adol001

> 预训练的loss你们最后超过1了吗，我在4左右降不下去了哇 降到4以下了，没降到1以下

你们用89万数据就降到4了？过拟合了吧

In the comments of RaftJournalSystem, it is mentioned that * This could happen if a server thinks it is the primary, and ignores a journal entry served from * the...

``` "Safety: the key safety property for Raft is the State Machine Safety Property in Figure 3: if any server has applied a particular log entry to its state machine,...

The core of the problem is that apply log to statemachine precedes the commit log. (preapply) This will cause: 1 The client to get inconsistent results after the master is...

Code from hdfs ``` boolean delete(String src, boolean recursive, boolean logRetryCache) throws IOException { final String operationName = "delete"; BlocksMapUpdateInfo toRemovedBlocks = null; checkOperation(OperationCategory.WRITE); final FSPermissionChecker pc = getPermissionChecker(); FSPermissionChecker.setOperationType(operationName);...

If all operations go through alluxio, the metadata should be based on alluxio. It should not load the missing metadata from ufs when alluxio restarts or re-selects the master. Synchronization...

@ggezer Inconsistency is inevitable for two separate and united systems, but the data in them need to be recoverable. Inconsistency but recoverable "Having a ghost metadata on Alluxio is acceptable",...