[IGNITE-16640] Peer class loading failure should not be treated as a critical node failure

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.12
• Fix Version/s: 2.13
• Component/s: compute, persistence
• Labels: ise.lts
• Release Note: Fixed peer class loading failure handling
• Ignite Flags: Release Notes Required
• Assignee: Roman Puchkovskiy
• Reporter: Roman Puchkovskiy
• Reviewer: Semyon Danilov
• Created: 03/Mar/22 07:39
• Updated: 13/May/22 09:11
• Resolved: 11/Mar/22 10:23

Description

The problematic scenario:

1. Enable p2p class-loading in the cluster
2. Create a cache
3. From a client, obtain a dataStreamer on that cache and set a receiver() which class is only present on the client (so that it is forced to be loaded using p2p)
4. The receiver class, during its operation (inside its receive() method), should initiate loading of another class only present on the client
5. If at the moment, when an attempt is made to load the class mentioned in item 4, the client is not available anymore, a class-loading exception happens on the server node, which is manifested as a NoClassDefFoundError, which is caught and processed by its Failure Handler. If the handler is 'stop-or-halt', the node is stopped.

So the scenario might cause a node failure, even though the original problem is local and transient. We should distinguish between p2p class load errors (which are non-critical) and non-p2p class load errors (which are critical).

Issue Links

• links to: GitHub Pull Request #9864