[ https://issues.apache.org/jira/browse/PHOENIX-1734?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15019065#comment-15019065 ]
Enis Soztutar commented on PHOENIX-1734: ---------------------------------------- We were discussing this with [~aliciashu], and she raised a good concern. I think with this patch, we are creating 1 shadow CF per table CF, but that CF contains data from multiple local indexes. We are using the column name for filtering out the other index's data. However, this means that the data from multiple local indices are intermixed at the physical layout making the scans for a single local index to scan over all the data. Thus a full table scan for a local index, still scans over the data for all local indices. Point index lookups will not be affected as much. I believe the query planner will never create a scan for scanning more than one index's data. Should we change the patch to create one CF per table CF per local index? It will be more similar to the one-table per local index setup, and scans for a particular covered index will only scan over the index data. > Local index improvements > ------------------------ > > Key: PHOENIX-1734 > URL: https://issues.apache.org/jira/browse/PHOENIX-1734 > Project: Phoenix > Issue Type: Improvement > Reporter: Rajeshbabu Chintaguntla > Assignee: Rajeshbabu Chintaguntla > Attachments: PHOENI-1734-WIP.patch, PHOENIX-1734_v1.patch, > TestAtomicLocalIndex.java > > > Local index design considerations: > 1. Colocation: We need to co-locate regions of local index regions and data > regions. The co-location can be a hard guarantee or a soft (best approach) > guarantee. The co-location is a performance requirement, and also maybe > needed for consistency(2). Hard co-location means that either both the data > region and index region are opened atomically, or neither of them open for > serving. > 2. Index consistency : Ideally we want the index region and data region to > have atomic updates. This means that they should either (a)use transactions, > or they should (b)share the same WALEdit and also MVCC for visibility. (b) is > only applicable if there is hard colocation guarantee. > 3. Local index clients : How the local index will be accessed from clients. > In case of the local index being managed in a table, the HBase client can be > used for doing scans, etc. If the local index is hidden inside the data > regions, there has to be a different mechanism to access the data through the > data region. > With the above considerations, we imagine three possible implementation for > the local index solution, each detailed below. > APPROACH 1: Current approach > (1) Current approach uses balancer as a soft guarantee. Because of this, in > some rare cases, colocation might not happen. > (2) The index and data regions do not share the same WALEdits. Meaning > consistency cannot be achieved. Also there are two WAL writes per write from > client. > (3) Regular Hbase client can be used to access index data since index is just > another table. > APPROACH 2: Shadow regions + shared WAL & MVCC > (1) Introduce a shadow regions concept in HBase. Shadow regions are not > assigned by AM. Phoenix implements atomic open (and split/merge) of region > opening for data regions and index regions so that hard co-location is > guaranteed. > (2) For consistency requirements, the index regions and data regions will > share the same WALEdit (and thus recovery) and they will also share the same > MVCC mechanics so that index update and data update is visible atomically. > (3) Regular Hbase client can be used to access index data since index is just > another table. > APPROACH 3: Storing index data in separate column families in the table. > (1) Regions will have store files for cfs, which is sorted using the primary > sort order. Regions may also maintain stores, sorted in secondary sort > orders. This approach is similar in vein how a RDBMS keeps data (a B-TREE in > primary sort order and multiple B-TREEs in secondary sort orders with > pointers to primary key). That means store the index data in separate column > families in the data region. This way a region is extended to be more similar > to a RDBMS (but LSM instead of BTree). This is sometimes called shadow cf’s > as well. This approach guarantees hard co-location. > (2) Since everything is in a single region, they automatically share the > same WALEdit and MVCC numbers. Atomicity is easily achieved. > (3) Current Phoenix implementation need to change in such a way that column > families selection in read/write path is based data table/index table(logical > table in phoenix). > I think that APPROACH 3 is the best one for long term, since it does not > require to change anything in HBase, mainly we don't need to muck around with > the split/merge stuff in HBase. It will be win-win. > However, APPROACH 2 still needs a “shadow regions” concept to be implemented > in HBase itself, and also a way to share WALEdits and MVCCs from multiple > regions. > APPROACH 1 is a good start for local indexes, but I think we are not getting > the full benefits for the feature. We can support this for the short term, > and decide on the next steps for a longer term implementation. > we won't be able to get to implementing it immediately, and want to start a > brainstorm. -- This message was sent by Atlassian JIRA (v6.3.4#6332)