There isn't functionality like this that exists for appending data to a file.  You could potentially open a file, read the footer back track to the footer location and start writing again.  This wouldn't be atomic but cold work (likely some lower level API would need to be exposed).  

The other option is to write out an ArrowStream (so no random access) and then start appending new record batches at the end (I would need to check but it is possible there isn't an option to write a new schema to the same file).

TL;DR; This would potentially require some refactoring of APIs.  I'm not sure it is a good idea, what is the motivating use-case?

Thanks,
Micah

On Wed, Feb 10, 2021 at 12:14 AM Priyanshu LNU <t-plnu@microsoft.com> wrote:

Hi,

I am using Apache arrow in Java. I want to append stream of record batches to an already existing file.

That means, I have a fileOutputStream. Using ArrowStreamWriter, I wrote some batches to this fileOutputStream. Now I close the fileOutputStream and then again start this stream to append some batches to it.

The issue is that metadata gets appended again, so the streamreader is not able to read all the batches. So how Can I append stream of recordBatches to an existing file correctly without repetition of metadata in between?

 

Below is part of Java code I am currently using:

 

public void setupWrite(String filename, boolean useCustom) throws Exception {

        File arrowFile = validateFile(filename, false);

        this.fileOutputStream = new FileOutputStream(arrowFile);

        Schema schema = makeSchema();

        this.root = VectorSchemaRoot.create(schema, this.ra);

        DictionaryProvider.MapDictionaryProvider provider = new DictionaryProvider.MapDictionaryProvider();

        this.arrowStreamWriter=new ArrowStreamWriter(root,provider,Channels.newChannel(this.fileOutputStream));

 

        // Just to show some stuff about the schema and layout

        System.out.println("Schema/Layout: ");

        for (Field field : root.getSchema().getFields()) {

            FieldVector vector = root.getVector(field.getName());

            showFieldLayout(field, vector);

        }

        System.out.println("Generated " + this.entries + " data entries , batch size " + batchSize + " usingCustomWriter: " + useCustom + " useNullValues " + this.useNullValues);

   

     // writing logic starts here

//      arrowFileWriter.start();

      arrowStreamWriter.start();

 

      for(int i = 0; i < this.entries;) {

          int toProcessItems = Math.min(this.batchSize, this.entries - i);

          // set the batch row count

          root.setRowCount(toProcessItems);

          for (Field field : root.getSchema().getFields()) {

              FieldVector vector = root.getVector(field.getName());

             // System.out.println(vector.getMinorType());

              switch (vector.getMinorType()) {

                  case INT:

                      writeFieldInt(vector, i, toProcessItems);

                      break;

                  case BIGINT:

                      writeFieldLong(vector, i, toProcessItems);

                      break;

                  case FLOAT4:

                      writeFieldFloat4(vector, i, toProcessItems);

                      break;

                  case VARCHAR:

                      writeFieldVarchar(vector, i, toProcessItems);

                      break;

                  case DATEDAY:

                      writeFieldDate(vector, i, toProcessItems);

                      break;   

               // case VARBINARY:

                 //   writeFieldVarBinary(vector, i, toProcessItems);

                   // break;

                  default:

                      throw new Exception(" Not supported yet type: " + vector.getMinorType());

              }

          }

//          arrowFileWriter.writeBatch();

          arrowStreamWriter.writeBatch();

          i+=toProcessItems;

      }

      arrowStreamWriter.end();

      arrowStreamWriter.close();

      fileOutputStream.flush();

      fileOutputStream.close();

 

     

    

  

      

       

      this.fileOutputStream = new FileOutputStream(arrowFile,true);

      this.arrowStreamWriter=new ArrowStreamWriter(root,provider,this.fileOutputStream.getChannel());

      arrowStreamWriter.start();

      for(int i = 0; i < this.entries;) {

          int toProcessItems = Math.min(this.batchSize, this.entries - i);

          // set the batch row count

          root.setRowCount(toProcessItems);

          for (Field field : root.getSchema().getFields()) {

              FieldVector vector = root.getVector(field.getName());

             // System.out.println(vector.getMinorType());

              switch (vector.getMinorType()) {

                  case INT:

                      writeFieldInt(vector, i, toProcessItems);

                      break;

                  case BIGINT:

                      writeFieldLong(vector, i, toProcessItems);

                      break;

                  case FLOAT4:

                      writeFieldFloat4(vector, i, toProcessItems);

                      break;

                  case VARCHAR:

                      writeFieldVarchar(vector, i, toProcessItems);

                      break;

                  case DATEDAY:

                      writeFieldDate(vector, i, toProcessItems);

                      break;   

               // case VARBINARY:

                 //   writeFieldVarBinary(vector, i, toProcessItems);

                   // break;

                  default:

                      throw new Exception(" Not supported yet type: " + vector.getMinorType());

              }

          arrowStreamWriter.writeBatch();

          i+=toProcessItems;

      }

      arrowStreamWriter.end();

      arrowStreamWriter.close();

      fileOutputStream.flush();

      fileOutputStream.close();

 }

 

 

Thanks,

Priyanshu