Reuse buffers while querying vector indexes #1980
Closed
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Currently we allocate 2 buffers of size K for every segment available in the snapshot. In a 'full-flow' indexing scenario, we can have a large number of segments, so for each query incident on the index, we allocate a total space of
S = numSegments * K * 12 Bytes
. Although the semaphore preventsS
from being allocated instantaneously, the total space will still be allocated and GC'd, leaving behind a large memory footprint for every query. With this approach, the space allocated per query isS' = min(10, numSegments) * K * 12 Bytes
which drastically reduces the memory footprint left behind per query as it becomes independent of the number of segments available in the systemi know this breaks interfaces; if it is bad, I will close the pull requests associated with this
please let me know what you think about this