S3 Select: Workaround java buffer size (#8312)

Updates #7475

The Java implementation has a 128KB buffer and a message must be emitted before that is used. #7475 therefore limits the message size to 128KB. But up to 256 bytes are written to the buffer in each call. This means we must emit a message before shorter than 128KB.

Therefore we change the limit to 128KB minus 256 bytes.

pkg/s3select/message.go

@@ -62,7 +62,12 @@ var recordsHeader = []byte{
 }
 
 const (
-	maxRecordMessageLength = 128 * 1024 // Chosen for compatibility with AWS JAVA SDK
+	// Chosen for compatibility with AWS JAVA SDK
+	// It has a a buffer size of 128K:
+	// https://github.com/aws/aws-sdk-java/blob/master/aws-java-sdk-s3/src/main/java/com/amazonaws/services/s3/internal/eventstreaming/MessageDecoder.java#L26
+	// but we must make sure there is always space to add 256 bytes:
+	// https://github.com/aws/aws-sdk-java/blob/master/aws-java-sdk-s3/src/main/java/com/amazonaws/services/s3/model/SelectObjectContentEventStream.java#L197
+	maxRecordMessageLength = (128 << 10) - 256
 )
 
 var (
@@ -83,7 +88,7 @@ func newRecordsMessage(payload []byte) []byte {
 }
 
 // payloadLenForMsgLen computes the length of the payload in a record
-// message given the length of the message.
+// message given the total length of the message.
 func payloadLenForMsgLen(messageLength int) int {
 	headerLength := len(recordsHeader)
 	payloadLength := messageLength - 4 - 4 - 4 - headerLength - 4