fix(cpp): aligned VECTOR row-offset skip only when time/value counts match (apache#778)

Whole-chunk and whole-page skips by statistic count previously used only the
value side for aligned series, which could desynchronize row_offset from
decoded rows when ChunkMeta or page header counts differed.

Require both time and value statistics to be present, positive, and equal
before applying count-based skip; otherwise decode and rely on page/row
handling.

Made-with: Cursor

Author: hongzhi-gao

cpp/src/reader/aligned_chunk_reader.cc

@@ -740,15 +740,21 @@ bool AlignedChunkReader::should_skip_page_by_offset(int& row_offset) {
     if (row_offset <= 0) {
         return false;
     }
-    // Use time page statistic for count.
-    Statistic* stat = cur_time_page_header_.statistic_;
-    if (stat == nullptr) {
-        stat = cur_value_page_header_.statistic_;
+    // Aligned TV pages: only skip a whole page by count when both page headers
+    // expose the same positive row count. Using a single side (or min) when
+    // the other is missing or unequal can desynchronize row_offset from
+    // decoded row order vs. the paired time/value stream.
+    Statistic* ts = cur_time_page_header_.statistic_;
+    Statistic* vs = cur_value_page_header_.statistic_;
+    if (ts == nullptr || vs == nullptr) {
+        return false;
     }
-    if (stat == nullptr || stat->count_ == 0) {
+    int32_t tc = ts->count_;
+    int32_t vc = vs->count_;
+    if (tc <= 0 || vc <= 0 || tc != vc) {
         return false;
     }
-    int32_t count = stat->count_;
+    int32_t count = tc;
     if (row_offset >= count) {
         row_offset -= count;
         return true;

cpp/src/reader/tsfile_series_scan_iterator.cc

@@ -60,6 +60,30 @@ bool TsFileSeriesScanIterator::should_skip_chunk_by_offset(ChunkMeta* cm) {
     return false;
 }
 
+bool TsFileSeriesScanIterator::should_skip_aligned_chunk_by_offset(
+    ChunkMeta* time_cm, ChunkMeta* value_cm) {
+    if (row_offset_ <= 0) {
+        return false;
+    }
+    if (time_cm->statistic_ == nullptr || value_cm->statistic_ == nullptr) {
+        return false;
+    }
+    int32_t tc = time_cm->statistic_->count_;
+    int32_t vc = value_cm->statistic_->count_;
+    if (tc <= 0 || vc <= 0) {
+        return false;
+    }
+    if (tc != vc) {
+        return false;
+    }
+    int32_t count = tc;
+    if (row_offset_ >= count) {
+        row_offset_ -= count;
+        return true;
+    }
+    return false;
+}
+
 int TsFileSeriesScanIterator::get_next(TsBlock*& ret_tsblock, bool alloc,
                                        Filter* oneshoot_filter,
                                        int64_t min_time_hint) {
@@ -106,7 +130,8 @@ int TsFileSeriesScanIterator::get_next(TsBlock*& ret_tsblock, bool alloc,
                                                       min_time_hint)) {
                             continue;
                         }
-                        if (should_skip_chunk_by_offset(value_cm)) {
+                        if (should_skip_aligned_chunk_by_offset(time_cm,
+                                                                value_cm)) {
                             continue;
                         }
                         chunk_reader_->reset();

cpp/src/reader/tsfile_series_scan_iterator.h

@@ -119,6 +119,14 @@ class TsFileSeriesScanIterator {
     }
     bool should_skip_chunk_by_time(ChunkMeta* cm, int64_t min_time_hint);
     bool should_skip_chunk_by_offset(ChunkMeta* cm);
+    /**
+     * Aligned (VECTOR): whole-chunk skip by row count is only safe when the
+     * time ChunkMeta and value ChunkMeta agree on statistic count (>0). If
+     * either side lacks count or counts differ, skip is disabled for this
+     * chunk; pages are loaded and page/row-level offset handling applies.
+     */
+    bool should_skip_aligned_chunk_by_offset(ChunkMeta* time_cm,
+                                             ChunkMeta* value_cm);
     common::TsBlock* alloc_tsblock();
 
    private: