[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14136886#comment-14136886 ] Hive QA commented on HIVE-8038: --- {color:green}Overall{color}: +1 all checks pass Here are the results of testing the latest attachment: https://issues.apache.org/jira/secure/attachment/12669113/HIVE-8038.3.patch {color:green}SUCCESS:{color} +1 6279 tests passed Test results: http://ec2-174-129-184-35.compute-1.amazonaws.com/jenkins/job/PreCommit-HIVE-TRUNK-Build/835/testReport Console output: http://ec2-174-129-184-35.compute-1.amazonaws.com/jenkins/job/PreCommit-HIVE-TRUNK-Build/835/console Test logs: http://ec2-174-129-184-35.compute-1.amazonaws.com/logs/PreCommit-HIVE-TRUNK-Build-835/ Messages: {noformat} Executing org.apache.hive.ptest.execution.PrepPhase Executing org.apache.hive.ptest.execution.ExecutionPhase Executing org.apache.hive.ptest.execution.ReportingPhase {noformat} This message is automatically generated. ATTACHMENT ID: 12669113 Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Assignee: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.2.patch, HIVE-8038.3.patch, HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14135101#comment-14135101 ] Hive QA commented on HIVE-8038: --- {color:green}Overall{color}: +1 all checks pass Here are the results of testing the latest attachment: https://issues.apache.org/jira/secure/attachment/12668843/HIVE-8038.2.patch {color:green}SUCCESS:{color} +1 6276 tests passed Test results: http://ec2-174-129-184-35.compute-1.amazonaws.com/jenkins/job/PreCommit-HIVE-TRUNK-Build/815/testReport Console output: http://ec2-174-129-184-35.compute-1.amazonaws.com/jenkins/job/PreCommit-HIVE-TRUNK-Build/815/console Test logs: http://ec2-174-129-184-35.compute-1.amazonaws.com/logs/PreCommit-HIVE-TRUNK-Build-815/ Messages: {noformat} Executing org.apache.hive.ptest.execution.PrepPhase Executing org.apache.hive.ptest.execution.ExecutionPhase Executing org.apache.hive.ptest.execution.ReportingPhase {noformat} This message is automatically generated. ATTACHMENT ID: 12668843 Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Assignee: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.2.patch, HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14135848#comment-14135848 ] Pankit Thapar commented on HIVE-8038: - Is .3.patch commited to trunk? Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Assignee: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.2.patch, HIVE-8038.3.patch, HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14135857#comment-14135857 ] Gopal V commented on HIVE-8038: --- No, there is a 24 hour waiting period after the +1. I will resolve the ticket once it is committed. Leave comments if you need to. Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Assignee: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.2.patch, HIVE-8038.3.patch, HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14134389#comment-14134389 ] Gopal V commented on HIVE-8038: --- [~pankit]: Updated patch looks good, left one comment about the exception type. Please attach latest patch here and I'll try to get this in (name it, with .2.patch). Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14134413#comment-14134413 ] Pankit Thapar commented on HIVE-8038: - [~gopalv] , Thanks for taking a look. I have changed the exception to IOException and uploaded the new patch here. Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.2.patch, HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14131648#comment-14131648 ] Pankit Thapar commented on HIVE-8038: - Hi, Can you please take a look at this cr : https://reviews.apache.org/r/25521/ Thanks, Pankit Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14132107#comment-14132107 ] Gopal V commented on HIVE-8038: --- Thanks [~pankit], left some minor comments on the RB. I cannot find any performance improvement due to this patch in my HDFS builds, but admittedly I do not have variable length blocks to trigger the need for this. Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14132263#comment-14132263 ] Pankit Thapar commented on HIVE-8038: - Hi Gopal, Thanks for taking a look. I have uploaded an updated diff on the cr with the changes recommended. And also, commented on the feedback. Please let me know your feedback on the same. Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14128768#comment-14128768 ] Gopal V commented on HIVE-8038: --- This is an interesting change-set. bq. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length For ORC to be really fast, we enforce that a stripe (the smallest split you can get) always fits within a block - this is true for HDFS at least, because it can specify a preferred block size when creating files. From an elegance point of view, I like the TreeMap.floorEntry() over a for loop - but I have never seen the 4A/4B scenarios when using Hive-13. bq. 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems This is something to be fixed anyway - as HDFS-3689 will allow variable length blocks in HDFS as well. Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14128871#comment-14128871 ] Pankit Thapar commented on HIVE-8038: - Hi, Thanks for the feedback. 1. The use case where the split may span more than one block would be when Math.min(MAX_BLOCK_SIZE, 2 * stripeSize) returns MAX_BLOCK_SIZE as the size of the block for the file. Example : stripe size 512MB and BLOCK SIZE is 400MB, in that case, split would span more than one block. 2. I see that HDFS wants to support variable length blocks but what I meant was to remove the usage of blockSize variable all together as that is not true for all the FileSystems. We want to generalize the usage for FileSystems apart from HDFS. Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14129029#comment-14129029 ] Hive QA commented on HIVE-8038: --- {color:red}Overall{color}: -1 at least one tests failed Here are the results of testing the latest attachment: https://issues.apache.org/jira/secure/attachment/12667782/HIVE-8038.patch {color:red}ERROR:{color} -1 due to 1 failed/errored test(s), 6195 tests executed *Failed tests:* {noformat} org.apache.hive.hcatalog.pig.TestOrcHCatLoader.testReadDataPrimitiveTypes {noformat} Test results: http://ec2-174-129-184-35.compute-1.amazonaws.com/jenkins/job/PreCommit-HIVE-TRUNK-Build/728/testReport Console output: http://ec2-174-129-184-35.compute-1.amazonaws.com/jenkins/job/PreCommit-HIVE-TRUNK-Build/728/console Test logs: http://ec2-174-129-184-35.compute-1.amazonaws.com/logs/PreCommit-HIVE-TRUNK-Build-728/ Messages: {noformat} Executing org.apache.hive.ptest.execution.PrepPhase Executing org.apache.hive.ptest.execution.ExecutionPhase Executing org.apache.hive.ptest.execution.ReportingPhase Tests exited with: TestsFailedException: 1 tests failed {noformat} This message is automatically generated. ATTACHMENT ID: 12667782 Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14129205#comment-14129205 ] Gopal V commented on HIVE-8038: --- bq. Example : stripe size 512MB and BLOCK SIZE is 400MB, in that case, split would span more than one block. I think that would be a case for reducing the stripe size, if performance is the primary goal. We already default to a 64Mb stripe in hive-14 down from 256Mb because of similar issues. That said, a TreeMap is probably the right way to do this when there is variable length support (also we can turn off orc.padding). bq. We want to generalize the usage for FileSystems apart from HDFS. My current ~30Tb tests are all off HDFS. Can you tell me what filesystem impl that you are targetting these fixes on? Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14129257#comment-14129257 ] Pankit Thapar commented on HIVE-8038: - We have a custom Filesystem implementation over S3. Our block allocation logic is a little different from hdfs. So, I will go ahead and see the failed test and try to fix it. Do you have comments on the code change ? Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14129266#comment-14129266 ] Pankit Thapar commented on HIVE-8038: - org.apache.hive.hcatalog.pig.TestOrcHCatLoader.testReadDataPrimitiveTypes fails even without the patch I submitted. Can someone , please confirm that? Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (HIVE-8038) Decouple ORC files split calculation logic from Filesystem's get file location implementation
[ https://issues.apache.org/jira/browse/HIVE-8038?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14129273#comment-14129273 ] Gopal V commented on HIVE-8038: --- Please open a reviews.apache.org reviewboard entry - also add [~prasanth_j] me as reviews on the entry. I will add this patch to my nightly perf runs and then take a look at it. Decouple ORC files split calculation logic from Filesystem's get file location implementation - Key: HIVE-8038 URL: https://issues.apache.org/jira/browse/HIVE-8038 Project: Hive Issue Type: Improvement Components: File Formats Affects Versions: 0.13.1 Reporter: Pankit Thapar Fix For: 0.14.0 Attachments: HIVE-8038.patch What is the Current Logic == 1.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using the array index (index = offset/blockSize), get the corresponding host having the blockLocation 4.If the split spans multiple blocks, then get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits Issue with Current Logic = Dependency on FileSystem API’s logic for block location calculations. It returns an array and we need to rely on FileSystem to make all blocks of same size if we want to directly access a block from the array. What is the Fix = 1a.get the file blocks from FileSystem.getFileBlockLocations() which returns an array of BlockLocation 1b.convert the array into a tree map offset, BlockLocation and return it through getLocationsWithOffSet() 2.In SplitGenerator.createSplit(), check if split only spans one block or multiple blocks. 3.If split spans just one block, then using Tree.floorEntry(key), get the highest entry smaller than offset for the split and get the corresponding host. 4a.If the split spans multiple blocks, get a submap, which contains all entries containing blockLocations from the offset to offset + length 4b.get all hosts that have at least 80% of the max of total data in split hosted by any host. 5.add the split to a list of splits What are the major changes in logic == 1. store BlockLocations in a Map instead of an array 2. Call SHIMS.getLocationsWithOffSet() instead of getLocations() 3. one block case is checked by if(offset + length = start.getOffset() + start.getLength()) instead of if((offset % blockSize) + length = blockSize) What is the affect on Complexity (Big O) = 1. We add a O(n) loop to build a TreeMap from an array but its a one time cost and would not be called for each split 2. In case of one block case, we can get the block in O(logn) worst case which was O(1) before 3. Getting the submap is O(logn) 4. In case of multiple block case, building the list of hosts is O(m) which was O(n) m n as previously we were iterating over all the block locations but now we are only iterating only blocks that belong to that range go offsets that we need. What are the benefits of the change == 1. With this fix, we do not depend on the blockLocations returned by FileSystem to figure out the block corresponding to the offset and blockSize 2. Also, it is not necessary that block lengths is same for all blocks for all FileSystems 3. Previously we were using blockSize for one block case and block.length for multiple block case, which is not the case now. We figure out the block depending upon the actual length and offset of the block -- This message was sent by Atlassian JIRA (v6.3.4#6332)