diff --git a/include/rocksdb/perf_context.h b/include/rocksdb/perf_context.h
index 7cae30aee..e2e6844cc 100644
--- a/include/rocksdb/perf_context.h
+++ b/include/rocksdb/perf_context.h
@@ -30,10 +30,33 @@ struct PerfContext {
   uint64_t block_read_time;           // total nanos spent on block reads
   uint64_t block_checksum_time;       // total nanos spent on block checksum
   uint64_t block_decompress_time;  // total nanos spent on block decompression
-  // total number of internal keys skipped over during iteration (overwritten or
-  // deleted, to be more specific, hidden by a put or delete of the same key)
+  // total number of internal keys skipped over during iteration.
+  // There are several reasons for it:
+  // 1. when calling Next(), iterator is in the position of the previous key,
+  //    so that we'll need to skip it. It means this counter will always
+  //    incrmented in Next().
+  // 2. when calling Next(), needs to skip internal entries for the previous
+  //    key that are overwritten.
+  // 3. when calling Next(), Seek() or SeekToFirst(), after previous key
+  //    before calling Next(), the seek key in Seek() or the beginning for
+  //    SeekToFirst(), there may be one or more deleted keys before the next
+  //    valid key that the operation should place the iterator to. We need
+  //    to skip both of the tombstone and updates hidden by the tombstones. The
+  //    tombstones are not included in this counter, while previous updates
+  //    hidden by the tombstones will be included here.
+  // 4. symmetric cases for Prev() and SeekToLast()
+  // We sometimes also skip entries of more recent updates than the snapshot
+  // we read from, but it is not counted.
+  //
   uint64_t internal_key_skipped_count;
-  // total number of deletes and single deletes skipped over during iteration
+  // Total number of deletes and single deletes skipped over during iteration
+  // When calling Next(), Seek() or SeekToFirst(), after previous key before
+  // calling Next(), the seek key in Seek() or the beginning for SeekToFirst(),
+  // there may be one or more deleted keys before the next valid key that the
+  // operation should place the iterator to. Every deleted key is counted once.
+  // We don't recount here if there are still older updates invalidated by the
+  // tombstones.
+  //
   uint64_t internal_delete_skipped_count;
 
   uint64_t get_snapshot_time;       // total nanos spent on getting snapshot