Memory leak detected.
jtong11 opened this issue · comments
Hello,
I use the crate-0.10.0 for our embedding recall, and have found some tips:
a. IndexImpl has not impl concurrentIndex, so IVFScalarQuantizerIndexImpl returned by into_ivf_scalar_quantizer builds error;
b. search increase the leaking of memory without upbound limit, so does the read_index, dropped but no release the RES memory.
What would the problem be? Anybody encount?
Thanks
Hello. Could you please provide a few examples that reproduce the problems reported? This would help to better understand the underlying issues with the bindings.
a. IndexImpl has not impl concurrentIndex, so IVFScalarQuantizerIndexImpl returned by into_ivf_scalar_quantizer builds error;
IndexImpl
should definitely not implement ConcurrentIndex
because it is unknown whether shared access to the index is safe. At best, there may be a poor integration of that with the IVF scalar quantizer.
Sry for reply so late. @Enet4
Here is an example code. Mock as a server, its memory will raises rapidly from 2M to 10+G and without upper bound.
I build it with "rustc 1.58.0-nightly (b426445c6 2021-11-24)",and faiss "c_api_v1.7.1" with "cmake -B build -DFAISS_ENABLE_C_API=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DFAISS_ENABLE_GPU=false" by "g++ (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0".
extern crate faiss;
extern crate rand; //rand = "0.3.17"
fn random_embedding(rng: &mut ThreadRng, dims: usize) -> Vec<f32> {
let mut embedding = Vec::with_capacity(dims);
for _ in 0..dims {
embedding.push(rng.gen_range(-1.0, 1.0));
}
return embedding;
}
fn main() {
let mut rng = rand::thread_rng();
let dims = 64u32;
let doc_count = 10000usize;
{
let mut ids = Vec::with_capacity(doc_count);
let mut embeddings = Vec::with_capacity(doc_count * dims as usize);
for i in 0..doc_count {
ids.push(Idx::new(i as u64));
embeddings.extend_from_slice(random_embedding(&mut rng, dims as usize).as_slice());
}
let mut index =
IVFFlatIndexImpl::new_ip(faiss::FlatIndex::new_ip(dims).unwrap(), dims, 10).unwrap();
index.set_verbose(true);
index.train(&embeddings).unwrap();
index.add_with_ids(&embeddings, &ids).unwrap();
write_index(&index, "/tmp/test.index").unwrap();
}
loop {
let query = random_embedding(&mut rng, dims as usize);
let mut index = read_index("/tmp/test.index").unwrap();
let k = 100;
let mut count = 0;
loop {
if count > 10000 {
break;
}
let result = index.search(&query, k);
count += 1;
if count % 1000 == 0 {
println!("{} {:?}", count, result);
}
}
sleep(Duration::from_secs(10));
}
}
This seems as a issue about memory leak in search case.
facebookresearch/faiss#2054