[jira] [Commented] (RATIS-1979) Correct LogEntryProto cache size calculation for stateMachineCachingEnabled=true

["Tsz-wo Sze (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/RATIS-1979?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17797868#comment-17797868 ] 

Tsz-wo Sze commented on RATIS-1979:
-----------------------------------

If both zero copy and stateMachineCaching are enabled, how about we copy the log proto before put it to cache?  Then, the zero copy buffers can be released earlier.

> Correct LogEntryProto cache size calculation for stateMachineCachingEnabled=true
> --------------------------------------------------------------------------------
>
>                 Key: RATIS-1979
>                 URL: https://issues.apache.org/jira/browse/RATIS-1979
>             Project: Ratis
>          Issue Type: Sub-task
>            Reporter: Duong
>            Priority: Major
>
> With zero-copy (https://github.com/apache/ratis/pull/990), we rely on RaftLog cache eviction to release the zero-copy input streams.
> There's a problems, with stateMachineCachingEnabled, the cache-size of LogEntryproto is calculated with StateMachineData trimmed. That leave the actual cached size if a log entry is too small comparing to the amount of direct memory backed for the StateMachineData. 
>  
> {code:java}
> static long getEntrySize(LogEntryProto entry, Op op) {
>   LogEntryProto e = entry;
>   if (op == Op.CHECK_SEGMENT_FILE_FULL) {
>     e = LogProtoUtils.removeStateMachineData(entry);
>   } else if (op == Op.LOAD_SEGMENT_FILE || op == Op.WRITE_CACHE_WITH_STATE_MACHINE_CACHE) {
>     Preconditions.assertTrue(entry == LogProtoUtils.removeStateMachineData(entry),
>         () -> "Unexpected LogEntryProto with StateMachine data: op=" + op + ", entry=" + entry);
>   } else {
>     Preconditions.assertTrue(op == Op.WRITE_CACHE_WITHOUT_STATE_MACHINE_CACHE || op == Op.REMOVE_CACHE,
>         () -> "Unexpected op " + op + ", entry=" + entry);
>   }
>   final int serialized = e.getSerializedSize();
>   return serialized + CodedOutputStream.computeUInt32SizeNoTag(serialized) + 4L;
> } {code}
>  
> With the default 200MB limit for raft log cache, cache eviction is likely never happens until the server run out of direct memory for zero-copy.
>  
> Maybe for cache calculation, we should take the StateMachineData size into account.



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