Reader, Processor & Writer

A chunk-oriented step has exactly three pluggable parts: an ItemReader that produces items one at a time, an optional ItemProcessor that transforms or filters each item, and an ItemWriter that persists a whole chunk at once. Spring Batch ships production-grade implementations for files, databases, and queues, so most of the time you configure rather than write them.

ItemReader: one item at a time

An ItemReader<T> returns the next item on each read() call, or null when the source is exhausted - that null is how the step knows to stop:

public interface ItemReader<T> {
    T read();   // next item, or null when done
}

You rarely implement this yourself. For BookVault's overdue report, a JdbcCursorItemReader (or the paging variant) streams active loans straight from the database without loading them all:

@Bean
JdbcPagingItemReader<Loan> loanReader(DataSource ds) {
    return new JdbcPagingItemReaderBuilder<Loan>()
        .name("loanReader")
        .dataSource(ds)
        .selectClause("SELECT id, book_id, member_id, due_date")
        .fromClause("FROM loans")
        .whereClause("WHERE returned_at IS NULL")
        .sortKeys(Map.of("id", Order.ASCENDING))   // stable order = safe restart
        .pageSize(100)
        .rowMapper(new LoanRowMapper())
        .build();
}

Built-ins cover most sources: FlatFileItemReader (CSV/fixed-width), JdbcPagingItemReader / JdbcCursorItemReader, JpaPagingItemReader, StaxEventItemReader (XML), and Kafka readers.

ItemProcessor: transform or filter

The ItemProcessor<I, O> takes one input item and returns the output to be written - or null to filter it out of this run entirely:

public class OverdueProcessor implements ItemProcessor<Loan, ReportRow> {
    public ReportRow process(Loan loan) {
        if (!loan.isOverdue()) {
            return null;              // not overdue → drop it, never written
        }
        long daysLate = DAYS.between(loan.dueDate(), LocalDate.now());
        return new ReportRow(loan.memberId(), loan.bookId(), daysLate);
    }
}

Returning null is the idiomatic way to skip items that don't belong in the output - no exception, no special case. The input type (Loan) and output type (ReportRow) can differ, which is where transformation happens. The processor is optional; a step that only moves data can omit it.

ItemWriter: a chunk at a time

The ItemWriter<T> receives the whole chunk as a list and writes it in one go - the batched write that makes the chunk model efficient:

public interface ItemWriter<T> {
    void write(Chunk<? extends T> items);   // the whole chunk, e.g. 100 rows
}

Because it gets a list, a JdbcBatchItemWriter issues a single batched INSERT for all 100 rows rather than 100 separate statements. Built-ins mirror the readers: FlatFileItemWriter, JdbcBatchItemWriter, JpaItemWriter, and Kafka writers. You can also compose several writers with a CompositeItemWriter.