[ https://issues.apache.org/jira/browse/HBASE-4081?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Ming Ma updated HBASE-4081:
---------------------------
Attachment: HBASE-4081-trunk.patch
Thanks, Stack, Ted.
1. It looks like there was some code rewrite for compaction and split if we compare trunk and 0.90.2.
2. compactStores method used to the main method for compaction in 0.90.2. It returns split key to so that split can happen if necessary.
3. In trunk, it has been rewritten to use thread pool to do compaction and split. There is no need to return split key in normal compaction process. compactStores are there still to support synchronous compaction need by utilities and testing.
Given these, I basically go with the original proposals and fix it so that test code can get split key with a separate method. Separating compaction and split on method level seems to be cleaner.
> Issues with HRegion.compactStores methods
> -----------------------------------------
>
> Key: HBASE-4081
> URL: https://issues.apache.org/jira/browse/HBASE-4081
> Project: HBase
> Issue Type: Bug
> Components: regionserver
> Reporter: Ming Ma
> Assignee: Ming Ma
> Attachments: HBASE-4081-trunk.patch
>
>
> HRegion.java,
> byte [] compactStores(final boolean majorCompaction)
> throws IOException {
> if (majorCompaction) {
> this.triggerMajorCompaction();
> }
> return compactStores();
> }
> /**
> * Compact all the stores and return the split key of the first store that needs
> * to be split.
> */
> public byte[] compactStores() throws IOException {
> for(Store s : getStores().values()) {
> CompactionRequest cr = s.requestCompaction();
> if(cr != null) {
> try {
> compact(cr);
> } finally {
> s.finishRequest(cr);
> }
> }
> byte[] splitRow = s.checkSplit();
> if (splitRow != null) {
> return splitRow;
> }
> }
> return null;
> }
> 1. It seems the second method's intention is to compact all the stores. However, if a store requires split, the process will stop.
> 2. Only MetaUtils, HRegion.merge, HRegion.processTable use these two methods. No caller uses the return value.
> 3. HRegion.merge expects major compaction for each store after the call and has code like below to check error condition.
> // Because we compacted the source regions we should have no more than two
> // HStoreFiles per family and there will be no reference store
> if (srcFiles.size() == 2)
> So it seems like the fixes are: a) take out s.CheckSplit() call inside compactStores. b) make the return type "void" for these two compactStores functions.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira