[ https://issues.apache.org/jira/browse/PHOENIX-4010?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16086911#comment-16086911 ]
Ankit Singhal commented on PHOENIX-4010: ---------------------------------------- Thanks [~sergey.soldatov] and [~giacomotaylor] for the review. bq. instead of a static Cache in HashCacheClient, how about storing it elsewhere, perhaps on TableResultIterator so you don't even need a Cache? Actually, we will need a cache at two different places(BaseResultIterators#getIterators and TableResultIterator#next()) depending upon the type of the query and way we are handling the iterators. bq. As far as the memory usage on the client, we track this with this code in ServerCacheClient. The idea with the MemoryManager off of QueryServices is to track memory usage at a course grain, potentially throttling or refusing allocation if too much is used. Yeah, thought of using the Memory Manager but this will fail the query if the memory limit is crossed which I think we don't want as we are keeping the cache on the client just to avoid failures for the join queries. so it is safe to evict the older entry rather than failing the current query. WDYT? > Hash Join cache may not be send to all regionservers when we have stale HBase > meta cache > ---------------------------------------------------------------------------------------- > > Key: PHOENIX-4010 > URL: https://issues.apache.org/jira/browse/PHOENIX-4010 > Project: Phoenix > Issue Type: Bug > Reporter: Ankit Singhal > Assignee: Ankit Singhal > Fix For: 4.12.0 > > Attachments: PHOENIX-4010.patch, PHOENIX-4010_v1.patch > > > If the region locations changed and our HBase meta cache is not updated then > we might not be sending hash join cache to all region servers hosting the > regions. > ConnectionQueryServicesImpl#getAllTableRegions > {code} > boolean reload =false; > while (true) { > try { > // We could surface the package projected > HConnectionImplementation.getNumberOfCachedRegionLocations > // to get the sizing info we need, but this would require a > new class in the same package and a cast > // to this implementation class, so it's probably not worth > it. > List<HRegionLocation> locations = Lists.newArrayList(); > byte[] currentKey = HConstants.EMPTY_START_ROW; > do { > HRegionLocation regionLocation = > connection.getRegionLocation( > TableName.valueOf(tableName), currentKey, reload); > locations.add(regionLocation); > currentKey = regionLocation.getRegionInfo().getEndKey(); > } while (!Bytes.equals(currentKey, HConstants.EMPTY_END_ROW)); > return locations; > {code} > Skipping duplicate servers in ServerCacheClient#addServerCache > {code} > List<HRegionLocation> locations = > services.getAllTableRegions(cacheUsingTable.getPhysicalName().getBytes()); > int nRegions = locations.size(); > > ..... > if ( ! servers.contains(entry) && > keyRanges.intersectRegion(regionStartKey, > regionEndKey, > cacheUsingTable.getIndexType() == > IndexType.LOCAL)) { > // Call RPC once per server > servers.add(entry); > {code} > For eg:- Table âTâ has two regions R1 and R2 originally hosted on > regionserver RS1. > while Phoenix/Hbase connection is still active, R2 is transitioned to RS2 , > but stale meta cache will still give old region locations i.e R1 and R2 on > RS1 and when we start copying hash table, we copy for R1 and skip R2 as they > are hosted on same regionserver. so, the query on a table will fail as it > will unable to find hash table cache on RS2 for processing regions R2. -- This message was sent by Atlassian JIRA (v6.4.14#64029)