Fix too much small row group issue in Velox parquet writer (oap-proje…

…ct#326)

Co-authored-by: youxiduo <[email protected]>

velox/dwio/parquet/writer/Writer.cpp

@@ -21,52 +21,95 @@
 
 namespace facebook::velox::parquet {
 
+void Writer::flush() {
+  if (stagingRows_ > 0) {
+    if (!arrowWriter_) {
+      stream_ = std::make_shared<DataBufferSink>(
+          finalSink_.get(),
+          pool_,
+          queryCtx_->queryConfig().dataBufferGrowRatio());
+      auto arrowProperties = ::parquet::ArrowWriterProperties::Builder().build();
+      PARQUET_ASSIGN_OR_THROW(
+          arrowWriter_,
+          ::parquet::arrow::FileWriter::Open(
+              *(schema_.get()),
+              arrow::default_memory_pool(),
+              stream_,
+              properties_,
+              arrowProperties));
+    }
+
+    auto fields = schema_->fields();
+    std::vector<std::shared_ptr<arrow::ChunkedArray>> chunks;
+    for (int colIdx = 0; colIdx < fields.size(); colIdx++) {
+      auto dataType = fields.at(colIdx)->type();
+      auto chunk = arrow::ChunkedArray::Make(std::move(stagingChunks_.at(colIdx)), dataType).ValueOrDie();
+      chunks.push_back(chunk);
+    }
+    auto table = arrow::Table::Make(schema_, std::move(chunks), stagingRows_);
+    PARQUET_THROW_NOT_OK(arrowWriter_->WriteTable(*table, maxRowGroupRows_));
+    if (queryCtx_->queryConfig().dataBufferGrowRatio() > 1) {
+      PARQUET_THROW_NOT_OK(stream_->Flush());
+    }
+    for (auto& chunk : stagingChunks_) {
+      chunk.clear();
+    }
+    stagingRows_ = 0;
+    stagingBytes_ = 0;
+  }
+}
+
+/**
+ * This method would cache input `ColumnarBatch` to make the size of row group big.
+ * It would flush when:
+ * - the cached numRows bigger than `maxRowGroupRows_`
+ * - the cached bytes bigger than `maxRowGroupBytes_`
+ *
+ * This method assumes each input `ColumnarBatch` have same schema.
+ */
 void Writer::write(const RowVectorPtr& data) {
   ArrowArray array;
   ArrowSchema schema;
   exportToArrow(data, array, &pool_);
   exportToArrow(data, schema);
   PARQUET_ASSIGN_OR_THROW(
       auto recordBatch, arrow::ImportRecordBatch(&array, &schema));
-  auto table = arrow::Table::Make(
-      recordBatch->schema(), recordBatch->columns(), data->size());
-  if (!arrowWriter_) {
-    stream_ = std::make_shared<DataBufferSink>(
-        finalSink_.get(),
-        pool_,
-        queryCtx_->queryConfig().dataBufferGrowRatio());
-    auto arrowProperties = ::parquet::ArrowWriterProperties::Builder().build();
-    PARQUET_ASSIGN_OR_THROW(
-        arrowWriter_,
-        ::parquet::arrow::FileWriter::Open(
-            *recordBatch->schema(),
-            arrow::default_memory_pool(),
-            stream_,
-            properties_,
-            arrowProperties));
+  if (!schema_) {
+    schema_ = recordBatch->schema();
+    for (int colIdx = 0; colIdx < schema_->num_fields(); colIdx++) {
+      stagingChunks_.push_back(std::vector<std::shared_ptr<arrow::Array>>());
+    }
   }
 
-  PARQUET_THROW_NOT_OK(arrowWriter_->WriteTable(*table, 10000));
-
-  if (queryCtx_->queryConfig().dataBufferGrowRatio() > 1) {
-    flush(); // No performance drop on 1TB dataset.
+  auto bytes = data->estimateFlatSize();
+  auto numRows = data->size();
+  if (stagingBytes_ + bytes > maxRowGroupBytes_ || stagingRows_ + numRows > maxRowGroupRows_) {
+    flush();
   }
-}
 
-void Writer::flush() {
-  PARQUET_THROW_NOT_OK(stream_->Flush());
+  for (int colIdx = 0; colIdx < recordBatch->num_columns(); colIdx++) {
+    auto array = recordBatch->column(colIdx);
+    stagingChunks_.at(colIdx).push_back(array);
+  }
+  stagingRows_ += numRows;
+  stagingBytes_ += bytes;
 }
 
 void Writer::newRowGroup(int32_t numRows) {
   PARQUET_THROW_NOT_OK(arrowWriter_->NewRowGroup(numRows));
 }
 
 void Writer::close() {
+  flush();
+
   if (arrowWriter_) {
     PARQUET_THROW_NOT_OK(arrowWriter_->Close());
     arrowWriter_.reset();
   }
+
   PARQUET_THROW_NOT_OK(stream_->Close());
+
+  stagingChunks_.clear();
 }
 
 } // namespace facebook::velox::parquet

velox/dwio/parquet/writer/Writer.h

@@ -90,18 +90,19 @@ class DataBufferSink : public arrow::io::OutputStream {
 class Writer {
  public:
   // Constructts a writer with output to 'sink'. A new row group is
-  // started every 'rowsInRowGroup' top level rows. 'pool' is used for
+  // started every 'maxRowGroupBytes' top level rows. 'pool' is used for
   // temporary memory. 'properties' specifies Parquet-specific
   // options.
   Writer(
       std::unique_ptr<dwio::common::DataSink> sink,
       memory::MemoryPool& pool,
-      int32_t rowsInRowGroup,
+      int64_t maxRowGroupBytes,
       std::shared_ptr<::parquet::WriterProperties> properties =
           ::parquet::WriterProperties::Builder().build(),
       std::shared_ptr<velox::core::QueryCtx> queryCtx =
           std::make_shared<velox::core::QueryCtx>(nullptr))
-      : rowsInRowGroup_(rowsInRowGroup),
+      : maxRowGroupBytes_(maxRowGroupBytes),
+        maxRowGroupRows_(properties->max_row_group_length()),
         pool_(pool),
         finalSink_(std::move(sink)),
         properties_(std::move(properties)),
@@ -121,11 +122,20 @@ class Writer {
   void close();
 
  private:
-  const int32_t rowsInRowGroup_;
+  const int64_t maxRowGroupBytes_;
+  const int64_t maxRowGroupRows_;
+
+  int64_t stagingRows_ = 0;
+  int64_t stagingBytes_ = 0;
 
   // Pool for 'stream_'.
   memory::MemoryPool& pool_;
 
+  std::shared_ptr<arrow::Schema> schema_;
+
+  // columns, Arrays
+  std::vector<std::vector<std::shared_ptr<arrow::Array>>> stagingChunks_;
+
   // Final destination of output.
   std::unique_ptr<dwio::common::DataSink> finalSink_;