You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ignite.apache.org by "Alex Abashev (Jira)" <ji...@apache.org> on 2026/04/28 10:51:00 UTC

[jira] [Updated] (IGNITE-28618) Move finishUnmarshal out of NIO communication thread

     [ https://issues.apache.org/jira/browse/IGNITE-28618?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Alex Abashev updated IGNITE-28618:
----------------------------------
    Description: 
h2. Summary

Receive-side mirror of IGNITE-28520. Same motivation, same mechanism — but for the {{finishUnmarshal}} direction.

The post-deserialisation pass already runs on the user-listener-dispatch thread: {{GridCacheIoManager.unmarshall}} calls {{cacheMsg.finishUnmarshal(cctx, ldr)}} on the dispatch thread before handing the message to {{processMessage}}. Just like on the send side, every {{GridCacheMessage}} subclass with {{CacheObject}} fields has a hand-written {{finishUnmarshal(GridCacheSharedContext, ClassLoader)}} override that walks those fields and calls {{co.finishUnmarshal(ctx, ldr)}}. Same maintenance hazard: a missed override silently shifts the {{Marshaller.unmarshal}} call onto the NIO reader (when something later asks for {{val}} via {{valueBytes(ctx)}}).

This ticket adds the symmetric generated method and auto-invoke so coverage is structural.

h2. What changes

* {{MessageSerializer}} gets {{default void finishUnmarshalCacheObjects(M msg, CacheObjectValueContext ctx, ClassLoader ldr)}} (no-op default).
* {{MessageSerializerGenerator}} emits {{finishUnmarshalCacheObjects}} on every concrete serializer with at least one traversable {{@Order}} field, with recursion into nested {{Message}} types via {{*_SER.finishUnmarshalCacheObjects(...)}}; map traversal walks both keys and values.
* {{GridCacheIoManager.unmarshall}} adds {{finishUnmarshalGeneratedCacheObjects(cacheMsg, ldr)}} *before* the existing {{cacheMsg.finishUnmarshal(cctx, ldr)}} (legacy code may rely on {{val}} being already populated for tx-key hashing / validation, so auto-invoke runs first). Helper gated the same way as the send-side: {{instanceof GridCacheIdMessage}}, context from the outer message's {{cacheId}}.
* Hand-written {{finishUnmarshal(GridCacheSharedContext, ClassLoader)}} overrides are removed where the body was a pure {{@Order}}-CO traversal — same set as the send-side cleanup:
** {{GridCacheTtlUpdateRequest}}, {{GridDhtForceKeysRequest}}, {{GridNearUnlockRequest}}, {{GridDhtUnlockRequest}}, {{GridDistributedLockRequest}}, {{GridDistributedLockResponse}}, {{GridDhtAtomicUpdateResponse}}, {{GridDhtAtomicNearResponse}}, {{GridNearGetRequest}}.
** Also dropped: {{ret.finishUnmarshal(cctx, ldr)}} call inside {{GridNearAtomicUpdateResponse.finishUnmarshal}} (covered by nested-message recursion).
* {{GridCacheReturn.finishUnmarshal}} delegates the {{CacheObject}} traversal to the generated {{GridCacheReturnSerializer.finishUnmarshalCacheObjects}} (mirror of the existing {{prepareMarshal}} bridge); the post-traversal binary-unwrapping ({{unwrapBinaryIfNeeded}} of {{cacheObj}} and {{invokeResCol}}, building the {{CacheInvokeResult}} map into {{v}}) stays inline — it is not part of the serializer contract.

h2. Out of scope

* Send-side migration is IGNITE-28520; this ticket only mirrors it.
* Extending auto-invoke beyond {{GridCacheIdMessage}} for either direction — IGNITE-ZZZZZ.

h2. Acceptance criteria

* {{GridCacheIoManager.unmarshall}} invokes {{finishUnmarshalGeneratedCacheObjects(cacheMsg, ldr)}} for every {{GridCacheIdMessage}} subclass with a non-{{null}} {{cacheObjectContext}}, before the legacy {{finishUnmarshal}} hook.
* {{MessageProcessorTest}} green with regenerated fixtures (each fixture now has both {{prepareMarshalCacheObjects}} and {{finishUnmarshalCacheObjects}}).
* Same targeted-suite gating as IGNITE-28520; in addition: {{IgniteCacheQueryMultiThreadedSelfTest}}, {{GridCacheReturnValueTransferSelfTest}} cover both directions on the same flow.
* TC-bot possible-blockers — no real regressions on read paths.

  was:
h2. Summary

Receive-side mirror of IGNITE-28520. Move the generic post-deserialisation {{CacheObject.finishUnmarshal}} traversal off the NIO worker by adding {{finishUnmarshalCacheObjects}} to {{MessageSerializer}} and auto-invoking it from {{GridCacheIoManager.unmarshall}} on the user-listener-dispatch thread, before the message is handed to {{processMessage}}.

After this lands the receive path looks like:

# {{ser.finishUnmarshalCacheObjects(msg, cacheObjCtx, ldr)}} — new auto-invoked traversal of {{@Order}}-annotated {{CacheObject}} / nested {{Message}} fields;
# {{cacheMsg.finishUnmarshal(cctx, ldr)}} — existing hand-written hook (legacy), runs after.

The same ordering rationale as send-side: legacy may rely on {{val}} being already populated (for tx-key hashing, validation, etc.), so auto-invoke must run first.

h2. Scope

* Add {{default void finishUnmarshalCacheObjects(M msg, CacheObjectValueContext ctx, ClassLoader ldr)}} to {{MessageSerializer}} (no-op default).
* {{MessageSerializerGenerator}}: emit symmetric {{finishUnmarshalCacheObjects}} method (with {{@Override}}) on every serializer with traversable {{@Order}} fields; recurse via per-class {{*_SER.finishUnmarshalCacheObjects(...)}}; map traversal walks both sides (key + value).
* {{GridCacheIoManager.unmarshall}}: add {{finishUnmarshalGeneratedCacheObjects(cacheMsg, ldr)}} call before {{cacheMsg.finishUnmarshal(cctx, ldr)}}; helper gated on {{instanceof GridCacheIdMessage}} mirror of the prepare-side.
* Phase 2 cleanup: where a {{GridCacheMessage}} subclass {{finishUnmarshal(GridCacheSharedContext, ClassLoader)}} override is a pure {{@Order}}-CO traversal, drop it. Mirror list from IGNITE-28520:
** {{GridCacheTtlUpdateRequest}}, {{GridDhtForceKeysRequest}}, {{GridNearUnlockRequest}}, {{GridDhtUnlockRequest}}, {{GridDistributedLockRequest}}, {{GridDistributedLockResponse}}, {{GridDhtAtomicUpdateResponse}}, {{GridDhtAtomicNearResponse}}, {{GridNearGetRequest}}, {{GridNearAtomicUpdateResponse.finishUnmarshal}} — line {{ret.finishUnmarshal(cctx, ldr)}}.
* Update {{GridCacheReturn.finishUnmarshal}} to delegate the {{CacheObject}} traversal to the generated {{GridCacheReturnSerializer.finishUnmarshalCacheObjects}} (mirror of current {{prepareMarshal}} bridge); the post-traversal binary-unwrapping ({{unwrapBinaryIfNeeded}} of {{cacheObj}} and {{invokeResCol}}, building the {{CacheInvokeResult}} map into {{v}}) stays inline — it is not part of the serializer contract.

h2. Out of scope

* The send-side migration is IGNITE-28520; this ticket only mirrors it for the receive side.
* Extending auto-invoke beyond {{GridCacheIdMessage}} for either direction — IGNITE-ZZZZZ.

h2. Acceptance criteria

* {{GridCacheIoManager.unmarshall}} invokes the new helper for {{GridCacheIdMessage}} subclasses, before the legacy {{finishUnmarshal}} call.
* {{MessageProcessorTest}} green with regenerated fixtures (each fixture now has both {{prepareMarshalCacheObjects}} and {{finishUnmarshalCacheObjects}}).
* Same targeted-suite gating as IGNITE-28520; in addition: {{IgniteCacheQueryMultiThreadedSelfTest}}, {{GridCacheReturnValueTransferSelfTest}} cover both directions.


> Move finishUnmarshal out of NIO communication thread
> ----------------------------------------------------
>
>                 Key: IGNITE-28618
>                 URL: https://issues.apache.org/jira/browse/IGNITE-28618
>             Project: Ignite
>          Issue Type: Sub-task
>            Reporter: Alex Abashev
>            Priority: Major
>
> h2. Summary
> Receive-side mirror of IGNITE-28520. Same motivation, same mechanism — but for the {{finishUnmarshal}} direction.
> The post-deserialisation pass already runs on the user-listener-dispatch thread: {{GridCacheIoManager.unmarshall}} calls {{cacheMsg.finishUnmarshal(cctx, ldr)}} on the dispatch thread before handing the message to {{processMessage}}. Just like on the send side, every {{GridCacheMessage}} subclass with {{CacheObject}} fields has a hand-written {{finishUnmarshal(GridCacheSharedContext, ClassLoader)}} override that walks those fields and calls {{co.finishUnmarshal(ctx, ldr)}}. Same maintenance hazard: a missed override silently shifts the {{Marshaller.unmarshal}} call onto the NIO reader (when something later asks for {{val}} via {{valueBytes(ctx)}}).
> This ticket adds the symmetric generated method and auto-invoke so coverage is structural.
> h2. What changes
> * {{MessageSerializer}} gets {{default void finishUnmarshalCacheObjects(M msg, CacheObjectValueContext ctx, ClassLoader ldr)}} (no-op default).
> * {{MessageSerializerGenerator}} emits {{finishUnmarshalCacheObjects}} on every concrete serializer with at least one traversable {{@Order}} field, with recursion into nested {{Message}} types via {{*_SER.finishUnmarshalCacheObjects(...)}}; map traversal walks both keys and values.
> * {{GridCacheIoManager.unmarshall}} adds {{finishUnmarshalGeneratedCacheObjects(cacheMsg, ldr)}} *before* the existing {{cacheMsg.finishUnmarshal(cctx, ldr)}} (legacy code may rely on {{val}} being already populated for tx-key hashing / validation, so auto-invoke runs first). Helper gated the same way as the send-side: {{instanceof GridCacheIdMessage}}, context from the outer message's {{cacheId}}.
> * Hand-written {{finishUnmarshal(GridCacheSharedContext, ClassLoader)}} overrides are removed where the body was a pure {{@Order}}-CO traversal — same set as the send-side cleanup:
> ** {{GridCacheTtlUpdateRequest}}, {{GridDhtForceKeysRequest}}, {{GridNearUnlockRequest}}, {{GridDhtUnlockRequest}}, {{GridDistributedLockRequest}}, {{GridDistributedLockResponse}}, {{GridDhtAtomicUpdateResponse}}, {{GridDhtAtomicNearResponse}}, {{GridNearGetRequest}}.
> ** Also dropped: {{ret.finishUnmarshal(cctx, ldr)}} call inside {{GridNearAtomicUpdateResponse.finishUnmarshal}} (covered by nested-message recursion).
> * {{GridCacheReturn.finishUnmarshal}} delegates the {{CacheObject}} traversal to the generated {{GridCacheReturnSerializer.finishUnmarshalCacheObjects}} (mirror of the existing {{prepareMarshal}} bridge); the post-traversal binary-unwrapping ({{unwrapBinaryIfNeeded}} of {{cacheObj}} and {{invokeResCol}}, building the {{CacheInvokeResult}} map into {{v}}) stays inline — it is not part of the serializer contract.
> h2. Out of scope
> * Send-side migration is IGNITE-28520; this ticket only mirrors it.
> * Extending auto-invoke beyond {{GridCacheIdMessage}} for either direction — IGNITE-ZZZZZ.
> h2. Acceptance criteria
> * {{GridCacheIoManager.unmarshall}} invokes {{finishUnmarshalGeneratedCacheObjects(cacheMsg, ldr)}} for every {{GridCacheIdMessage}} subclass with a non-{{null}} {{cacheObjectContext}}, before the legacy {{finishUnmarshal}} hook.
> * {{MessageProcessorTest}} green with regenerated fixtures (each fixture now has both {{prepareMarshalCacheObjects}} and {{finishUnmarshalCacheObjects}}).
> * Same targeted-suite gating as IGNITE-28520; in addition: {{IgniteCacheQueryMultiThreadedSelfTest}}, {{GridCacheReturnValueTransferSelfTest}} cover both directions on the same flow.
> * TC-bot possible-blockers — no real regressions on read paths.



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