Rets-Io/docs/asciidoc/common-patterns.adoc

728 lines
25 KiB
Plaintext

:batch-asciidoc: ./
:toc: left
:toclevels: 4
[[commonPatterns]]
== Common Batch Patterns
ifndef::onlyonetoggle[]
include::toggle.adoc[]
endif::onlyonetoggle[]
Some batch jobs can be assembled purely from off-the-shelf components in Spring Batch.
For instance, the `ItemReader` and `ItemWriter` implementations can be configured to
cover a wide range of scenarios. However, for the majority of cases, custom code must be
written. The main API entry points for application developers are the `Tasklet`, the
`ItemReader`, the `ItemWriter`, and the various listener interfaces. Most simple batch
jobs can use off-the-shelf input from a Spring Batch `ItemReader`, but it is often the
case that there are custom concerns in the processing and writing that require developers
to implement an `ItemWriter` or `ItemProcessor`.
In this chapter, we provide a few examples of common patterns in custom business logic.
These examples primarily feature the listener interfaces. It should be noted that an
`ItemReader` or `ItemWriter` can implement a listener interface as well, if appropriate.
[[loggingItemProcessingAndFailures]]
=== Logging Item Processing and Failures
A common use case is the need for special handling of errors in a step, item by item,
perhaps logging to a special channel or inserting a record into a database. A
chunk-oriented `Step` (created from the step factory beans) lets users implement this use
case with a simple `ItemReadListener` for errors on `read` and an `ItemWriteListener` for
errors on `write`. The following code snippet illustrates a listener that logs both read
and write failures:
[source, java]
----
public class ItemFailureLoggerListener extends ItemListenerSupport {
private static Log logger = LogFactory.getLog("item.error");
public void onReadError(Exception ex) {
logger.error("Encountered error on read", e);
}
public void onWriteError(Exception ex, List<? extends Object> items) {
logger.error("Encountered error on write", ex);
}
}
----
Having implemented this listener, it must be registered with a step, as shown in the
following example:
.XML Configuration
[source, xml, role="xmlContent"]
----
<step id="simpleStep">
...
<listeners>
<listener>
<bean class="org.example...ItemFailureLoggerListener"/>
</listener>
</listeners>
</step>
----
.Java Configuration
[source, java, role="javaContent"]
----
@Bean
public Step simpleStep() {
return this.stepBuilderFactory.get("simpleStep")
...
.listener(new ItemFailureLoggerListener())
.build();
}
----
IMPORTANT: if your listener does anything in an `onError()` method, it must be inside
a transaction that is going to be rolled back. If you need to use a transactional
resource, such as a database, inside an `onError()` method, consider adding a declarative
transaction to that method (see Spring Core Reference Guide for details), and giving its
propagation attribute a value of `REQUIRES_NEW`.
[[stoppingAJobManuallyForBusinessReasons]]
=== Stopping a Job Manually for Business Reasons
Spring Batch provides a `stop()` method through the `JobLauncher` interface, but this is
really for use by the operator rather than the application programmer. Sometimes, it is
more convenient or makes more sense to stop a job execution from within the business
logic.
The simplest thing to do is to throw a `RuntimeException` (one that is neither retried
indefinitely nor skipped). For example, a custom exception type could be used, as shown
in the following example:
[source, java]
----
public class PoisonPillItemProcessor<T> implements ItemProcessor<T, T> {
@Override
public T process(T item) throws Exception {
if (isPoisonPill(item)) {
throw new PoisonPillException("Poison pill detected: " + item);
}
return item;
}
}
----
Another simple way to stop a step from executing is to return `null` from the
`ItemReader`, as shown in the following example:
[source, java]
----
public class EarlyCompletionItemReader implements ItemReader<T> {
private ItemReader<T> delegate;
public void setDelegate(ItemReader<T> delegate) { ... }
public T read() throws Exception {
T item = delegate.read();
if (isEndItem(item)) {
return null; // end the step here
}
return item;
}
}
----
The previous example actually relies on the fact that there is a default implementation
of the `CompletionPolicy` strategy that signals a complete batch when the item to be
processed is `null`. A more sophisticated completion policy could be implemented and
injected into the `Step` through the `SimpleStepFactoryBean`, as shown in the following
example:
.XML Configuration
[source, xml, role="xmlContent"]
----
<step id="simpleStep">
<tasklet>
<chunk reader="reader" writer="writer" commit-interval="10"
chunk-completion-policy="completionPolicy"/>
</tasklet>
</step>
<bean id="completionPolicy" class="org.example...SpecialCompletionPolicy"/>
----
.Java Configuration
[source, java, role="javaContent"]
----
@Bean
public Step simpleStep() {
return this.stepBuilderFactory.get("simpleStep")
.<String, String>chunk(new SpecialCompletionPolicy())
.reader(reader())
.writer(writer())
.build();
}
----
An alternative is to set a flag in the `StepExecution`, which is checked by the `Step`
implementations in the framework in between item processing. To implement this
alternative, we need access to the current `StepExecution`, and this can be achieved by
implementing a `StepListener` and registering it with the `Step`. The following example
shows a listener that sets the flag:
[source, java]
----
public class CustomItemWriter extends ItemListenerSupport implements StepListener {
private StepExecution stepExecution;
public void beforeStep(StepExecution stepExecution) {
this.stepExecution = stepExecution;
}
public void afterRead(Object item) {
if (isPoisonPill(item)) {
stepExecution.setTerminateOnly();
}
}
}
----
When the flag is set, the default behavior is for the step to throw a
`JobInterruptedException`. This behavior can be controlled through the
`StepInterruptionPolicy`. However, the only choice is to throw or not throw an exception,
so this is always an abnormal ending to a job.
[[addingAFooterRecord]]
=== Adding a Footer Record
Often, when writing to flat files, a "footer" record must be appended to the end of the
file, after all processing has be completed. This can be achieved using the
`FlatFileFooterCallback` interface provided by Spring Batch. The `FlatFileFooterCallback`
(and its counterpart, the `FlatFileHeaderCallback`) are optional properties of the
`FlatFileItemWriter` and can be added to an item writer as shown in the following
example:
.XML Configuration
[source, xml, role="xmlContent"]
----
<bean id="itemWriter" class="org.spr...FlatFileItemWriter">
<property name="resource" ref="outputResource" />
<property name="lineAggregator" ref="lineAggregator"/>
<property name="headerCallback" ref="headerCallback" />
<property name="footerCallback" ref="footerCallback" />
</bean>
----
.Java Configuration
[source, java, role="javaContent"]
----
@Bean
public FlatFileItemWriter<String> itemWriter(Resource outputResource) {
return new FlatFileItemWriterBuilder<String>()
.name("itemWriter")
.resource(outputResource)
.lineAggregator(lineAggregator())
.headerCallback(headerCallback())
.footerCallback(footerCallback())
.build();
}
----
The footer callback interface has just one method that is called when the footer must be
written, as shown in the following interface definition:
[source, java]
----
public interface FlatFileFooterCallback {
void writeFooter(Writer writer) throws IOException;
}
----
[[writingASummaryFooter]]
==== Writing a Summary Footer
A common requirement involving footer records is to aggregate information during the
output process and to append this information to the end of the file. This footer often
serves as a summarization of the file or provides a checksum.
For example, if a batch job is writing `Trade` records to a flat file, and there is a
requirement that the total amount from all the `Trades` is placed in a footer, then the
following `ItemWriter` implementation can be used:
[source, java]
----
public class TradeItemWriter implements ItemWriter<Trade>,
FlatFileFooterCallback {
private ItemWriter<Trade> delegate;
private BigDecimal totalAmount = BigDecimal.ZERO;
public void write(List<? extends Trade> items) throws Exception {
BigDecimal chunkTotal = BigDecimal.ZERO;
for (Trade trade : items) {
chunkTotal = chunkTotal.add(trade.getAmount());
}
delegate.write(items);
// After successfully writing all items
totalAmount = totalAmount.add(chunkTotal);
}
public void writeFooter(Writer writer) throws IOException {
writer.write("Total Amount Processed: " + totalAmount);
}
public void setDelegate(ItemWriter delegate) {...}
}
----
This `TradeItemWriter` stores a `totalAmount` value that is increased with the `amount`
from each `Trade` item written. After the last `Trade` is processed, the framework calls
`writeFooter`, which puts the `totalAmount` into the file. Note that the `write` method
makes use of a temporary variable, `chunkTotal`, that stores the total of the
`Trade` amounts in the chunk. This is done to ensure that, if a skip occurs in the
`write` method, the `totalAmount` is left unchanged. It is only at the end of the `write`
method, once we are guaranteed that no exceptions are thrown, that we update the
`totalAmount`.
In order for the `writeFooter` method to be called, the `TradeItemWriter` (which
implements `FlatFileFooterCallback`) must be wired into the `FlatFileItemWriter` as the
`footerCallback`. The following example shows how to do so:
.XML Configuration
[source, xml, role="xmlContent"]
----
<bean id="tradeItemWriter" class="..TradeItemWriter">
<property name="delegate" ref="flatFileItemWriter" />
</bean>
<bean id="flatFileItemWriter" class="org.spr...FlatFileItemWriter">
<property name="resource" ref="outputResource" />
<property name="lineAggregator" ref="lineAggregator"/>
<property name="footerCallback" ref="tradeItemWriter" />
</bean>
----
.Java Configuration
[source, java, role="javaContent"]
----
@Bean
public TradeItemWriter tradeItemWriter() {
TradeItemWriter itemWriter = new TradeItemWriter();
itemWriter.setDelegate(flatFileItemWriter(null));
return itemWriter;
}
@Bean
public FlatFileItemWriter<String> flatFileItemWriter(Resource outputResource) {
return new FlatFileItemWriterBuilder<String>()
.name("itemWriter")
.resource(outputResource)
.lineAggregator(lineAggregator())
.footerCallback(tradeItemWriter())
.build();
}
----
The way that the `TradeItemWriter` has been written so far functions correctly only if
the `Step` is not restartable. This is because the class is stateful (since it stores the
`totalAmount`), but the `totalAmount` is not persisted to the database. Therefore, it
cannot be retrieved in the event of a restart. In order to make this class restartable,
the `ItemStream` interface should be implemented along with the methods `open` and
`update`, as shown in the following example:
[source, java]
----
public void open(ExecutionContext executionContext) {
if (executionContext.containsKey("total.amount") {
totalAmount = (BigDecimal) executionContext.get("total.amount");
}
}
public void update(ExecutionContext executionContext) {
executionContext.put("total.amount", totalAmount);
}
----
The update method stores the most current version of `totalAmount` to the
`ExecutionContext` just before that object is persisted to the database. The open method
retrieves any existing `totalAmount` from the `ExecutionContext` and uses it as the
starting point for processing, allowing the `TradeItemWriter` to pick up on restart where
it left off the previous time the `Step` was run.
[[drivingQueryBasedItemReaders]]
=== Driving Query Based ItemReaders
In the link:readersAndWriters.html[chapter on readers and writers], database input using
paging was discussed. Many database vendors, such as DB2, have extremely pessimistic
locking strategies that can cause issues if the table being read also needs to be used by
other portions of the online application. Furthermore, opening cursors over extremely
large datasets can cause issues on databases from certain vendors. Therefore, many
projects prefer to use a 'Driving Query' approach to reading in data. This approach works
by iterating over keys, rather than the entire object that needs to be returned, as the
following image illustrates:
.Driving Query Job
image::{batch-asciidoc}images/drivingQueryExample.png[Driving Query Job, scaledwidth="60%"]
As you can see, the example shown in the preceding image uses the same 'FOO' table as was
used in the cursor-based example. However, rather than selecting the entire row, only the
IDs were selected in the SQL statement. So, rather than a `FOO` object being returned
from `read`, an `Integer` is returned. This number can then be used to query for the
'details', which is a complete `Foo` object, as shown in the following image:
.Driving Query Example
image::{batch-asciidoc}images/drivingQueryJob.png[Driving Query Example, scaledwidth="60%"]
An `ItemProcessor` should be used to transform the key obtained from the driving query
into a full 'Foo' object. An existing DAO can be used to query for the full object based
on the key.
[[multiLineRecords]]
=== Multi-Line Records
While it is usually the case with flat files that each record is confined to a single
line, it is common that a file might have records spanning multiple lines with multiple
formats. The following excerpt from a file shows an example of such an arrangement:
----
HEA;0013100345;2007-02-15
NCU;Smith;Peter;;T;20014539;F
BAD;;Oak Street 31/A;;Small Town;00235;IL;US
FOT;2;2;267.34
----
Everything between the line starting with 'HEA' and the line starting with 'FOT' is
considered one record. There are a few considerations that must be made in order to
handle this situation correctly:
* Instead of reading one record at a time, the `ItemReader` must read every line of the
multi-line record as a group, so that it can be passed to the `ItemWriter` intact.
* Each line type may need to be tokenized differently.
Because a single record spans multiple lines and because we may not know how many lines
there are, the `ItemReader` must be careful to always read an entire record. In order to
do this, a custom `ItemReader` should be implemented as a wrapper for the
`FlatFileItemReader`, as shown in the following example:
.XML Configuration
[source, xml, role="xmlContent"]
----
<bean id="itemReader" class="org.spr...MultiLineTradeItemReader">
<property name="delegate">
<bean class="org.springframework.batch.item.file.FlatFileItemReader">
<property name="resource" value="data/iosample/input/multiLine.txt" />
<property name="lineMapper">
<bean class="org.spr...DefaultLineMapper">
<property name="lineTokenizer" ref="orderFileTokenizer"/>
<property name="fieldSetMapper" ref="orderFieldSetMapper"/>
</bean>
</property>
</bean>
</property>
</bean>
----
.Java Configuration
[source, java, role="javaContent"]
----
@Bean
public MultiLineTradeItemReader itemReader() {
MultiLineTradeItemReader itemReader = new MultiLineTradeItemReader();
itemReader.setDelegate(flatFileItemReader());
return itemReader;
}
@Bean
public FlatFileItemReader flatFileItemReader() {
FlatFileItemReader<Trade> reader = new FlatFileItemReaderBuilder<>()
.name("flatFileItemReader")
.resource(new ClassPathResource("data/iosample/input/multiLine.txt"))
.lineTokenizer(orderFileTokenizer())
.fieldSetMapper(orderFieldSetMapper())
.build();
return reader;
}
----
To ensure that each line is tokenized properly, which is especially important for
fixed-length input, the `PatternMatchingCompositeLineTokenizer` can be used on the
delegate `FlatFileItemReader`. See
link:readersAndWriters.html#flatFileItemReader[`FlatFileItemReader` in the Readers and
Writers chapter] for more details. The delegate reader then uses a
`PassThroughFieldSetMapper` to deliver a `FieldSet` for each line back to the wrapping
`ItemReader`, as shown in the following example:
.XML Content
[source, xml, role="xmlContent"]
----
<bean id="orderFileTokenizer" class="org.spr...PatternMatchingCompositeLineTokenizer">
<property name="tokenizers">
<map>
<entry key="HEA*" value-ref="headerRecordTokenizer" />
<entry key="FOT*" value-ref="footerRecordTokenizer" />
<entry key="NCU*" value-ref="customerLineTokenizer" />
<entry key="BAD*" value-ref="billingAddressLineTokenizer" />
</map>
</property>
</bean>
----
.Java Content
[source, java, role="javaContent"]
----
@Bean
public PatternMatchingCompositeLineTokenizer orderFileTokenizer() {
PatternMatchingCompositeLineTokenizer tokenizer =
new PatternMatchingCompositeLineTokenizer();
Map<String, LineTokenizer> tokenizers = new HashMap<>(4);
tokenizers.put("HEA*", headerRecordTokenizer());
tokenizers.put("FOT*", footerRecordTokenizer());
tokenizers.put("NCU*", customerLineTokenizer());
tokenizers.put("BAD*", billingAddressLineTokenizer());
tokenizer.setTokenizers(tokenizers);
return tokenizer;
}
----
This wrapper has to be able to recognize the end of a record so that it can continually
call `read()` on its delegate until the end is reached. For each line that is read, the
wrapper should build up the item to be returned. Once the footer is reached, the item can
be returned for delivery to the `ItemProcessor` and `ItemWriter`, as shown in the
following example:
[source, java]
----
private FlatFileItemReader<FieldSet> delegate;
public Trade read() throws Exception {
Trade t = null;
for (FieldSet line = null; (line = this.delegate.read()) != null;) {
String prefix = line.readString(0);
if (prefix.equals("HEA")) {
t = new Trade(); // Record must start with header
}
else if (prefix.equals("NCU")) {
Assert.notNull(t, "No header was found.");
t.setLast(line.readString(1));
t.setFirst(line.readString(2));
...
}
else if (prefix.equals("BAD")) {
Assert.notNull(t, "No header was found.");
t.setCity(line.readString(4));
t.setState(line.readString(6));
...
}
else if (prefix.equals("FOT")) {
return t; // Record must end with footer
}
}
Assert.isNull(t, "No 'END' was found.");
return null;
}
----
[[executingSystemCommands]]
=== Executing System Commands
Many batch jobs require that an external command be called from within the batch job.
Such a process could be kicked off separately by the scheduler, but the advantage of
common metadata about the run would be lost. Furthermore, a multi-step job would also
need to be split up into multiple jobs as well.
Because the need is so common, Spring Batch provides a `Tasklet` implementation for
calling system commands, as shown in the following example:
.XML Configuration
[source, xml, role="xmlContent"]
----
<bean class="org.springframework.batch.core.step.tasklet.SystemCommandTasklet">
<property name="command" value="echo hello" />
<!-- 5 second timeout for the command to complete -->
<property name="timeout" value="5000" />
</bean>
----
.Java Configuration
[source, java, role="javaContent"]
----
@Bean
public SystemCommandTasklet tasklet() {
SystemCommandTasklet tasklet = new SystemCommandTasklet();
tasklet.setCommand("echo hello");
tasklet.setTimeout(5000);
return tasklet;
}
----
[[handlingStepCompletionWhenNoInputIsFound]]
=== Handling Step Completion When No Input is Found
In many batch scenarios, finding no rows in a database or file to process is not
exceptional. The `Step` is simply considered to have found no work and completes with 0
items read. All of the `ItemReader` implementations provided out of the box in Spring
Batch default to this approach. This can lead to some confusion if nothing is written out
even when input is present (which usually happens if a file was misnamed or some similar
issue arises). For this reason, the metadata itself should be inspected to determine how
much work the framework found to be processed. However, what if finding no input is
considered exceptional? In this case, programmatically checking the metadata for no items
processed and causing failure is the best solution. Because this is a common use case,
Spring Batch provides a listener with exactly this functionality, as shown in
the class definition for `NoWorkFoundStepExecutionListener`:
[source, java]
----
public class NoWorkFoundStepExecutionListener extends StepExecutionListenerSupport {
public ExitStatus afterStep(StepExecution stepExecution) {
if (stepExecution.getReadCount() == 0) {
return ExitStatus.FAILED;
}
return null;
}
}
----
The preceding `StepExecutionListener` inspects the `readCount` property of the
`StepExecution` during the 'afterStep' phase to determine if no items were read. If that
is the case, an exit code of FAILED is returned, indicating that the `Step` should fail.
Otherwise, `null` is returned, which does not affect the status of the `Step`.
[[passingDataToFutureSteps]]
=== Passing Data to Future Steps
It is often useful to pass information from one step to another. This can be done through
the `ExecutionContext`. The catch is that there are two `ExecutionContexts`: one at the
`Step` level and one at the `Job` level. The `Step` `ExecutionContext` remains only as
long as the step, while the `Job` `ExecutionContext` remains through the whole `Job`. On
the other hand, the `Step` `ExecutionContext` is updated every time the `Step` commits a
chunk, while the `Job` `ExecutionContext` is updated only at the end of each `Step`.
The consequence of this separation is that all data must be placed in the `Step`
`ExecutionContext` while the `Step` is executing. Doing so ensures that the data is
stored properly while the `Step` runs. If data is stored to the `Job` `ExecutionContext`,
then it is not persisted during `Step` execution. If the `Step` fails, that data is lost.
[source, java]
----
public class SavingItemWriter implements ItemWriter<Object> {
private StepExecution stepExecution;
public void write(List<? extends Object> items) throws Exception {
// ...
ExecutionContext stepContext = this.stepExecution.getExecutionContext();
stepContext.put("someKey", someObject);
}
@BeforeStep
public void saveStepExecution(StepExecution stepExecution) {
this.stepExecution = stepExecution;
}
}
----
To make the data available to future `Steps`, it must be "promoted" to the `Job`
`ExecutionContext` after the step has finished. Spring Batch provides the
`ExecutionContextPromotionListener` for this purpose. The listener must be configured
with the keys related to the data in the `ExecutionContext` that must be promoted. It can
also, optionally, be configured with a list of exit code patterns for which the promotion
should occur (`COMPLETED` is the default). As with all listeners, it must be registered
on the `Step` as shown in the following example:
.XML Configuration
[source, xml, role="xmlContent"]
----
<job id="job1">
<step id="step1">
<tasklet>
<chunk reader="reader" writer="savingWriter" commit-interval="10"/>
</tasklet>
<listeners>
<listener ref="promotionListener"/>
</listeners>
</step>
<step id="step2">
...
</step>
</job>
<beans:bean id="promotionListener" class="org.spr....ExecutionContextPromotionListener">
<beans:property name="keys">
<list>
<value>someKey</value>
</list>
</beans:property>
</beans:bean>
----
.Java Configuration
[source, java, role="javaContent"]
----
@Bean
public Job job1() {
return this.jobBuilderFactory.get("job1")
.start(step1())
.next(step1())
.build();
}
@Bean
public Step step1() {
return this.stepBuilderFactory.get("step1")
.<String, String>chunk(10)
.reader(reader())
.writer(savingWriter())
.listener(promotionListener())
.build();
}
@Bean
public ExecutionContextPromotionListener promotionListener() {
ExecutionContextPromotionListener listener = new ExecutionContextPromotionListener();
listener.setKeys(new String[] {"someKey" });
return listener;
}
----
Finally, the saved values must be retrieved from the `Job` `ExecutionContext`, as shown
in the following example:
[source, java]
----
public class RetrievingItemWriter implements ItemWriter<Object> {
private Object someObject;
public void write(List<? extends Object> items) throws Exception {
// ...
}
@BeforeStep
public void retrieveInterstepData(StepExecution stepExecution) {
JobExecution jobExecution = stepExecution.getJobExecution();
ExecutionContext jobContext = jobExecution.getExecutionContext();
this.someObject = jobContext.get("someKey");
}
}
----