Segmentation Fault when using join_algorithm='partial_merge' #11791
Description
Describe the bug
Segmentation fault when using join_algorithm='partial_merge'
How to reproduce
Trying to merge large tables using merge join.
Problem does NOT happen when limits are applied on the "right" table.
Could not understand specific threshold on when it start to appear.
Right table is heavy and has long strings of 50-400 char wide.
Version: 20.3.11.97
(select from fact group by) -- uses two-stage grouping due to memory peakmem 2GB
join
(select from bigdimension) — ~67 mln + fat strings, peakmem 6GB
UPD:
Problem is happening when using LowCardinality on the right side table.
Query to reproduce (although depending on conditions it may fail with failure to allocate memory):
select
attr.k1,
attr.dim1,
fact.k1,
fact.gk1val,
fact.gk2val,
fact.gk3val from
(
select k1, k2,
(groupArrayInsertAt(Null, 3)(tuple(toNullable(val)), toUInt32(indexOf(range(3), gk) - 1)).1 as arr)[1] gk1val,
arr[2] gk2val,arr[3] gk3val
from (
select cityHash64(number) as k1, toString(cityHash64(number,2)) as k2, toString(cityHash64(number,gk)) as val, gk from numbers(1000000)
array join range(3) as gk
) group by k1, k2
) as fact
JOIN
(
select cityHash64(number) as k1, toLowCardinality(arrayStringConcat(arrayMap(x-> toString(x),range(cityHash64(number)%100)),'-')) dim1 from numbers(10000000)
)
as attr using k1
SETTINGS max_bytes_before_external_group_by='500M',join_algorithm='partial_merge'
Format Null;
Same query works correctly without having "toLowCardinality" (but too slow), and does not utilize much memory.
Expected behavior
clickhouse query shall not result in segmentation faults and query shall be completed successfully.
Error message and/or stacktrace
2020.06.19 15:13:30.829653 [ 14927 ] {} <Fatal> BaseDaemon: (version 20.3.11.97 (official build)) (from thread 14872) (query_id: 864f09d2-8cc0-44df-8a44-72fc974269e8) Received signal Segmentation fault (11).
2020.06.19 15:13:30.829695 [ 14927 ] {} <Fatal> BaseDaemon: Address: 0x1 Access: read. Address not mapped to object.
2020.06.19 15:13:30.829718 [ 14927 ] {} <Fatal> BaseDaemon: Stack trace: 0xd5c7718 0xd445dbe 0xd446c24 0xd4540a0 0xd0c8f83 0xd0ccc7e 0xde3f812 0xde428c2 0xdb75075 0xdbb9581 0xdbbd70d 0xdbbe0d2 0x8f757e7 0x8f73c33 0x7fdf6df78ea5 0x7fdf6e7958dd
2020.06.19 15:13:30.879340 [ 14927 ] {} <Fatal> BaseDaemon: 3. DB::ColumnString::insertFrom(DB::IColumn const&, unsigned long) @ 0xd5c7718 in /usr/bin/clickhouse
2020.06.19 15:13:30.879432 [ 14927 ] {} <Fatal> BaseDaemon: 4. ? @ 0xd445dbe in /usr/bin/clickhouse
2020.06.19 15:13:30.879490 [ 14927 ] {} <Fatal> BaseDaemon: 5. DB::MergeJoin::allInnerJoin(DB::MergeJoinCursor&, DB::Block const&, DB::Block const&, std::__1::vector<COW<DB::IColumn>::mutable_ptr<DB::IColumn>, std::__1::allocator<COW<DB::IColumn>::mutable_ptr<DB::IColumn> > >&, std::__1::vector<COW<DB::IColumn>::mutable_ptr<DB::IColumn>, std::__1::allocator<COW<DB::IColumn>::mutable_ptr<DB::IColumn> > >&, unsigned long&, unsigned long&) @ 0xd446c24 in /usr/bin/clickhouse
2020.06.19 15:13:30.879541 [ 14927 ] {} <Fatal> BaseDaemon: 6. void DB::MergeJoin::joinSortedBlock<false, true>(DB::Block&, std::__1::shared_ptr<DB::ExtraBlock>&) @ 0xd4540a0 in /usr/bin/clickhouse
2020.06.19 15:13:30.879577 [ 14927 ] {} <Fatal> BaseDaemon: 7. DB::ExpressionAction::execute(DB::Block&, bool, std::__1::shared_ptr<DB::ExtraBlock>&) const @ 0xd0c8f83 in /usr/bin/clickhouse
2020.06.19 15:13:30.879606 [ 14927 ] {} <Fatal> BaseDaemon: 8. DB::ExpressionActions::execute(DB::Block&, std::__1::shared_ptr<DB::ExtraBlock>&, unsigned long&) const @ 0xd0ccc7e in /usr/bin/clickhouse
2020.06.19 15:13:30.879639 [ 14927 ] {} <Fatal> BaseDaemon: 9. DB::InflatingExpressionTransform::readExecute(DB::Chunk&) @ 0xde3f812 in /usr/bin/clickhouse
2020.06.19 15:13:30.879664 [ 14927 ] {} <Fatal> BaseDaemon: 10. DB::InflatingExpressionTransform::transform(DB::Chunk&) @ 0xde428c2 in /usr/bin/clickhouse
2020.06.19 15:13:30.879727 [ 14927 ] {} <Fatal> BaseDaemon: 11. DB::ISimpleTransform::work() @ 0xdb75075 in /usr/bin/clickhouse
2020.06.19 15:13:30.879751 [ 14927 ] {} <Fatal> BaseDaemon: 12. ? @ 0xdbb9581 in /usr/bin/clickhouse
2020.06.19 15:13:30.879782 [ 14927 ] {} <Fatal> BaseDaemon: 13. DB::PipelineExecutor::executeSingleThread(unsigned long, unsigned long) @ 0xdbbd70d in /usr/bin/clickhouse
2020.06.19 15:13:30.879805 [ 14927 ] {} <Fatal> BaseDaemon: 14. ? @ 0xdbbe0d2 in /usr/bin/clickhouse
2020.06.19 15:13:30.879834 [ 14927 ] {} <Fatal> BaseDaemon: 15. ThreadPoolImpl<std::__1::thread>::worker(std::__1::__list_iterator<std::__1::thread, void*>) @ 0x8f757e7 in /usr/bin/clickhouse
2020.06.19 15:13:30.879866 [ 14927 ] {} <Fatal> BaseDaemon: 16. ? @ 0x8f73c33 in /usr/bin/clickhouse
2020.06.19 15:13:30.879914 [ 14927 ] {} <Fatal> BaseDaemon: 17. start_thread @ 0x7ea5 in /usr/lib64/libpthread-2.17.so
2020.06.19 15:13:30.879959 [ 14927 ] {} <Fatal> BaseDaemon: 18. __clone @ 0xfe8dd in /usr/lib64/libc-2.17.so
Additional context
partial_merge was tried to workaround out of memory problem when right table is too big. Seems bad luck...