repair: Use merge_to_gently to merge two lists

During a performance test, test_latency_read_with_nemesis during manager
repair, it experienced a stall of 73 ms:

```
 (inlined by) std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > >::operator=(repair_row const&) at /usr/include/c++/9/bits/stl_iterator.h:515
 (inlined by) std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > std::__copy_move<false, false, std::bidirectional_iterator_tag>::__copy_m<std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > >(std::_List_iterator<repair_row>, std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > >) at /usr/include/c++/9/bits/stl_algobase.h:312
 (inlined by) std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > std::__copy_move_a<false, std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > >(std::_List_iterator<repair_row>, std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > >) at /usr/include/c++/9/bits/stl_algobase.h:404
 (inlined by) std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > std::__copy_move_a2<false, std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > >(std::_List_iterator<repair_row>, std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > >) at /usr/include/c++/9/bits/stl_algobase.h:440
 (inlined by) std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > std::copy<std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > >(std::_List_iterator<repair_row>, std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > >) at /usr/include/c++/9/bits/stl_algobase.h:474
 (inlined by) std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > std::__merge<std::_List_iterator<repair_row>, std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > >, __gnu_cxx::__ops::_Iter_comp_iter<repair_meta::apply_rows_on_master_in_thread(std::__cxx11::list<partition_key_and_mutation_fragments, std::allocator<partition_key_and_mutation_fragments> >, gms::inet_address, seastar::bool_class<update_working_row_buf_tag>, seastar::bool_class<update_peer_row_hash_sets_tag>, unsigned int)::{lambda(repair_row const&, repair_row const&)#1}> >(std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > >, std::_List_iterator<repair_row>, std::_List_iterator<repair_row>, __gnu_cxx::__ops::_Iter_comp_iter<repair_meta::apply_rows_on_master_in_thread(std::__cxx11::list<partition_key_and_mutation_fragments, std::allocator<partition_key_and_mutation_fragments> >, gms::inet_address, seastar::bool_class<update_working_row_buf_tag>, seastar::bool_class<update_peer_row_hash_sets_tag>, unsigned int)::{lambda(repair_row const&, repair_row const&)#1}>, __gnu_cxx::__ops::_Iter_comp_iter<repair_meta::apply_rows_on_master_in_thread(std::__cxx11::list<partition_key_and_mutation_fragments, std::allocator<partition_key_and_mutation_fragments> >, gms::inet_address, seastar::bool_class<update_working_row_buf_tag>, seastar::bool_class<update_peer_row_hash_sets_tag>, unsigned int)::{lambda(repair_row const&, repair_row const&)#1}>) at /usr/include/c++/9/bits/stl_algo.h:4923
 (inlined by) std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > > std::merge<std::_List_iterator<repair_row>, std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > >, repair_meta::apply_rows_on_master_in_thread(std::__cxx11::list<partition_key_and_mutation_fragments, std::allocator<partition_key_and_mutation_fragments> >, gms::inet_address, seastar::bool_class<update_working_row_buf_tag>, seastar::bool_class<update_peer_row_hash_sets_tag>, unsigned int)::{lambda(repair_row const&, repair_row const&)#1}>(std::_List_iterator<repair_row>, std::back_insert_iterator<std::__cxx11::list<repair_row, std::allocator<repair_row> > >, std::_List_iterator<repair_row>, std::_List_iterator<repair_row>, repair_meta::apply_rows_on_master_in_thread(std::__cxx11::list<partition_key_and_mutation_fragments, std::allocator<partition_key_and_mutation_fragments> >, gms::inet_address, seastar::bool_class<update_working_row_buf_tag>, seastar::bool_class<update_peer_row_hash_sets_tag>, unsigned int)::{lambda(repair_row const&, repair_row const&)#1}, repair_meta::apply_rows_on_master_in_thread(std::__cxx11::list<partition_key_and_mutation_fragments, std::allocator<partition_key_and_mutation_fragments> >, gms::inet_address, seastar::bool_class<update_working_row_buf_tag>, seastar::bool_class<update_peer_row_hash_sets_tag>, unsigned int)::{lambda(repair_row const&, repair_row const&)#1}) at /usr/include/c++/9/bits/stl_algo.h:5018
 (inlined by) repair_meta::apply_rows_on_master_in_thread(std::__cxx11::list<partition_key_and_mutation_fragments, std::allocator<partition_key_and_mutation_fragments> >, gms::inet_address, seastar::bool_class<update_working_row_buf_tag>, seastar::bool_class<update_peer_row_hash_sets_tag>, unsigned int) at ./repair/row_level.cc:1242
repair_meta::get_row_diff_source_op(seastar::bool_class<update_peer_row_hash_sets_tag>, gms::inet_address, unsigned int, seastar::rpc::sink<repair_hash_with_cmd>&, seastar::rpc::source<repair_row_on_wire_with_cmd>&) at ./repair/row_level.cc:1608
repair_meta::get_row_diff_with_rpc_stream(std::unordered_set<repair_hash, std::hash<repair_hash>, std::equal_to<repair_hash>, std::allocator<repair_hash> >, seastar::bool_class<needs_all_rows_tag>, seastar::bool_class<update_peer_row_hash_sets_tag>, gms::inet_address, unsigned int) at ./repair/row_level.cc:1674
row_level_repair::get_missing_rows_from_follower_nodes(repair_meta&) at ./repair/row_level.cc:2413
```

The problem was that when std::merge() ran out of one range, it copied the second range.

To fix, use the new merge_to_gently helper.

Fixes #6976
This commit is contained in:
Asias He
2020-08-10 10:47:45 +08:00
parent 0bf0019eeb
commit 53fee789f0

View File

@@ -48,6 +48,7 @@
#include "gms/gossiper.hh"
#include "repair/row_level.hh"
#include "mutation_source_metadata.hh"
#include "utils/stall_free.hh"
extern logging::logger rlogger;
@@ -1261,13 +1262,11 @@ private:
stats().rx_row_nr += row_diff.size();
stats().rx_row_nr_peer[from] += row_diff.size();
if (update_buf) {
std::list<repair_row> tmp;
tmp.swap(_working_row_buf);
// Both row_diff and _working_row_buf and are ordered, merging
// two sored list to make sure the combination of row_diff
// and _working_row_buf are ordered.
std::merge(tmp.begin(), tmp.end(), row_diff.begin(), row_diff.end(), std::back_inserter(_working_row_buf),
[this] (const repair_row& x, const repair_row& y) { thread::maybe_yield(); return _cmp(x.boundary(), y.boundary()) < 0; });
utils::merge_to_gently(_working_row_buf, row_diff,
[this] (const repair_row& x, const repair_row& y) { return _cmp(x.boundary(), y.boundary()) < 0; });
}
if (update_hash_set) {
_peer_row_hash_sets[node_idx] = boost::copy_range<repair_hash_set>(row_diff |