Patching parquet archive writer for high memory usage

pkg/metricstore/parquetArchive.go

@@ -31,15 +31,15 @@ type ParquetMetricRow struct {
 	Value     float32 `parquet:"value"`
 }
 
-// flattenCheckpointFile recursively converts a CheckpointFile tree into Parquet rows.
+// flattenCheckpointFile recursively converts a CheckpointFile tree into Parquet rows via a channel.
 // The scope path is built from the hierarchy: host level is "node", then child names
 // map to scope/scope_id (e.g., "socket0" → scope="socket", scope_id="0").
-func flattenCheckpointFile(cf *CheckpointFile, cluster, hostname, scope, scopeID string, rows []ParquetMetricRow) []ParquetMetricRow {
+func flattenCheckpointFile(cf *CheckpointFile, cluster, hostname, scope, scopeID string, rowChan chan<- *ParquetMetricRow) {
 	for metricName, cm := range cf.Metrics {
 		ts := cm.Start
 		for _, v := range cm.Data {
 			if !v.IsNaN() {
-				rows = append(rows, ParquetMetricRow{
+				rowChan <- &ParquetMetricRow{
 					Cluster:   cluster,
 					Hostname:  hostname,
 					Metric:    metricName,
@@ -48,7 +48,7 @@ func flattenCheckpointFile(cf *CheckpointFile, cluster, hostname, scope, scopeID
 					Timestamp: ts,
 					Frequency: cm.Frequency,
 					Value:     float32(v),
-				})
+				}
 			}
 			ts += cm.Frequency
 		}
@@ -56,10 +56,8 @@ func flattenCheckpointFile(cf *CheckpointFile, cluster, hostname, scope, scopeID
 
 	for childName, childCf := range cf.Children {
 		childScope, childScopeID := parseScopeFromName(childName)
-		rows = flattenCheckpointFile(childCf, cluster, hostname, childScope, childScopeID, rows)
+		flattenCheckpointFile(childCf, cluster, hostname, childScope, childScopeID, rowChan)
 	}
-
-	return rows
 }
 
 // parseScopeFromName infers scope and scope_id from a child level name.
@@ -91,15 +89,23 @@ func parseScopeFromName(name string) (string, string) {
 	return name, ""
 }
 
-// writeParquetArchive writes rows to a Parquet file with Zstd compression.
-func writeParquetArchive(filename string, rows []ParquetMetricRow) error {
+// writeParquetArchiveStream writes rows from a channel to a Parquet file with Zstd compression in batches.
+func writeParquetArchiveStream(filename string, rowChan <-chan *ParquetMetricRow) (int, error) {
 	if err := os.MkdirAll(filepath.Dir(filename), CheckpointDirPerms); err != nil {
-		return fmt.Errorf("creating archive directory: %w", err)
+		go func() {
+			for range rowChan {
+			}
+		}()
+		return 0, fmt.Errorf("creating archive directory: %w", err)
 	}
 
 	f, err := os.OpenFile(filename, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, CheckpointFilePerms)
 	if err != nil {
-		return fmt.Errorf("creating parquet file: %w", err)
+		go func() {
+			for range rowChan {
+			}
+		}()
+		return 0, fmt.Errorf("creating parquet file: %w", err)
 	}
 	defer f.Close()
 
@@ -115,19 +121,45 @@ func writeParquetArchive(filename string, rows []ParquetMetricRow) error {
 		)),
 	)
 
-	if _, err := writer.Write(rows); err != nil {
-		return fmt.Errorf("writing parquet rows: %w", err)
+	batchSize := 4096
+	batch := make([]ParquetMetricRow, 0, batchSize)
+	rowCount := 0
+	var writeErr error
+
+	for rowPtr := range rowChan {
+		if writeErr != nil {
+			continue // Drain the channel to prevent worker deadlock
+		}
+		batch = append(batch, *rowPtr)
+		if len(batch) >= batchSize {
+			if _, err := writer.Write(batch); err != nil {
+				writeErr = fmt.Errorf("writing parquet batch: %w", err)
+			}
+			rowCount += len(batch)
+			batch = batch[:0]
+		}
+	}
+
+	if writeErr != nil {
+		return rowCount, writeErr
+	}
+
+	if len(batch) > 0 {
+		if _, err := writer.Write(batch); err != nil {
+			return rowCount, fmt.Errorf("writing remaining parquet batch: %w", err)
+		}
+		rowCount += len(batch)
 	}
 
 	if err := writer.Close(); err != nil {
-		return fmt.Errorf("closing parquet writer: %w", err)
+		return rowCount, fmt.Errorf("closing parquet writer: %w", err)
 	}
 
 	if err := bw.Flush(); err != nil {
-		return fmt.Errorf("flushing parquet file: %w", err)
+		return rowCount, fmt.Errorf("flushing parquet file: %w", err)
 	}
 
-	return nil
+	return rowCount, nil
 }
 
 // loadCheckpointFileFromDisk reads a JSON or binary checkpoint file and returns
@@ -180,24 +212,22 @@ func loadCheckpointFileFromDisk(filename string) (*CheckpointFile, error) {
 }
 
 // archiveCheckpointsToParquet reads checkpoint files for a host directory,
-// converts them to Parquet rows. Returns the rows and filenames that were processed.
-func archiveCheckpointsToParquet(dir, cluster, host string, from int64) ([]ParquetMetricRow, []string, error) {
+// converts them to Parquet rows. Returns the filenames that were processed.
+func archiveCheckpointsToParquet(dir, cluster, host string, from int64, rowChan chan<- *ParquetMetricRow) ([]string, error) {
 	entries, err := os.ReadDir(dir)
 	if err != nil {
-		return nil, nil, err
+		return nil, err
 	}
 
 	files, err := findFiles(entries, from, false)
 	if err != nil {
-		return nil, nil, err
+		return nil, err
 	}
 
 	if len(files) == 0 {
-		return nil, nil, nil
+		return nil, nil
 	}
 
-	var rows []ParquetMetricRow
-
 	for _, checkpoint := range files {
 		filename := filepath.Join(dir, checkpoint)
 		cf, err := loadCheckpointFileFromDisk(filename)
@@ -206,8 +236,8 @@ func archiveCheckpointsToParquet(dir, cluster, host string, from int64) ([]Parqu
 			continue
 		}
 
-		rows = flattenCheckpointFile(cf, cluster, host, "node", "", rows)
+		flattenCheckpointFile(cf, cluster, host, "node", "", rowChan)
 	}
 
-	return rows, files, nil
+	return files, nil
 }