[jira] [Commented] (KAFKA-14490) Consider using UncheckedIOException instead of IOException in the log layer

["Canxuan Wang (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/KAFKA-14490?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18082137#comment-18082137 ] 

Canxuan Wang commented on KAFKA-14490:
--------------------------------------

I'd like to take this on. I have measured the current scope on trunk: 209 occurrences of {{throws IOException}} / {{throw new IOException}} across 18 files under {{{}storage/src/main/java/org/apache/kafka/storage/internals/log/{}}}, against only 2 existing usages of {{UncheckedIOException}} in the same package. The raft module already follows this pattern ({{{}FileRawSnapshotReader{}}}, {{{}FileRawSnapshotWriter{}}}, {{{}Snapshots{}}}, {{{}FileQuorumStateStore{}}}, {{{}KafkaRaftLog{}}}), so there is a directly comparable precedent and no KIP needed.

Plan to keep this reviewable:
 * One PR per file or small cluster, each independently mergeable.
 * First PR targets {{{}LazyIndex{}}}, which is called out in the description. Alongside the IOException-to-UncheckedIOException conversion, it folds in the constructor cleanup mentioned in the description: replacing {{(IndexWrapper, long, int, IndexType)}} with {{(IndexWrapper, factoryFunction)}} so {{loadIndex}} no longer needs the {{@SuppressWarnings("unchecked")}} switch on {{{}IndexType{}}}. The PR is small enough to act as a pattern PR for reviewer sign-off before the larger files land.
 * Subsequent PRs walk the remaining 17 files in dependency order: leaf indexes ({{{}AbstractIndex{}}}, {{{}OffsetIndex{}}}, {{{}TimeIndex{}}}, {{{}TransactionIndex{}}}, {{{}SnapshotFile{}}}) first, then {{LogSegment}} / {{LogSegments}} / {{LocalLog}} / {{LogLoader}} / {{UnifiedLog}} / {{LogManager}} / {{Cleaner}} / {{LogCleanerManager}} / {{ProducerStateManager}} / {{RemoteIndexCache}} / {{AsyncOffsetReader}} / {{{}SegmentDeletionReason{}}}.
 * Externally visible method signatures stay declared as-is in the first pass. {{UncheckedIOException}} extends {{RuntimeException}} so dropping a {{throws}} clause is source- and binary-compatible, but I will keep public callers' signatures untouched to limit blast radius. Public-surface cleanup belongs in a separate ticket.

> Consider using UncheckedIOException instead of IOException in the log layer
> ---------------------------------------------------------------------------
>
>                 Key: KAFKA-14490
>                 URL: https://issues.apache.org/jira/browse/KAFKA-14490
>             Project: Kafka
>          Issue Type: Sub-task
>            Reporter: Ismael Juma
>            Assignee: Canxuan Wang
>            Priority: Major
>
> IOException is a checked exception, which makes it difficult to use with lambdas. We should consider using UncheckdIOException instead.
> I'll add notes below with code that could be simplified if we did this:
> 1. The private constructor of LazyIndex could take a factory method instead of all parameters and IndexType (like the Scala code used to). This would avoid some repetition and some unsafe casts.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)