Document Processing

edit page

This document describes how to develop and deploy Document Processors, often called docproc in this documentation. Document processing is a framework to create chains of configurable components, that read and modify document operations.

The input source splits the input data into logical units called documents. A feeder application sends the documents into a document processing chain. This chain is an ordered list of document processors. Document processing examples range from language detection, HTML removal and natural language processing to mail attachment processing, character set transcoding and image thumbnailing. At the end of the processing chain, extracted data will typically be set in some fields in the document.

The motivation for document processing is that code and configuration is atomically deployed, as like all Vespa components. It is also easy to build components that access data in Vespa as part of processing.

To get started, see the sample application. Read indexing to understand deployment and routing. As document processors are chained components just like Searchers, read Searcher Development. For reference, see the Javadoc, and services.xml.

Deploying a Document Processor

Refer to album-recommendation-docproc to get started, LyricsDocumentProcessor.java is a document processor example. Add the document processor in services.xml, using "default" to include it in the default indexing chain - find more details in indexing:

  <container version="1.0" id="default">
    <document-processing>
      <chain id="default" inherits="indexing">
        <documentprocessor id="ai.vespa.example.album.LyricsDocumentProcessor" bundle="albums-docproc"/>
      </chain>
    </document-processing>
  </container>
</services>

Document Processors

A document processor is a component extending com.yahoo.docproc.DocumentProcessor. All document processors must implement process():

public Progress process(Processing processing);

A worker thread will retrieve a Processing from the input queue, and run its document operations through its call stack. A minimal, no-op document processor implementation is thus:

import com.yahoo.docproc.*;

public class SimpleDocumentProcessor extends DocumentProcessor {
    public Progress process(Processing processing) {
        return Progress.DONE;
    }
}

The process() method should loop through all document operations in Processing.getDocumentOperations(), do whatever it sees fit to them, and return a Progress:

public Progress process(Processing processing) {
    for (DocumentOperation op : processing.getDocumentOperations()) {
        if (op instanceof DocumentPut) {
            DocumentPut put = (DocumentPut) op;
            // TODO do something to 'put here
        } else if (op instanceof DocumentUpdate) {
            DocumentUpdate update = (DocumentUpdate) op;
            // TODO do something to 'update' here
        } else if (op instanceof DocumentRemove) {
            DocumentRemove remove = (DocumentRemove) op;
            // TODO do something to 'remove' here
        }
    }
    return Progress.DONE;
}

if (op instanceof DocumentPut) {
    return Progress.FAILED.withReason("PUT is not supported");
}

Chains

The call stack mentioned above is another name for a document processor chain. Document processor chains are a special case of the general component chains - to avoid confusion some concepts are explained here as well. A document processor chain is nothing more than a list of document processor instances, having an id, and represented as a stack. The document processor chains are typically not created for every processing, but are part of the configuration. Multiple ones may exist at the same time, the chain to execute will be specified by the message bus destination of the incoming message. The same document processor instance may exist in multiple document processor chains, which is why the CallStack of the Processing is responsible for knowing the next document processor to invoke in a particular message.

The execution order of the document processors in a chain are not ordered explicitly, but by ordering constraints declared in the document processors or their configuration.

Execution model

The Document Processing Framework works like this:

1. A thread from the message bus layer appends an incoming message to an internal priority queue, shared between all document processing chains configured on a node. The priority is set based on the message bus priority of the message. Messages of the same priority are ordered FIFO
2. One worker thread from the docproc thread pool picks one message from the head of the queue, deserializes it, copies the call stack (chain) in question, and runs it through the document processors
3. Processing finishes if (a) the document(s) has passed successfully through the whole chain, or (b) a document processor in the chain has returned Progress.FAILED or thrown an exception
4. The same thread passes the message on to the message bus layer for further transport on to its destination

As is evident from the model above, multiple worker threads execute the document processors in a chain concurrently. Thus; many threads of execution can be going through process() in a document processor, at the same time.

This model places a very important constraint on document processor classes: instance variables are not safe. They must be eliminated, or made threadsafe somehow.

Also see Resource management, use deconstruct() in order to not leak resources.

Asynchronous execution

The execution model outlined above also shows one important restriction: If a document processor performs any high-latency operation in its process() method, a docproc worker thread will be occupied. With all n worker threads blocking on an external resource, throughput will be limited. This can be fixed by saving the state in the Processing object, and returning Progress.LATER. A document processor doing a high-latency operation should use a pattern like this:

1. Check a self-defined context variable in Processing for status. Basically, have we seen this Processing before?
2. If no:
   1. We have been given a Processing object fresh off the network, we have not seen this before. Process it up until the high-latency operation.
   2. Start the high-latency operation (possibly in a separate thread).
   3. Save the state of the operation in a self-defined context variable in the Processing.
   4. Return Progress.LATER. This Processing is the appended to the back of the input queue, and we will be called again later.
3. If yes:
   1. Retrieve the reference that we set in our self-defined context variable in Processing
   2. Is the high-latency operation done? If so, return Progress.DONE
   3. Is it not yet done? Return Progress.LATER again

State

Any state in the document processor for the particular Processing should be kept as local variables in the process method, while state which should be shared by all Processings should be kept as member variables. As the latter kind will be accessed by multiple threads at any one time, the state of such member variables must be threadsafe. This critical restriction is similar to those of e.g. the Servlet API. Options for implementing a multithread-safe document processor with instance variables:

1. Use immutable (and preferably final) objects: they never change after they are constructed; no modifications to their state occurs after the DocumentProcessor constructor returns
2. Use a single instance of a thread-safe class
3. Create a single instance and synchronize access to it across all threads (but this will severely limit scalability)
4. Arrange for each thread to have its own instance, e.g. with a ThreadLocal

Processing Context Variables

Processing has a map String -> Object that can be used to pass information between document processors. It is also useful when using Progress.LATER to save the state of a processing - see Processing.java for get/setVariable and more.

The sample application uses such context variables, too.

Operation ordering

Feed ordering

Ordering of feed operations is not guaranteed. Operations on different documents will be done concurrently and hence not ordered. However Vespa guarantees that operations on the same document are processed in the order they were fed if they enter vespa at the same feed endpoint.

Document processing ordering

Document operations that are produced inside a document processor obey the same rules as at feed time. If you either split the input into other documents or into multiple operations to the same document, Vespa will ensure that operations to the same document id are sequenced and are delivered in the order they enter.

(Re)configuring Document Processing

Consider the following configuration:

<?xml version="1.0" encoding="utf-8" ?>
<services version="1.0">
  <container version="1.0" id="default">
    <document-processing>
      <chain id="default">
        <documentprocessor id="SomeDocumentProcessor">
          <config name="foo.something">
            <variable>value</variable>
          </config>
        </documentprocessor>
      </chain>
    </document-processing>
  </container>
</services>

Note when adding or modifying a processing chain in a running cluster; if at the same time deploying a new document processor (i.e. a document processor that was unknown to Vespa at the time the cluster was started), the container must be restarted:

$ vespa-sentinel-cmd restart container

Class diagram

The framework core supports asynchronous processing, processing one or multiple documents or document updates at the same time, document processors that makes dynamic decisions about the processing flow and passing of information between processors outside the document or document update:

• One or more named Docproc Services may be created. One of the services is the default
• A service accepts subclasses of DocumentOperation for processing, meaning DocumentPuts, DocumentUpdates and DocumentRemoves. It has a Call Stack which lists the calls to make to various Document Processors to process each DocumentOperation handed to the service
• Call Stacks consist of Calls, which refer to the Document Processor instance to call
• DocumentPuts and document updates are processed asynchronously, the state is kept in a Processing for its duration (instead of in a thread or process). A Document Processor may make some asynchronous calls (typically to remote services) and return to the framework that it should be called again later for the same Processing to handle the outcome of the calls
• A processing contains its own copy of the Call Stack of the Docproc Service to keep track of what to call next. Document Processors may modify this Call Stack to dynamically decide the processing steps required to process a DocumentOperation
• A Processing may contain one or more DocumentOperations to be processed as a unit
• A Processing has a context, which is a Map of named values which can be used to pass arguments between processors
• Processings are prepared to be stored to disk, to allow a high number of ongoing long-term processings per node