Fault Tolerance, Skip, and Retry Policies

Batch jobs process huge volumes of data, and real-world data is messy. A single malformed row, a momentary database deadlock, or a flaky downstream call can blow up a job that has already processed millions of records.

Spring Batch gives a chunk-oriented step fault tolerance so it can survive these issues without aborting the whole run. The three core tools are:

• Skip — discard records that cause unrecoverable errors (e.g. bad data) and keep going.
• Retry — re-attempt an operation that failed due to a transient error (e.g. a lock timeout).
• Restart — resume a failed job instance from where it stopped instead of starting over.

Used together, these turn a brittle job into a resilient one.

Fault tolerance is opt-in. When you build a chunk step, call .faultTolerant() on the step builder to switch to the fault-tolerant variant. Only then can you declare skip and retry rules.

Without .faultTolerant(), any exception thrown by a reader, processor, or writer rolls back the chunk and fails the step immediately.

Skipping lets the step throw away an individual item that can never succeed — typically a parsing or validation failure — and continue with the next one.

You declare which exceptions are skippable and a global limit:

• .skip(Exception.class) — mark an exception type as skippable.
• .noSkip(Exception.class) — explicitly exclude a subtype from skipping.
• .skipLimit(n) — total number of skips allowed before the step fails.

Once the cumulative skip count exceeds skipLimit, the step aborts. This prevents a job from silently swallowing thousands of bad records.

Skip behaviour depends on where the exception is thrown:

• Reader skip: the bad item is dropped and reading continues — cheap, no rollback.
• Processor skip: the chunk transaction rolls back, then Spring Batch re-processes the chunk item-by-item, skipping only the offending item.
• Writer skip: same scan-and-retry — the chunk rolls back and items are re-written one at a time so the single bad item can be isolated and skipped.

Because processor/writer skips trigger a rollback and a single-item replay, they are far more expensive than reader skips. Keep validation in the reader/processor where possible so failures are caught early.

The declarative .skip()/.skipLimit() API covers most cases, but you can implement SkipPolicy for fully custom logic — for example, allow more skips for one exception type than another, or inspect the exception message.

The shouldSkip method receives the thrown Throwable and the current skip count; return true to skip, or throw SkipLimitExceededException to fail the step.

The decision between skip and retry comes down to the nature of the error:

• Retry a transient error that may succeed if attempted again: deadlock victim, lock timeout, optimistic locking conflict, brief network blip.
• Skip a deterministic error that will always fail: malformed input, a failed business validation, a constraint that the data itself violates.

Retrying a deterministic error just wastes attempts before failing; skipping a transient error throws away data that would have succeeded. Classify your exceptions correctly — this is the key design decision of the lesson.

Retry re-attempts the failing operation up to a configured number of times before giving up. On a fault-tolerant step you declare:

• .retry(Exception.class) — exception types that are retryable.
• .noRetry(Exception.class) — exclude a subtype.
• .retryLimit(n) — maximum attempts per item (including the first try).

When an item fails with a retryable exception, the chunk transaction rolls back and the item is replayed up to retryLimit times. If it still fails, the exception propagates — at which point it may be skipped if it is also declared skippable.

Hammering a contended resource with immediate retries often makes contention worse. A backoff policy inserts a delay between attempts. ExponentialBackOffPolicy grows the wait multiplicatively, spreading out load.

You attach a custom RetryPolicy or BackOffPolicy via .retryPolicy(...) / by configuring a RetryTemplate. Below, the wait starts at 200ms and doubles each attempt up to 5s.

Silently skipping records is dangerous — you need an audit trail. SkipListener callbacks fire for each skipped item so you can log it, write it to a dead-letter table, or alert.

• onSkipInRead — a read failure was skipped.
• onSkipInProcess — gives you the item and the exception.
• onSkipInWrite — the item that could not be written.

Register it with .listener(skipListener) on the step builder. There is also RetryListener for observing retry attempts.

Skip and retry handle errors within a run; restart handles a run that failed completely. Because Spring Batch persists each step's ExecutionContext and read/write counts in the job repository, relaunching the same JobInstance resumes from the last committed chunk rather than reprocessing everything.

Key rules:

• A JobInstance is identified by its identifying job parameters; reuse them to restart, change them to start a fresh instance.
• Only jobs in a non-COMPLETED state (e.g. FAILED, STOPPED) can be restarted.
• Mark a step .allowStartIfComplete(true) to force already-completed steps to re-run on restart.
• Cap retries with .startLimit(n) so a broken step is not relaunched forever.

A production-grade resilient step combines all three concerns: retry transient failures with backoff, skip deterministic bad data within a bounded limit, audit every skip, and rely on the job repository for restart.

Note the layering: an item that fails is first retried; if it still fails and the exception is skippable, it is skipped (and the listener records it). Exceptions can be both retryable and skippable — retry exhausts first, then skip applies.

Test your understanding of skip vs. retry semantics.

You made a Spring Batch step resilient to transient and deterministic failures:

• Enable tolerance with .faultTolerant() before declaring any skip/retry rules.
• Skip deterministic bad data with .skip() + .skipLimit(); processor/writer skips cost a rollback and single-item replay, so validate early.
• Retry transient errors with .retry() + .retryLimit(), adding an exponential backoff to ease contention.
• Classify carefully: retry transient (deadlock, lock timeout), skip deterministic (parse/validation). When an exception is both, retry runs first, then skip.
• Audit every skip with a SkipListener so nothing disappears silently.
• Restart failed instances from the last committed chunk via the job repository; control re-runs with allowStartIfComplete and startLimit.