[ https://issues.apache.org/jira/browse/HIVE-23880?focusedWorklogId=470577&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-470577 ]
ASF GitHub Bot logged work on HIVE-23880: ----------------------------------------- Author: ASF GitHub Bot Created on: 14/Aug/20 05:29 Start Date: 14/Aug/20 05:29 Worklog Time Spent: 10m Work Description: abstractdog commented on pull request #1280: URL: https://github.com/apache/hive/pull/1280#issuecomment-673892379 > @abstractdog > I am almost ok with this patch. However I still dont understand how this integrates with `ProcessingModeHashAggregate`. Since there are multiple VectorAggregationBufferRows in hash mode, I think we should `finish` each of them as we process them. Otherwise, we pass to the next operator in the pipeline without completing the bloom filter. Also, since hash mode dynamically allocates and frees VectorAggregationBufferRows these `finish`es should happen as we deallocate each of them, rather than only at the end of the operator. Good point. I was creating this patch by focusing on finishing buffers correctly, I think I've already taken care of by this, please take a look: https://github.com/apache/hive/pull/1280/commits/0ada66534a937b8f4492d14f508903fa98402aed#diff-07c28d3f5c72db581b9cd4fa424a0ecbR675 As you can see, I'm calling finish before every instance of writeSingleRow. I'm assuming that writeSingleRow is a point where a buffer should be finished for writing. In ProcessingModeHashAggregate, the above part is enclosed in an iteration on buffers in flush method. Are you aware of any other places where I should finish a buffer? ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org Issue Time Tracking ------------------- Worklog Id: (was: 470577) Time Spent: 7h 40m (was: 7.5h) > Bloom filters can be merged in a parallel way in VectorUDAFBloomFilterMerge > --------------------------------------------------------------------------- > > Key: HIVE-23880 > URL: https://issues.apache.org/jira/browse/HIVE-23880 > Project: Hive > Issue Type: Improvement > Reporter: László Bodor > Assignee: László Bodor > Priority: Major > Labels: pull-request-available > Attachments: lipwig-output3605036885489193068.svg > > Time Spent: 7h 40m > Remaining Estimate: 0h > > Merging bloom filters in semijoin reduction can become the main bottleneck in > case of large number of source mapper tasks (~1000, Map 1 in below example) > and a large amount of expected entries (50M) in bloom filters. > For example in TPCDS Q93: > {code} > select /*+ semi(store_returns, sr_item_sk, store_sales, 70000000)*/ > ss_customer_sk > ,sum(act_sales) sumsales > from (select ss_item_sk > ,ss_ticket_number > ,ss_customer_sk > ,case when sr_return_quantity is not null then > (ss_quantity-sr_return_quantity)*ss_sales_price > else > (ss_quantity*ss_sales_price) end act_sales > from store_sales left outer join store_returns on (sr_item_sk = > ss_item_sk > and > sr_ticket_number = ss_ticket_number) > ,reason > where sr_reason_sk = r_reason_sk > and r_reason_desc = 'reason 66') t > group by ss_customer_sk > order by sumsales, ss_customer_sk > limit 100; > {code} > On 10TB-30TB scale there is a chance that from 3-4 mins of query runtime 1-2 > mins are spent with merging bloom filters (Reducer 2), as in: > [^lipwig-output3605036885489193068.svg] > {code} > ---------------------------------------------------------------------------------------------- > VERTICES MODE STATUS TOTAL COMPLETED RUNNING PENDING > FAILED KILLED > ---------------------------------------------------------------------------------------------- > Map 3 .......... llap SUCCEEDED 1 1 0 0 > 0 0 > Map 1 .......... llap SUCCEEDED 1263 1263 0 0 > 0 0 > Reducer 2 llap RUNNING 1 0 1 0 > 0 0 > Map 4 llap RUNNING 6154 0 207 5947 > 0 0 > Reducer 5 llap INITED 43 0 0 43 > 0 0 > Reducer 6 llap INITED 1 0 0 1 > 0 0 > ---------------------------------------------------------------------------------------------- > VERTICES: 02/06 [====>>----------------------] 16% ELAPSED TIME: 149.98 s > ---------------------------------------------------------------------------------------------- > {code} > For example, 70M entries in bloom filter leads to a 436 465 696 bits, so > merging 1263 bloom filters means running ~ 1263 * 436 465 696 bitwise OR > operation, which is very hot codepath, but can be parallelized. -- This message was sent by Atlassian Jira (v8.3.4#803005)