chenglei created PHOENIX-2970: --------------------------------- Summary: SpoolingResultIterator using memory too conservative Key: PHOENIX-2970 URL: https://issues.apache.org/jira/browse/PHOENIX-2970 Project: Phoenix Issue Type: Bug Affects Versions: 4.7.0 Reporter: chenglei
Even if SpoolingResultIterator will be deprecated, but HBase older than than 0.98.17 would continue use it, and I think the DeferredByteBufferSegmentQueue class simialr to SpoolingResultIterator in the future version may also has the same problem. In SpoolingResultIterator's ctor, it tries to allocate "phoenix.query. spoolThresholdBytes" bytes from MemoryManager,and uses the allocated MemoryChunk's size as DeferredFileOutputStream's threshold,just as the following code: {code:borderStyle=solid} final MemoryChunk chunk = mm.allocate(0, thresholdBytes); long waitTime = System.currentTimeMillis() - startTime; GLOBAL_MEMORY_WAIT_TIME.update(waitTime); memoryMetrics.getMemoryWaitTimeMetric().change(waitTime); DeferredFileOutputStream spoolTo = null; try { // Can't be bigger than int, since it's the max of the above allocation int size = (int)chunk.getSize(); spoolTo = new DeferredFileOutputStream(size, "ResultSpooler",".bin", new File(spoolDirectory)) { @Override protected void thresholdReached() throws IOException { try { super.thresholdReached(); } finally { chunk.close(); } } }; {code} SpoolingResultIterator assumes that the wrapped ResultIterator would always fetch "phoenix.query.spoolThresholdBytes" bytes,but most of time it won't. For example, if we execute "select count(*) from table" on a big table with many regions, the ScanPlan will parallel too many SpoolingResultIterators to fetch the result, and each SpoolingResultIterator tries to allocate "phoenix.query. spoolThresholdBytes" bytes from MemoryManager.If we don't have too many memory, lots of SpoolingResultIterators will allocate 0 bytes from MemoryManager,and the DeferredFileOutputStream will unnecessaryly use temp file to put the results, even if the result is just a single count value. This behavior will slow the query. Can we remove the MemoryManager,just as the MappedByteBufferSortedQueue does? -- This message was sent by Atlassian JIRA (v6.3.4#6332)