Elek, Marton created HDDS-1926:
----------------------------------

             Summary: The new caching layer is used for old OM requests but not 
updated
                 Key: HDDS-1926
                 URL: https://issues.apache.org/jira/browse/HDDS-1926
             Project: Hadoop Distributed Data Store
          Issue Type: Bug
          Components: om
            Reporter: Elek, Marton


HDDS-1499 introduced a new caching layer together with a double-buffer based db 
writer to support OM HA.

TLDR: I think the caching layer is not updated for new volume creation. And 
(slightly related to this problem) I suggest to separated the TypedTable and 
the caching layer.

## How to reproduce the problem?

1. Start a docker compose cluster
2. Create one volume (let's say `/vol1`)
3. Restart the om (!)
4. Try to create an _other_ volume twice!

```
bash-4.2$ ozone sh volume create /vol2
2019-08-07 12:29:47 INFO  RpcClient:288 - Creating Volume: vol2, with hadoop as 
owner.
bash-4.2$ ozone sh volume create /vol2
2019-08-07 12:29:50 INFO  RpcClient:288 - Creating Volume: vol2, with hadoop as 
owner.
```

Expected behavior is an error:

{code}
bash-4.2$ ozone sh volume create /vol1
2019-08-07 09:48:39 INFO  RpcClient:288 - Creating Volume: vol1, with hadoop as 
owner.
bash-4.2$ ozone sh volume create /vol1
2019-08-07 09:48:42 INFO  RpcClient:288 - Creating Volume: vol1, with hadoop as 
owner.
VOLUME_ALREADY_EXISTS 
{code}

The problem is that the new cache is used even for the old code path 
(TypedTable):

{code}
 @Override
  public VALUE get(KEY key) throws IOException {
    // Here the metadata lock will guarantee that cache is not updated for same
    // key during get key.

    CacheResult<CacheValue<VALUE>> cacheResult =
        cache.lookup(new CacheKey<>(key));

    if (cacheResult.getCacheStatus() == EXISTS) {
      return cacheResult.getValue().getCacheValue();
    } else if (cacheResult.getCacheStatus() == NOT_EXIST) {
      return null;
    } else {
      return getFromTable(key);
    }
  }
{code}

For volume table after the FIRST start it always returns with 
`getFromTable(key)` due to the condition in the `TableCacheImpl.lookup`:

{code}

  public CacheResult<CACHEVALUE> lookup(CACHEKEY cachekey) {

    if (cache.size() == 0) {
      return new CacheResult<>(CacheResult.CacheStatus.MAY_EXIST,
          null);
    }
{code}

But after a restart the cache is pre-loaded by the TypedTable.constructor. 
After the restart, the real caching logic will be used (as cache.size()>0), 
which cause a problem as the cache is NOT updated from the old code path.

An additional problem is that the cache is turned on for all the metadata table 
even if the cache is not required... 

## Proposed solution

As I commented at HDDS-1499 this caching layer is not a "traditional cache". 
It's not updated during the typedTable.put() call but updated by a separated 
component during double-buffer flash.

I would suggest to remove the cache related methods from TypedTable (move to a 
separated implementation). I think this kind of caching can be independent from 
the TypedTable implementation. We can continue to use the simple TypedTable 
everywhere where we don't need to use any kind of caching.

For caching we can use a separated object. It would make it more visible that 
the cache should always be updated manually all the time. This separated 
caching utility may include a reference to the original TypedTable/Table. With 
this approach we can separate the different responsibilities but provide the 
same functionality.



--
This message was sent by Atlassian JIRA
(v7.6.14#76016)

---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-dev-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-dev-h...@hadoop.apache.org

Reply via email to