Uh oh!
There was an error while loading. Please reload this page.
[fix](move-memtable) fix move memtable core when use multi table load - #35458
Conversation
doris-robot
commented
May 27, 2024
Thank you for your contribution to Apache Doris. Since 2024-03-18, the Document has been moved to doris-website. |
sollhui
commented
May 27, 2024
run buildall |
clang-tidy review says "All clean, LGTM! 👍" |
PR approved by at least one committer and no changes requested. |
PR approved by anyone and no changes requested. |
doris-robot
commented
May 27, 2024
TPC-H: Total hot run time: 42153 ms |
doris-robot
commented
May 27, 2024
TPC-DS: Total hot run time: 168886 ms |
doris-robot
commented
May 27, 2024
TeamCity be ut coverage result: |
doris-robot
commented
May 27, 2024
ClickBench: Total hot run time: 30.6 s |
clang-tidy review says "All clean, LGTM! 👍" |
sollhui
commented
May 30, 2024
run buildall |
clang-tidy review says "All clean, LGTM! 👍" |
sollhui
commented
May 30, 2024
run buildall |
clang-tidy review says "All clean, LGTM! 👍" |
sollhui
commented
May 30, 2024
run buildall |
clang-tidy review says "All clean, LGTM! 👍" |
doris-robot
commented
May 30, 2024
TPC-H: Total hot run time: 40034 ms |
doris-robot
commented
May 30, 2024
TPC-DS: Total hot run time: 169100 ms |
doris-robot
commented
May 30, 2024
ClickBench: Total hot run time: 30.58 s |
sollhui
commented
May 31, 2024
run buildall |
clang-tidy review says "All clean, LGTM! 👍" |
doris-robot
commented
May 31, 2024
TPC-H: Total hot run time: 42156 ms |
PR approved by at least one committer and no changes requested. |
Uh oh!
There was an error while loading. Please reload this page.
dataroaring
commented
Jun 4, 2024
run buildall |
clang-tidy review says "All clean, LGTM! 👍" |
doris-robot
commented
Jun 4, 2024
TeamCity be ut coverage result: |
doris-robot
commented
Jun 4, 2024
TPC-H: Total hot run time: 41673 ms |
doris-robot
commented
Jun 4, 2024
TPC-DS: Total hot run time: 173022 ms |
doris-robot
commented
Jun 4, 2024
ClickBench: Total hot run time: 30.77 s |
…#35458) ## Proposed changes Move memtable core when use multi table load: ``` 0x51f000c73860 is located 3040 bytes inside of 3456-byte region [0x51f000c72c80,0x51f000c73a00) freed by thread T4867 (FragmentMgrThre) here: #0 0x558f6ad7f43d in operator delete(void*) (/mnt/hdd01/STRESS_ENV/be/lib/doris_be+0x22eec43d) (BuildId: b46f73d1f76dfcd6) #1 0x558f6e6cea2c in std::__new_allocator<doris::PTabletID>::deallocate(doris::PTabletID*, unsigned long) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/new_allocator.h:168:2 #2 0x558f6e6ce9e7 in std::allocator<doris::PTabletID>::deallocate(doris::PTabletID*, unsigned long) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/allocator.h:210:25 #3 0x558f6e6ce9e7 in std::allocator_traits<std::allocator<doris::PTabletID>>::deallocate(std::allocator<doris::PTabletID>&, doris::PTabletID*, unsigned long) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/alloc_traits.h:516:13 #4 0x558f6e6ce9e7 in std::_Vector_base<doris::PTabletID, std::allocator<doris::PTabletID>>::_M_deallocate(doris::PTabletID*, unsigned long) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/stl_vector.h:387:4 #5 0x558f6e6d0780 in void std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>::_M_range_insert<__gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>>(__gnu_cxx::__normal_iterator<doris::PTabletID*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, __gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, __gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, std::forward_iterator_tag) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/vector.tcc:832:3 #6 0x558f6e6c54c5 in __gnu_cxx::__normal_iterator<doris::PTabletID*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>> std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>::insert<__gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, void>(__gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, __gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, __gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/stl_vector.h:1483:4 #7 0x558f9b4b214f in doris::LoadStreamMap::save_tablets_to_commit(long, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>> const&) /mnt/disk2/xujianxu/doris/be/src/vec/sink/load_stream_map_pool.cpp:90:13 #8 0x558f9b7258dd in doris::vectorized::VTabletWriterV2::_calc_tablets_to_commit() /mnt/disk2/xujianxu/doris/be/src/vec/sink/writer/vtablet_writer_v2.cpp:650:27 #9 0x558f9b7229f1 in doris::vectorized::VTabletWriterV2::close(doris::Status) /mnt/disk2/xujianxu/doris/be/src/vec/sink/writer/vtablet_writer_v2.cpp:547:9 ``` Multiple sinks with different table loads use the load id, causing confusion in the use of shared data structures between sinks.
We meet OOM when using single stream multi table  It exist memory leak, and heap profile like:  The stream load context will not release in some exception conditions as plan failed for high concurrency causing timeout when obtaining read lock. It is introduced by #35458 The solution effect is shown in the following figure, which can run stably with a small amount of memory 
We meet OOM when using single stream multi table  It exist memory leak, and heap profile like:  The stream load context will not release in some exception conditions as plan failed for high concurrency causing timeout when obtaining read lock. It is introduced by #35458 The solution effect is shown in the following figure, which can run stably with a small amount of memory 
…#38824) pick (#38255) We meet OOM when using single stream multi table  It exist memory leak, and heap profile like:  The stream load context will not release in some exception conditions as plan failed for high concurrency causing timeout when obtaining read lock. It is introduced by #35458 The solution effect is shown in the following figure, which can run stably with a small amount of memory 
…apache#35458) ## Proposed changes Move memtable core when use multi table load: ``` 0x51f000c73860 is located 3040 bytes inside of 3456-byte region [0x51f000c72c80,0x51f000c73a00) freed by thread T4867 (FragmentMgrThre) here: #0 0x558f6ad7f43d in operator delete(void*) (/mnt/hdd01/STRESS_ENV/be/lib/doris_be+0x22eec43d) (BuildId: b46f73d1f76dfcd6) #1 0x558f6e6cea2c in std::__new_allocator<doris::PTabletID>::deallocate(doris::PTabletID*, unsigned long) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/new_allocator.h:168:2 #2 0x558f6e6ce9e7 in std::allocator<doris::PTabletID>::deallocate(doris::PTabletID*, unsigned long) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/allocator.h:210:25 #3 0x558f6e6ce9e7 in std::allocator_traits<std::allocator<doris::PTabletID>>::deallocate(std::allocator<doris::PTabletID>&, doris::PTabletID*, unsigned long) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/alloc_traits.h:516:13 #4 0x558f6e6ce9e7 in std::_Vector_base<doris::PTabletID, std::allocator<doris::PTabletID>>::_M_deallocate(doris::PTabletID*, unsigned long) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/stl_vector.h:387:4 #5 0x558f6e6d0780 in void std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>::_M_range_insert<__gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>>(__gnu_cxx::__normal_iterator<doris::PTabletID*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, __gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, __gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, std::forward_iterator_tag) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/vector.tcc:832:3 #6 0x558f6e6c54c5 in __gnu_cxx::__normal_iterator<doris::PTabletID*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>> std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>::insert<__gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, void>(__gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, __gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>, __gnu_cxx::__normal_iterator<doris::PTabletID const*, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>>>) /mnt/disk2/xujianxu/ldb_toolchain/bin/../lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/bits/stl_vector.h:1483:4 #7 0x558f9b4b214f in doris::LoadStreamMap::save_tablets_to_commit(long, std::vector<doris::PTabletID, std::allocator<doris::PTabletID>> const&) /mnt/disk2/xujianxu/doris/be/src/vec/sink/load_stream_map_pool.cpp:90:13 #8 0x558f9b7258dd in doris::vectorized::VTabletWriterV2::_calc_tablets_to_commit() /mnt/disk2/xujianxu/doris/be/src/vec/sink/writer/vtablet_writer_v2.cpp:650:27 #9 0x558f9b7229f1 in doris::vectorized::VTabletWriterV2::close(doris::Status) /mnt/disk2/xujianxu/doris/be/src/vec/sink/writer/vtablet_writer_v2.cpp:547:9 ``` Multiple sinks with different table loads use the load id, causing confusion in the use of shared data structures between sinks.
…38255) We meet OOM when using single stream multi table  It exist memory leak, and heap profile like:  The stream load context will not release in some exception conditions as plan failed for high concurrency causing timeout when obtaining read lock. It is introduced by apache#35458 The solution effect is shown in the following figure, which can run stably with a small amount of memory 
Proposed changes
Move memtable core when use multi table load:
Multiple sinks with different table loads use the load id, causing confusion in the use of shared data structures between sinks.
Further comments
If this is a relatively large or complex change, kick off the discussion at dev@doris.apache.org by explaining why you chose the solution you did and what alternatives you considered, etc...