xmlaminar

Utilities for working with streaming XML pipelines

Background and motivation

XML is a widely used format for information serialization, transfer, and processing. Unfortunately, because of its document-centric nature, XML is unnecessarily inconvenient to work with in contexts where content consists of multiple independent records, and where streaming and/or fault-tolerant processing are desired. In practice, records are usually grouped together arbitrarily into documents for transfer and processing; this approach often necessitates workarounds to approximate streaming and fault-tolerance, and pointlessly increases the complexity and inefficiency of many workflows involving XML.

Functionality

xmlaminar defines and implements a clean java and command-line API, providing a flexible level of abstraction for managing the document-based nature of XML representations of data. Functionality enabled includes:

1. Joining multiple XML streams together
2. Splitting XML streams into smaller chunks
3. Fault-tolerant, order-preserving parallel processing of arbitrarily large XML streams (taking advantage of multiple cores to speed processing)
4. Streaming XML representation of information from databases (including MARC->MARCXML)
5. Integration of multiple corresponding XML streams into a single hierarchical XML stream
6. Configurable modules for handling XML output (write to file, stdout, POST to URL, etc.)

Design considerations

Main design considerations include:

1. Transparency, accessibility, and flexibility of configuration and use
2. Ease of extensibility (for supporting new functionality)
3. Emphasis on efficiency (memory and CPU) and support for streaming use cases

Build and use

Building

Build using Maven (i.e., mvn clean install from top-level project); all dependencies should be available in Maven Central, with the exception of a runtime dependency on the Oracle JDBC driver (see below). Top level pom file builds all modules. Command-line runnable jar file is generated at: cli/dist/xmlaminar-cli-[version]-jar-with-dependencies.jar

JDBC driver dependency

For local/historical reasons, there is an Oracle/JDBC bias in the database-oriented portion of the codebase. This dependency should be refactored out, but for now the dependency is referred to in core/pom.xml as [email protected] Version 10.2.0.5 is used because it is the most recent driver that maintains compatibility with legacy Oracle database versions (8.x?). This dependency must be fetched manually (currently@2015-11-12 from Oracle, and added to local maven repo according to instructions from Maven, i.e.:

mvn install:install-file -Dfile=<path-to-file> -DgroupId=com.oracle \
        -DartifactId=ojdbc4 -Dversion=10.2.0.5 -Dpackaging=jar

Use

Command-line help

Referring to the jar-with-dependencies as xmlaminar.jar, top-level cli help may be accessed via:

[user@host]$ java -jar xmlaminar.jar --help
For help with a specific command: 
        --command --help
available commands: 
        [process, integrate, input, marcdb-to-xml, pipeline, tee, config, 
        split, join, output, db-to-xml]

As indicated in the output for the top-level --help command, many (though not all) of the available commands have their sub-options documented and accessible via java -jar xmlaminar.jar --input --help, e.g.:

[user@host]$ java -jar xmlaminar.jar --input --help
Option                              Description
------                              -----------
-0, --from0                         indirect input file null-delimited
-d, --input-delimiter               directly specify input delimiter
                                      (default:
                                    )
--files-from <File: '-' for stdin>  indirect input
-h, --help                          show help
-i, --input <File: '-' for stdin>   input; default to stdin if no --files-from,
                                      otherwise CWD
-v, --verbose                       be more verbose

Command-line syntax

Command-line syntax is designed to enable pipelineing within a single JVM. Syntax was to some degree inspired by the robustness and configurability of the find and rsync commands, and the fluency of the xrandr command-line utility.

Command-line structure follows the pattern:

java -jar xmlaminar.jar [[--command] [--option [arg]]*]+

Each command is configured by 0 or more options (some of which accept arguments), and output from each command is "piped" (as SAX events) as input to the subsequent command. As a simple example, the following command takes a list of similarly-formatted XML files on stdin, joins them all together into one large XML file, splits them into documents containing 10 records each, and writes each of the resulting documents as gzipped files with incrementing file names (/tmp/out.xml-00000.gz,-00001.gz,-00002.gz...)

find . -name '*.xml' -print0 | \
  java -jar xmlaminar.jar --input --from0 --files-from - \
                          --join -a \
                          --split -n 10 \
                          --output -z -b /tmp/out.xml

Note that when working with large amounts of XML, the ability to write directly to gzipped output files can save a significant amount of disk space and i/o.

Input

There are three main types of input that can be specified (including on stdin):

1. Direct (XML content)
2. Indirect (list of Files/URIs containing XML content; compare rsync "files-from" option)
3. Parameters (for commands that support generation of XML content based on input parameters)

Of these three, the first two should be fairly intuitive. The third case is designed to be used when XML output is generated as a linear function of a stream of input parameters. In a way input type 2 is a special case of input type 3, where the "parameters" consist of resource references that define the generated output XML.

The motivating use case for input type 3 was the *db-to-xml-type commands, where a parameterized SQL query is used to configure an XMLReader, which dynamically generates XML as specified by a stream of input parameters. For instance, for pseudo-SQL configuration query SELECT * FROM TABLE WHERE ID IN (<\INTEGER*\>), if we pass in delimited stream of a million ids (cat ids.txt | java -jar xmlaminar.jar [...]), we would expect to receive output analogous to:

<root>
  <record id="[first-id]">
    <col1>val</col1>
    <col2>val</col2>
  </record>
  ...
  <record id="[millionth-id]">
    <col1>val</col1>
    <col2>val</col2>
  </record>
</root>

To facilitate JVM-internal pipelining of the generation and consumption of input-type-3 parameter streams, the pipeline automatically pivots according to each command's expected input type, so that the following command works as one might intuitively hope (generating flat XML containing all columns for each row that has been updated in the last half day):

echo "0.5" | java -jar xmlaminar.jar \
     --db-to-xml --sql 'SELECT ID FROM TABLE WHERE LAST_UPDATE > SYSDATE - <\DECIMAL\>' \
     --db-to-xml --sql 'SELECT * FROM TABLE WHERE ID IN (<\INTEGER*\>)'

Configuration options

Most commands (excluding, e.g., the integrate command) can be fully configured directly on the command line. However, to do so would quickly become unwieldy, because of:

1. The profusion of options for some commands
2. The usefulness of chaining (potentially very) many commands together

Also, the inherent composability of pipelines would make it very useful to define a modular alias for a frequently-used configuration of a command (or pipeline of commands).

These concerns are addressed by the config pseudo-command, whose single XML-configuration-file argument bundles command and pipeline configurations together into a single modular component that may be used as part of other pipelines or commands -- either on the command-line, or nested within other configuration files.

The below configuration file (invoked echo 20 | java -jar xmlaminar.jar --config pipeline.xml reads ids for records modified during the past 20 days, and uses them to retrieve associated marc records as MARCXML, group them into documents containing 1000 records each (plus remainder), and write the resulting documents out (gzipped) as out/marc-00000.xml.gz, out/marc-00001.xml.gz, etc.

pipeline.xml:

<config:source type="pipeline" xmlns:config="http://library.upenn.edu/xmlaminar/config">
  <config:source type="config">
    incremental.xml
  </config:source>
  <config:filter type="config">
    records.xml
  </config:filter>
  <config:filter type="split">
    chunk-size=1000
  </config:filter>
  <config:filter type="output">
    output-extension=.xml
    output-basename=out/marc
    gzip=true
  </config:filter>
</config:source>

incremental.xml:

<config:source type="db-to-xml" xmlns:config="http://library.upenn.edu/xmlaminar/config">
  connection-config-file=connection.properties
  id-field-labels=BIB_ID
  <config:property name="sql">
SELECT DISTINCT BIB_ID
FROM BIB_HISTORY
WHERE ACTION_DATE > SYSDATE - &lt;\DECIMAL\&gt;
ORDER BY 1
  </config:property>
</config:source>

records.xml:

<config:source type="marcdb-to-xml" xmlns:config="http://library.upenn.edu/xmlaminar/config">
  connection-config-file=connection.properties
  marc-binary-field-label=RECORD_SEGMENT
  id-field-labels=BIB_ID
  <config:property name="sql">
SELECT BIB_DATA.BIB_ID, BIB_DATA.SEQNUM, BIB_DATA.RECORD_SEGMENT
FROM PENNDB.BIB_DATA, BIB_MASTER
WHERE BIB_DATA.BIB_ID = BIB_MASTER.BIB_ID
  AND BIB_DATA.BIB_ID in (&lt;\INTEGER*\&gt;)
ORDER BY 1,2
  </config:property>
</config:source>

Note that properties specified in the config may be contained directly as top-level text content (which is parsed as a Java properties file) or within a dedicated <config:source/> element. This allows the terseness of a simple properties file to coexist alongside the richness and explicitness of XML, without sacrificing clarity or introducing ambiguity.

Parallel, fault-tolerant XSLT processing

The process command is one of the most powerful xmlaminar commands. Its handful of options allow for XML streams (consisting of independent records) to be parsed and processed in parallel, greatly speeding what has traditionally been a fragile, fundamentally serial task.

Option                       Description
------                       -----------
-r, --split-depth <Integer>  set record element depth (default: 1)
--record-xpath               xpath specifying record id location
-s, --subdivide-on-failure   define behavior on processing failure
-x, --xsl <File>             xsl file defining processing templates

By passing a list of files to be processed, or by invoking the split command before the process command, chunks of records will be parsed serially, processed in parallel, and output corresponding to input order. If the subdivide-on-failure flag is specified, a failure during processing of a chunk of records will cause the chunk to be recursively subdivided (according to the split-depth option) and re-processed, isolating the individual problem record, logging its id (according to the setting of the record-xpath option), and allowing the remainder of the records to complete the transformation successfully.

Output files corresponding to input

In the common use case of transforming a directory full of XML files and writing output files in an analogous directory structure, xmlaminar may be configured to behave like "'rsync --files-from' plus transformation".

The following command takes all *.xml files in the input/ directory, splits them to ensure that XSLT processing isn't attempted on chunks of more than 1000 records, processes the resulting chunks according to transform.xsl (subdividing on transformation failure and logging failed record ids), joins successfully transformed chunks back together (grouping by input systemId), and writes the resulting transformed documents (gzipped) to a path rooted in the output/ directory, analogous to the path of the corresponding input document (as rooted in the input/ directory):

find input/ -name '*.xml' | \
  java -jar xmlaminar.jar --input -i input/ --files-from - --split -n 1000 \
       --process -s -x transform.xsl --record-xpath '/root/record/@id' 
       --join --output -z -b output/

Extensions

Two main approaches are available for extending the functionality of xmlaminar:

1. Depend on xmlaminar-cli and package your own additional functionality along with all dependencies in a single jar file, with the unmodified edu.upenn.library.xmlaminar.cli.Driver main class, or
2. Package your plugin in a thin jar file (with compile-time dependencies on xmlaminar-cli) and invoke the stock xmlaminar jar file with a --plugins <colon-separated-paths-to-plugin-jar-files> initial argument.

Both approaches allow you to take advantage of the command line arg parsing logic and framework, making functionality (including help messages) available directly on the command-line.

Several steps should be taken when extending functionality with either of the above approaches:

1. Include one or more edu.upenn.library.xmlaminar.xli.CommandFactory instances to define command-line arguments and certain aspects of command behavior.
   • your CommandFactory should include a static initializer block to register itself with the top-level command-line parser:

        class MyCommandFactory extends CommandFactory {
            static {
                registerCommandFactory(new MyCommandFactory());
            }

            @Override
            public String getKey() {
                return "mycommand";
            }
        }

2. Include one or more instances of org.xml.sax.helpers.XMLFilterImpl, or edu.upenn.library.xmlaminar.parallel.QueueSourceXMLFilter if you wish to support multi-input; optionally extending edu.upenn.library.xmlaminar.parallel.callback.OutputCallback if you wish to support multi-output.
3. If you wish to support the second (i.e., dynamic/runtime) approach to extending functionality, you should take the extra step of enumerating desired CommandFactory instances in a special classpath resource file (which you would normally bundl in your plugin/extension jar file): edu/upenn/library/xmlaminar/cli/load-external-command-factories.ini. Each non-empty, non-whitespace, non-#-prefixed line in each such file will be parsed as a full-qualified Java class name, and an attempt will be made to load these classes by name at runtime. On class load, each named CommandFactory should (in its static initializer block) register an instance of itself with the main command-line parser, making the added functionality directly available on the command line.
   • The xmlaminar-solr module (which defines a command that unmarshals SAX events representing solr input documents into Java SolrInputDocument objects and posts them in parallel to a solr server via a ConcurrentUpdateSolrServer) is an example of this approach. The benefit of this dynamic/runtime approach is that it allows users to swap in specialized modular extensions of functionality without requiring them to compile and distribute custom-modified versions of the main code.