsequent Next() calls will return nil.
+//
+// # TrySeekUsingNext optimization
+//
+// The TrySeekUsingNext flag in flags indicates that the caller knows the seek
+// key is greater than or equal to the iterator's current position. When set:
+//
+//   - At the file level: the iterator may scan forward through files rather
+//     than performing a binary search through the file metadata.
+//
+//   - At the sstable level: the optimization is automatically disabled when a
+//     new file is loaded, since the new file's iterator is not yet positioned.
+//
+// Note: The caller must ensure that key is greater than or equal to the lower
+// bound. SeekPrefixGE checks the upper bound but not the lower bound.
 func (l *levelIter) SeekPrefixGE(prefix, key []byte, flags base.SeekGEFlags) (kv *base.InternalKV) {
 	if treesteps.Enabled && treesteps.IsRecording(l) {
 		op := treesteps.StartOpf(l, "SeekPrefixGE(%q, %q, %d)", prefix, key, flags)

sstable/reader_iter_single_lvl.go

@@ -839,9 +839,44 @@ func (i *singleLevelIterator[I, PI, D, PD]) seekGEHelper(
 	return i.skipForward(shouldReturnMeta)
 }
 
-// SeekPrefixGE implements internalIterator.SeekPrefixGE, as documented in the
-// pebble package. Note that SeekPrefixGE only checks the upper bound. It is up
-// to the caller to ensure that key is greater than or equal to the lower bound.
+// SeekPrefixGE implements InternalIterator.SeekPrefixGE. It positions the
+// iterator at the first key greater than or equal to key, as long as that key
+// has a prefix matching prefix. It returns the key-value pair at that position,
+// or nil if no such key exists in the table.
+//
+// The prefix argument is used exclusively for bloom filter checking. If the
+// table has a bloom filter and the filter indicates that prefix is not present,
+// SeekPrefixGE returns nil without positioning the iterator. The key argument
+// is used for the actual seek positioning when the bloom filter check passes or
+// is not applicable.
+//
+// The prefix is typically the prefix of key (i.e. Split.Prefix(key) == prefix),
+// but this is not required. The only requirement is that prefix be less than or
+// equal to Split.Prefix(key). This flexibility is used by higher-layer
+// iterators.
+//
+// Return values and iterator state:
+//
+//   - If a key is found: returns the key-value pair and positions the iterator
+//     at that key. Subsequent Next() calls will return following keys.
+//
+//   - If no key is found because the bloom filter excluded the prefix: returns
+//     nil. The iterator is not positioned; calling Next() or Prev() after this
+//     is not permitted.
+//
+//   - If no key is found but the bloom filter matched (or was not used): returns
+//     nil. The iterator may be exhausted (reached bounds or end of table).
+//     Calling Next() when exhausted forward will panic.
+//
+// The TrySeekUsingNext flag in flags indicates that the caller knows the seek
+// key is greater than or equal to the iterator's current position. When set,
+// the iterator may optimize by scanning forward from its current position
+// rather than performing a full seek. This optimization is automatically
+// disabled when the previous SeekPrefixGE returned nil due to a bloom filter
+// miss, since the iterator was not repositioned in that case.
+//
+// Note: SeekPrefixGE only checks the upper bound. It is up to the caller to
+// ensure that key is greater than or equal to the lower bound.
 func (i *singleLevelIterator[I, PI, D, PD]) SeekPrefixGE(
 	prefix, key []byte, flags base.SeekGEFlags,
 ) (kv *base.InternalKV) {