Page History

...

Samza jobs produce to a System Stream Partition (SSP) of the destination "system". This SystemProducer maps the Azure container to the System and blob name is formatted as Stream/Partition/timestamp. 

The flow of data is as follows.

Samza SystemProducers get the message as an OutgoingMessageEnvelope which has an SSP. This producer, uses those fields to identify the Azure container and form a blob. The partition field of the envelope is optional and in this case the blob name is formatted as Stream/timestamp. If two tasks produce to the same SSP then two different blobs will be created with the same prefix of stream/partition but different timestamps. However, its recommended to turn on the config to add a random string suffix to avoid blob name collisions.

For example a blob with SSP: system = "oss-testcontainer", stream = "OSSWikipediaFeedStreamEvent" and partition = empty gives blob name as below

Similarly, with SSP: system = "oss-testcontainer", stream = "OSSWikipediaFeedStreamEvent" and partition = "partition-0" gives blob name as below

The flow of data is as follows.

1. Messages to be sent to Azure Blob Storage are buffered in-memory. 

2. Every time the buffer reaches the maxBlock

3. Messages to be sent to Azure Blob Storage are buffered in-memory. 

4. Every time the buffer reaches the maxBlock size, the block is uploaded to Azure Blob Storage asynchronously using stageBlock API provided by Azure.  

5. List of all blocks uploaded for a blob is maintained in memory. 

6. During flush of SystemProducer, for each stream-partition, remaining buffer is uploaded as a new block and the block list is committed which creates the Block Blob with list of blocks uploaded using the commitBlockList API provided by Azure. 

7. Committing block list has to be a blocking call to wait for all pending block uploads of the blob to finish before the commit as this is expected by Azure Blob Storage. 

8. Messages sent through the SystemProducer after a flush, are part of a new blob. Hence timestamp is added to the blob name as there could be multiple blobs (one per flush of SystemProducer) for the same SSP. This timestamp corresponds to the time when the first message is received by the SystemProducer after a commit.

9. Optionally, a random string can be suffixed to the blob name to avoid blob name collisions for cases when two tasks write to the same SSP at the same time. 

...

1. Source: is the source of the message. It indicates the origin of this message.

1. Example: for messages coming from Samza job’s task source is the task name, for coordinator messages source is “SamzaContainer”, for checkpoint messages source is task name, for diagnostics messges it is “DiagnosticManager”, for metrics related messages it is either “ApplicationMaster” or “StreamProcessor” and so on.

2. ome.systemStream: 

1. SystemStream.system: Azure container name. Azure account name is provided as a config.

2. SystemStream.stream: Azure blob name prefix.

3. ome.partition : empty or partition. If not empty, then blob name is “ome.stream/ome.partition/timestamp” else it is “ome.stream/timestamp". 

4. Example: to send message to SSP system = Azure container name = “azure-container”, Stream = “azure-stream”, the blob with name “azure-stream/..” is created in Azure container “azure-container”. The blob can be found at https://<azure-account-name>.blob.core.windows.net/azure-container/azure-stream/.../timestamp where the blob name itself is azure-stream/../timestamp

3. ome.keySerializerName: no-op serde

4. ome.messageSerializerName: no-op serde

5. Ome.message: supports a single message. 

6. Ome.key: this will be ignoredkey for the message. 

SystemProducer.flush(String source): flush for all the streams associated with that source. During flush, the SystemProducer waits till all the pending blocks for all streams are asynchronously uploaded and then blobs are created. All messages coming after a flush will go into a new blob.

...

This SystemProducer allows the following properties to be configurable.

Implementation / Test Plan 

...

public interface AzureBlobWriter {
	/**
	* write given message envelope to the blob opened
	*/
	void write(OutgoingMessageEnvelope messageEnvelope);

	/** 
	* Asynchronously upload messages written as a blob. 
	* Messages written after this go as a new block.
	*/
	void flush();
 
	/**
	* Close the writer and all of its underlying components.
	* At the end of close, all the messages sent to the writer should be persisted in a blob.
	* flush should be called explicitly before close.
	* It is not the responsibility of close to upload blocks.
	* After close, no other operations can be performed.
	*/
	void close();	
}

AvroAzureBlobWriter is an AzureBlobWriter implementation to write messages which are Avro records.

...

AvroAzureBlobWriter is an AzureBlobWriter implementation to write messages which are Avro records.

1. Schema of the incoming Avro records is expected to be the same for all messages between two flushes of the SystemProducer. This is because an Avro file expects all of its entries to conform to a schema and here blob is an Avro file.
2. OutgoingMessageEnvelope.key will be ignored here. This is because the Avro blobs created consist of only the schema taken from the OutgoingMessageEnvelope.message.getSchema and all the messages which are Avro records. There is no way to insert a key into the records as that would make the record incompatible with the schema. Other implementations of AzureBlobWriter can utilize the key in a meaning fashion.

public class AvroAzureBlobWriter implements AzureBlobWriter { 
	// Avro's DataFileWriter is used to write the incoming Avro records. The sink for DataFileWriter is AzureBlobOutputStream.
    /**
     * write the message to blob via DataFileWriter which writes to AzureBlobOutputStream
     */
	public void write(OutgoingMessageEnvelope messageEnvelope) {
	}
    
	/**
     * flush the contents of DataFileWriter and underlying AzureBlobOutputStream 
     * this operation uploads a block.
     */
	public void flush() {
	}

   /**
    * Close the DataFileWriter and underlying AzureBlobOutputStream
    * this operation commits the blob with all blocks uploaded so far.
    */
	public void close() {
	}
}

...