Beihao-Zhou
commented
1 month ago Thanks for helping benchmark it! I'll fix the bug ASAP
Open PragmaTwice opened 1 month ago
Beihao-Zhou
commented
1 month ago Thanks for helping benchmark it! I'll fix the bug ASAP
Beihao-Zhou
commented
1 month ago Hiii @PragmaTwice,
I have some updates on the issue. <3
I tried debugging using gdb with the following steps:
➜ kvrocks git:(unstable) ✗ gdb ./build/kvrocks
(gdb) run -c kvrocks.confAfter running python3 run.py --engines redis-default --datasets 'glove-25-angular', I reproduced the error and obtained the following backtrace:
(gdb) bt
[......]
#9 0x000055555561893a in std::__throw_bad_alloc() ()
#10 0x000055555651e6be in handleOOM (size=14411505886035969, nothrow=<optimized out>)
at /home/ubuntu/kvrocks/build/_deps/jemalloc-src/src/jemalloc_cpp.cpp:90
#11 0x00005555556c4fb4 in std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >::_M_construct<char*> (
this=0x7ffff19f28a0, __beg=0x33010000000200 <error: Cannot access memory at address 0x33010000000200>,
__end=0x66343031040200 <error: Cannot access memory at address 0x66343031040200>)
at /usr/include/c++/11/bits/basic_string.tcc:219
#12 0x000055555568925a in redis::VectorItem::VectorItem (this=<optimized out>, this=<optimized out>)
at /home/ubuntu/kvrocks/src/search/hnsw_indexer.h:57
#13 0x00005555559bb5d3 in operator() (__closure=0x7ffff19f25f0, candidate_neighbours=..., selected_neighbours=...)
at /home/ubuntu/kvrocks/src/search/hnsw_indexer.cc:432
#14 0x00005555559bc4c3 in redis::HnswIndex::InsertVectorEntryInternal (this=0x7ffff19f2a90, key=..., vector=..., batch=...,
target_level=0) at /home/ubuntu/kvrocks/src/search/hnsw_indexer.cc:436
#15 0x00005555559bd6f9 in redis::HnswIndex::InsertVectorEntry (this=0x7ffff19f2a90, key=..., vector=..., batch=...)
at /home/ubuntu/kvrocks/src/search/hnsw_indexer.cc:498
#16 0x00005555559d3e7a in redis::IndexUpdater::UpdateHnswVectorIndex (this=0x7ffff402f000, key=..., original=..., current=...,
search_key=..., vector=0x7ffff5c08380) at /home/ubuntu/kvrocks/src/search/indexer.cc:287
#17 0x00005555559d4633 in redis::IndexUpdater::UpdateIndex (this=0x7ffff402f000, field=..., key=..., original=..., current=...)
at /home/ubuntu/kvrocks/src/search/indexer.cc:314
[......]From line 13, we can see that there is a closure attempting to access a VectorItem but failing, causing a memory leak. It turns out that it is trying to dereference a null iterator. The patch at the end will temporarily fix the crash.
However, logically, there should be no situation where deleted_node_it gets a null iterator. I suspect this issue is due to concurrent requests for index updating, where individual updates access inconsistent snapshots and mess up the node metadata. Currently, HnswIndex doesn't handle concurrency issues.
# Kvrocks log
I20240812 00:40:37.942303 185593 worker.cc:111] [worker] New connection: fd=56 from port: 6666 thread #140737224099392
I20240812 00:40:37.997772 185593 worker.cc:111] [worker] New connection: fd=57 from port: 6666 thread #140737224099392
...
...
I20240812 00:40:39.056308 185592 worker.cc:111] [worker] New connection: fd=71 from port: 6666 thread #140737232492096Given this, I guess addressing the consistency issue will probably in turn resolve the issue here. I saw there is a PR #2310 to solve this? And should we solve this issue after the PR is finalized and merged? Let me know if you believe there are other possibilities/solutions we should consider, as I'm not entirely certain this is the underlying cause. Thanks!
diff --git a/src/search/hnsw_indexer.cc b/src/search/hnsw_indexer.cc
index 3618ad8..e8271b8 100644
--- a/src/search/hnsw_indexer.cc
+++ b/src/search/hnsw_indexer.cc
@@ -385,7 +385,7 @@ Status HnswIndex::InsertVectorEntryInternal(std::string_view key, const kqir::Nu
// Check if candidate node has room after some other nodes' are pruned in current batch
auto has_room_after_deletions = [&](const HnswNode& candidate_node, uint16_t candidate_node_num_neighbours) {
- auto it = deleted_edges_map.find(candidate_node.key);
+ const auto& it = deleted_edges_map.find(candidate_node.key);
if (it != deleted_edges_map.end()) {
auto num_deleted_edges = static_cast<uint16_t>(it->second.size());
return (candidate_node_num_neighbours - num_deleted_edges) < m_max;
@@ -417,27 +417,30 @@ Status HnswIndex::InsertVectorEntryInternal(std::string_view key, const kqir::Nu
std::find(sorted_neighbours_by_distance.begin(), sorted_neighbours_by_distance.end(),
inserted_vector_item) != sorted_neighbours_by_distance.end();
- if (inserted_node_is_selected) {
- // Add the edge between candidate and inserted node
- GET_OR_RET(AddEdge(inserted_vector_item.key, candidate_node.key, level, batch));
- connected_edges_set.insert(candidate_node.key);
+ if (!inserted_node_is_selected) {
+ continue;
+ }
- auto find_deleted_item = [&](const std::vector<VectorItem>& candidate_neighbours,
- const std::vector<VectorItem>& selected_neighbours) -> VectorItem {
- auto it =
- std::find_if(candidate_neighbours.begin(), candidate_neighbours.end(), [&](const VectorItem& item) {
- return std::find(selected_neighbours.begin(), selected_neighbours.end(), item) ==
- selected_neighbours.end();
- });
- return *it;
- };
-
- // Remove the edge for candidate and the pruned node
- auto deleted_node = find_deleted_item(candidate_node_neighbour_vec_items, sorted_neighbours_by_distance);
- GET_OR_RET(RemoveEdge(deleted_node.key, candidate_node.key, level, batch));
- deleted_edges_map[candidate_node.key].insert(deleted_node.key);
- deleted_edges_map[deleted_node.key].insert(candidate_node.key);
+ auto deleted_node_it =
+ std::find_if(candidate_node_neighbour_vec_items.begin(), candidate_node_neighbour_vec_items.end(),
+ [&](const VectorItem& item) {
+ return std::find(sorted_neighbours_by_distance.begin(), sorted_neighbours_by_distance.end(),
+ item) == sorted_neighbours_by_distance.end();
+ });
+
+ if (deleted_node_it == candidate_node_neighbour_vec_items.end()) {
+ continue;
}
+ auto deleted_node = *deleted_node_it;
+
+ // Remove the edge for candidate and the pruned node
+ GET_OR_RET(RemoveEdge(deleted_node.key, candidate_node.key, level, batch));
+ deleted_edges_map[candidate_node.key].insert(deleted_node.key);
+ deleted_edges_map[deleted_node.key].insert(candidate_node.key);
+
+ // Add the edge between candidate and inserted node
+ GET_OR_RET(AddEdge(inserted_vector_item.key, candidate_node.key, level, batch));
+ connected_edges_set.insert(candidate_node.key);
}
+ auto deleted_node = *deleted_node_it;
+
+ // Remove the edge for candidate and the pruned node
+ GET_OR_RET(RemoveEdge(deleted_node.key, candidate_node.key, level, batch));
+ deleted_edges_map[candidate_node.key].insert(deleted_node.key);
+ deleted_edges_map[deleted_node.key].insert(candidate_node.key);
+
+ // Add the edge between candidate and inserted node
+ GET_OR_RET(AddEdge(inserted_vector_item.key, candidate_node.key, level, batch));
+ connected_edges_set.insert(candidate_node.key);
}
// Update inserted node metadata
PragmaTwice
commented
1 month ago Thank you for looking into this!
I agree that it's likely related to rocksdb concurrency and transaction management. However, besides ensuring the consistency of rocksdb snapshots, do we need to consider anything else when constructing HNSW indexes in multiple threads? Should we also introduce some locks?
Beihao-Zhou
commented
1 month ago Thank you for looking into this!
I agree that it's likely related to rocksdb concurrency and transaction management. However, besides ensuring the consistency of rocksdb snapshots, do we need to consider anything else when constructing HNSW indexes in multiple threads? Should we also introduce some locks?
Yea introducing locks sounds good to me. Also I think this is probably a common issue for all index rather than only HNSW (including text index in the future if also using quite complicated algorithm?), so we probably want to have a lock mechanism applicable for all. I'll look into it these two days, and open a new issue to discuss my idea & ask for suggestions if needed
Search before asking
Motivation
We can try to use this repo: https://github.com/qdrant/vector-db-benchmark
After some simple patching:
We can start a kvrocks instance and run this benchmark:
cc @Beihao-Zhou
Solution
Currently we'll get some coredumps in the vector indexing phase, e.g.:
Maybe we can try to solve it first.
Are you willing to submit a PR?