feat: update for inmemory index - #1724
Conversation
Sorry, something went wrong.
Signed-off-by: jupyterjazz <saba.sturua@jina.ai>
Codecov ReportPatch coverage: 46.15% and project coverage change: -0.04 ⚠️ Additional details and impacted files @@ Coverage Diff @@
## main #1724 +/- ##
==========================================
- Coverage 85.57% 85.54% -0.04%
==========================================
Files 133 133
Lines 8592 8608 +16
==========================================
+ Hits 7353 7364 +11
- Misses 1239 1244 +5
Flags with carried forward coverage won't be shown. Click here to find out more.
☔ View full report in Codecov by Sentry. |
Sorry, something went wrong.
|
Can you also add how much more memory is used? I am not sure this is desired. |
Sorry, something went wrong.
|
Not much change in terms of memory too. Here's the experiment code from docarray.index import InMemoryExactNNIndex
from docarray import BaseDoc
from docarray.typing import NdArray
import numpy as np
import tracemalloc
tracemalloc.start()
class MyDoc(BaseDoc):
text: str
embedding: NdArray[128]
data = [MyDoc(text=f'text {i}', embedding=np.random.rand(128)) for i in range(200000)]
doc_index = InMemoryExactNNIndex[MyDoc]()
doc_index.index(data)
docs, scores = doc_index.find(data[0], search_field='embedding')
if data[10] in doc_index:
print('wohoo')
ids_to_get = [data[200].id, data[250].id, data[350].id]
docs = doc_index[ids_to_get]
ids_to_del = [data[200].id, data[250].id, data[350].id]
del doc_index[ids_to_del]
print(tracemalloc.get_traced_memory())
tracemalloc.stop()Before change prints (768773718, 1032141402) showing current memory usage and max memory usage |
Sorry, something went wrong.
|
📝 Docs are deployed on https://ft-feat-inmemory-update--jina-docs.netlify.app 🎉 |
Sorry, something went wrong.
Implements Update for InMemoryExactNNIndex by creating a mapping between document IDs and their corresponding positions in the DocList.
In terms of performance, we see significant ~130x (from 0.0266s to 0.0002s) improvement in get, and a slight increase in the index time - approx. from 3.50s to 3.57s; Other operation times stayed very similar. Experiment was made on 200 000 docs, vector dim 128.