Uh oh!
There was an error while loading. Please reload this page.
[fix](merge-on-write) segcompaction should process delete bitmap if necessary - #38369
Conversation
doris-robot
commented
Jul 25, 2024
Thank you for your contribution to Apache Doris. Since 2024-03-18, the Document has been moved to doris-website. |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
zhannngchen
commented
Jul 25, 2024
run buildall |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
zhannngchen
commented
Jul 25, 2024
run buildall |
6865700 to
7759455Comparezhannngchen
commented
Jul 25, 2024
run buildall |
zhannngchen
commented
Jul 25, 2024
run buildall |
doris-robot
commented
Jul 25, 2024
TPC-H: Total hot run time: 39487 ms |
doris-robot
commented
Jul 25, 2024
TPC-DS: Total hot run time: 171977 ms |
doris-robot
commented
Jul 25, 2024
ClickBench: Total hot run time: 30.94 s |
zhannngchen
commented
Jul 25, 2024
run buildall |
doris-robot
commented
Jul 25, 2024
TPC-H: Total hot run time: 39742 ms |
doris-robot
commented
Jul 25, 2024
TPC-DS: Total hot run time: 174282 ms |
doris-robot
commented
Jul 25, 2024
ClickBench: Total hot run time: 30.2 s |
PR approved by at least one committer and no changes requested. |
PR approved by anyone and no changes requested. |
…ecessary (apache#38369) ## Proposed changes Issue Number: close #xxx When loading data to a unique key table with sequence column, some data in current load job might be marked as delete due to a lower sequence value. If there's many segments in such load job, segcompaction might be triggered, which don't process the delete bitmap currently, will cause data correctness issue For example: 1. we have 4 segments in current load job initially, and due to seq column, some rows are marked as deleted 2. after segcompaction, if we don't process the delete bitmap, it's content is still corresponding to the old segment layout, and row 7,14,15 is not mark deleted correctly on new generated segment 1. 3. in this PR, we convert old delete bitmap to fit new segment layout, it use similar way as base/cumulative compaction to convert delete bitmaps on old layout to new one, but the rowid conversion is simpler 
…ecessary (#38369) ## Proposed changes Issue Number: close #xxx When loading data to a unique key table with sequence column, some data in current load job might be marked as delete due to a lower sequence value. If there's many segments in such load job, segcompaction might be triggered, which don't process the delete bitmap currently, will cause data correctness issue For example: 1. we have 4 segments in current load job initially, and due to seq column, some rows are marked as deleted 2. after segcompaction, if we don't process the delete bitmap, it's content is still corresponding to the old segment layout, and row 7,14,15 is not mark deleted correctly on new generated segment 1. 3. in this PR, we convert old delete bitmap to fit new segment layout, it use similar way as base/cumulative compaction to convert delete bitmaps on old layout to new one, but the rowid conversion is simpler 
## Proposed changes Issue Number: close #xxx introduced by #38369
…ecessary (#38369) ## Proposed changes Issue Number: close #xxx When loading data to a unique key table with sequence column, some data in current load job might be marked as delete due to a lower sequence value. If there's many segments in such load job, segcompaction might be triggered, which don't process the delete bitmap currently, will cause data correctness issue For example: 1. we have 4 segments in current load job initially, and due to seq column, some rows are marked as deleted 2. after segcompaction, if we don't process the delete bitmap, it's content is still corresponding to the old segment layout, and row 7,14,15 is not mark deleted correctly on new generated segment 1. 3. in this PR, we convert old delete bitmap to fit new segment layout, it use similar way as base/cumulative compaction to convert delete bitmaps on old layout to new one, but the rowid conversion is simpler 
## Proposed changes Issue Number: close #xxx introduced by #38369
…ecessary (apache#38369) Issue Number: close #xxx When loading data to a unique key table with sequence column, some data in current load job might be marked as delete due to a lower sequence value. If there's many segments in such load job, segcompaction might be triggered, which don't process the delete bitmap currently, will cause data correctness issue For example: 1. we have 4 segments in current load job initially, and due to seq column, some rows are marked as deleted 2. after segcompaction, if we don't process the delete bitmap, it's content is still corresponding to the old segment layout, and row 7,14,15 is not mark deleted correctly on new generated segment 1. 3. in this PR, we convert old delete bitmap to fit new segment layout, it use similar way as base/cumulative compaction to convert delete bitmaps on old layout to new one, but the rowid conversion is simpler 
## Proposed changes Issue Number: close #xxx introduced by apache#38369
…ecessary (apache#38369) (apache#39707) Issue Number: close #xxx cherry-pick apache#38369 and apache#38800
…ecessary (apache#38369) (apache#39749) cherry-pick apache#38369 and apache#38800
… input segments before converting delete bitmaps on them (#53198) ### What problem does this PR solve? When the table has sequence column, load may generate delete bitmap marks on its own rowset. Segment compaction should wait for delete bitmaps generation before converting these delete bitmaps(#38369).
… input segments before converting delete bitmaps on them (apache#53198) ### What problem does this PR solve? When the table has sequence column, load may generate delete bitmap marks on its own rowset. Segment compaction should wait for delete bitmaps generation before converting these delete bitmaps(apache#38369).
… input segments before converting delete bitmaps on them (apache#53198) ### What problem does this PR solve? When the table has sequence column, load may generate delete bitmap marks on its own rowset. Segment compaction should wait for delete bitmaps generation before converting these delete bitmaps(apache#38369).
… input segments before converting delete bitmaps on them (apache#53198) ### What problem does this PR solve? When the table has sequence column, load may generate delete bitmap marks on its own rowset. Segment compaction should wait for delete bitmaps generation before converting these delete bitmaps(apache#38369).
… input segments before converting delete bitmaps on them (apache#53198) ### What problem does this PR solve? When the table has sequence column, load may generate delete bitmap marks on its own rowset. Segment compaction should wait for delete bitmaps generation before converting these delete bitmaps(apache#38369).
… input segments before converting delete bitmaps on them (apache#53198) ### What problem does this PR solve? When the table has sequence column, load may generate delete bitmap marks on its own rowset. Segment compaction should wait for delete bitmaps generation before converting these delete bitmaps(apache#38369).
… input segments before converting delete bitmaps on them (apache#53198) ### What problem does this PR solve? When the table has sequence column, load may generate delete bitmap marks on its own rowset. Segment compaction should wait for delete bitmaps generation before converting these delete bitmaps(apache#38369).
…ecessary (apache#38369) ## Proposed changes Issue Number: close #xxx When loading data to a unique key table with sequence column, some data in current load job might be marked as delete due to a lower sequence value. If there's many segments in such load job, segcompaction might be triggered, which don't process the delete bitmap currently, will cause data correctness issue For example: 1. we have 4 segments in current load job initially, and due to seq column, some rows are marked as deleted 2. after segcompaction, if we don't process the delete bitmap, it's content is still corresponding to the old segment layout, and row 7,14,15 is not mark deleted correctly on new generated segment 1. 3. in this PR, we convert old delete bitmap to fit new segment layout, it use similar way as base/cumulative compaction to convert delete bitmaps on old layout to new one, but the rowid conversion is simpler 
## Proposed changes Issue Number: close #xxx introduced by apache#38369
Proposed changes
Issue Number: close #xxx
When loading data to a unique key table with sequence column, some data in current load job might be marked as delete due to a lower sequence value.
If there's many segments in such load job, segcompaction might be triggered, which don't process the delete bitmap currently, will cause data correctness issue
For example: