[
https://issues.apache.org/jira/browse/HDDS-1992?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Bharat Viswanadham updated HDDS-1992:
-------------------------------------
Parent: HDDS-505
Issue Type: Sub-task (was: Bug)
> Have a separate CacheTable class for cache based tables.
> --------------------------------------------------------
>
> Key: HDDS-1992
> URL: https://issues.apache.org/jira/browse/HDDS-1992
> Project: Hadoop Distributed Data Store
> Issue Type: Sub-task
> Components: om
> Reporter: Bharat Viswanadham
> Priority: Major
> Fix For: 0.4.1
>
>
> 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
(v8.3.2#803003)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
