Uh oh!
There was an error while loading. Please reload this page.
HBASE-28584 RS SIGSEGV under heavy replication load - #6124
Conversation
Deep clone the cells that are used to apply mutations on the local cluster. Some operations may still be in flight even as we fail to apply some other in-flight mutations and trigger failure handling including a release of the buffer underlying the cellScanner that is sourcing the cells.
Apache-HBase
commented
Jul 28, 2024
🎊 +1 overall
This message was automatically generated. |
Apache9
left a comment
There was a problem hiding this comment.
Let's get this in to fix the crash issue first.
IIRC, we have a similar problem when async wal is enabled. The rpc call will finish before we actually finish the WAL writing and cause we write out corrupt wal entries.
Let me check how we deal with the problem there. I guess the same trick can be used here too.
Thanks @apurtell for the analyzing.
Apache9
commented
Jul 28, 2024
OK, in ServerCall class, we have a retainByWAL method, where we will count the extra references of the ServerCall, mainly the CellScanners. I think we can just change the method to retain, which means we want to retain it for other usage even after the rpc call is done, and also use it here. Anyway, since the current PR has been well tested in producation, I think we can apply it first, and open another issue for optimizing. Thanks. |
Apache-HBase
commented
Jul 28, 2024
🎊 +1 overall
This message was automatically generated. |
| * Deep clones the given cell if the cell supports deep cloning | ||
| * @param cell the cell to be cloned | ||
| * @return the cloned cell |
There was a problem hiding this comment.
nit: javadoc to include @throws
Apache9
commented
Sep 4, 2024
Any updates here? Thanks. |
Apache9
commented
Sep 17, 2024
I think we should fix this in newer releases. If you are all OK, I could try to implement the reference counting way to solve the problem. @apurtell@virajjasani Thoughts? Thanks. |
apurtell
commented
Sep 17, 2024
I thought refcounting would be complex but am not opposed to it as a different solution. When and if we have that we could remove the copying. |
apurtell
commented
Sep 18, 2024
We would not need this change if #6263 solves the problem instead. |
apurtell
commented
Sep 20, 2024
Fixed by #6263 |
Clone the cells that are used to apply mutations on the local cluster. Some operations may still be in flight even as we fail to apply some other in-flight mutations and trigger failure handling including a release of the buffer underlying the cellScanner that is sourcing the cells.