Increase parquet read buffer size (#2924)

Author: aleks-p

pkg/phlaredb/block_querier.go

@@ -49,8 +49,14 @@ import (
 )
 
 const (
-	defaultBatchSize      = 4096
-	parquetReadBufferSize = 256 << 10 // 256KB
+	defaultBatchSize = 4096
+
+	// This controls the buffer size for reads to a parquet io.Reader. This value should be small for memory or
+	// disk backed readers, but when the reader is backed by network storage a larger size will be advantageous.
+	//
+	// The chosen value should be larger than the page size. Page sizes depend on the write buffer size as well as
+	// on how well the data is encoded. In practice, they tend to be around 1MB.
+	parquetReadBufferSize = 2 << 20
 )
 
 type tableReader interface {

pkg/phlaredb/profile_store.go

@@ -26,6 +26,10 @@ import (
 	"github.com/grafana/pyroscope/pkg/util/build"
 )
 
+const (
+	parquetWriteBufferSize = 3 << 20 // 3MB
+)
+
 type profileStore struct {
 	size      atomic.Uint64
 	totalSize atomic.Uint64
@@ -63,7 +67,7 @@ type profileStore struct {
 }
 
 func newParquetProfileWriter(writer io.Writer, options ...parquet.WriterOption) *parquet.GenericWriter[*schemav1.Profile] {
-	options = append(options, parquet.PageBufferSize(3*1024*1024))
+	options = append(options, parquet.PageBufferSize(parquetWriteBufferSize))
 	options = append(options, parquet.CreatedBy("github.com/grafana/pyroscope/", build.Version, build.Revision))
 	options = append(options, schemav1.ProfilesSchema)
 	return parquet.NewGenericWriter[*schemav1.Profile](