Add compile time feature-gate for lookups and removals #170
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Ngalstyan4
commented
Jul 31, 2023
| @@ -609,12 +638,14 @@ class index_punned_dense_gt { | |||
| auto typed_result = typed_->copy(config); | |||
| if (!typed_result) | |||
| return result.failed(std::move(typed_result.error)); | |||
| #if USEARCH_LOOKUP_LABEL | |||
| if (!result.index.free_ids_.reserve(free_ids_.size())) | |||
| return result.failed(std::move(typed_result.error)); | |||
| for (std::size_t i = 0; i != free_ids_.size(); ++i) | |||
| result.index.free_ids_.push(free_ids_[i]); | |||
|
|
|||
| result.index.labeled_lookup_ = labeled_lookup_; | |||
There was a problem hiding this comment.
where would copy be used?
Can the old index be used while/after copying?
if so, should this not keep track of the old lookup table lock as well?
|
Thanks, @Ngalstyan4! It’s a valid suggestion. We can also solve it differently, by adding a single template argument. Will discuss that today ;) |
3 tasks
ashvardanian
added a commit
to ashvardanian/usearch
that referenced
this pull request
Aug 4, 2023
|
@Ngalstyan4, just to clarify, memory savings from removing the lookup table are more important than the ability to remove entries? |
|
Right |
ashvardanian
pushed a commit
that referenced
this pull request
Aug 5, 2023
# [0.23.0](v0.22.3...v0.23.0) (2023-08-05) ### Add * `Matches` and `BatchMatches` simple API ([1b40f13](1b40f13)) * Add node offsets in a serialized file ([c600ffd](c600ffd)) * Batch add ([74860d6](74860d6)) * Batch add test ([5f99b05](5f99b05)) * Changing the metric at runtime ([d7bfac7](d7bfac7)) * Compactions ([434c1da](434c1da)) * efficiency estimate in `recall_members` ([64a60b4](64a60b4)) * Exact search shortcut ([a005084](a005084)), closes [#176](#176) * Multi-`Index` lookups ([c5b7ccd](c5b7ccd)) * Parallel View ([ed3f845](ed3f845)) * Prefetching functionality for external memory ([b544ddb](b544ddb)), closes [#170](#170) [#171](#171) * Streaming and in-memory serialization in C++ ([7da44a2](7da44a2)) * Vector alignment ([ea230e0](ea230e0)) ### Break * Final patches for 1.0 release ([8d557e2](8d557e2)) ### Docs * add descriptions of match-related classes ([637e5ef](637e5ef)) * Annotating C 99 and GoLang interfaces ([4b910a8](4b910a8)) * Documenting Python tests ([1f89e0a](1f89e0a)) * Shorten name ([9a6a01c](9a6a01c)) * Spelling and details ([6f25ed9](6f25ed9)) * Spelling and links ([20566e0](20566e0)) * TypeScript docs factual errors ([fe8103c](fe8103c)) * Update benchmarking sections ([96baa09](96baa09)) ### Fix * `reset` and serialization code ([11d7844](11d7844)) * Avoid exception in `.xbin` file is missing ([4863bea](4863bea)) * Avoid spawning needless threads ([9dff0fb](9dff0fb)) * Concurrent file access issues in tests ([5ae6db1](5ae6db1)) * Dead-lock on post-removal insertions ([284b058](284b058)), closes [#175](#175) * Excpetion handling for `index_dense_metadata` ([d9627ba](d9627ba)) * Heap overflow for fractional-size scalars ([459abcd](459abcd)) * Imports in Python benchmarks ([cffe507](cffe507)) * Inferring OS-dependant file path in Python ([7743709](7743709)), closes [#174](#174) * JavaScript bindings ([ee04856](ee04856)) * JS keys should be `bigint` ([e1fbec4](e1fbec4)), closes [#178](#178) * Memory leak and multi-index lookup overflow ([597b0d5](597b0d5)) * Narrowing conversions for WASM 32-bit builds ([79add97](79add97)) * Portable way of matching 32-bit builds ([604e634](604e634)) * Progress reporting issue ([b2565e5](b2565e5)) * Reclaiming file descriptor ([05e908f](05e908f)) * Report error if `reserve` hasn't been called ([f94f358](f94f358)) * Typo in metric name ([34f5530](34f5530)) * Undefined behaviour on duplicate labels ([c04a5cc](c04a5cc)) ### Improve * `usearch_remove` C99 interface ([2072540](2072540)) * Align allocations to page size by default ([134a6f0](134a6f0)) * Broader types support in `usearch.io` ([b1a1439](b1a1439)) * Exposing search stats to users ([2779ffc](2779ffc)) * Feature-complete GoLang bindings ([e2058d1](e2058d1)) * More flexibility for Python args ([6aa06cb](6aa06cb)) * Out-of-bounds checks ([54cecb6](54cecb6)) * Task scheduling with STL threads ([9131287](9131287)) ### Make * Add CMake for C builds ([4d2127b](4d2127b)) * All targets enabled for debugging ([ea0f835](ea0f835)) * Build only WASM tests ([372738b](372738b)) * Typescript ([dacfbed](dacfbed)) * Upgrade to the newest SimSIMD ([368d853](368d853)) ### Refactor * `label_t` to `key_t` ([0d6c800](0d6c800)) * Add ([5d62180](5d62180)) * Index serialization in a file ([ba72585](ba72585)) * JS and GoLang tests ([a45fc40](a45fc40)) * Keep only batch requests in CPython ([44c0318](44c0318)) * Rename `f8` to `i8` to match IEEE ([c37f80b](c37f80b)) * Revert `Matches` ([5731e70](5731e70)) * Splitting proximity-graphs and vectors ([e996b38](e996b38)) * Use Executor instead of std::thread ([c3a3693](c3a3693)) * Vector alignment issue ([b02d0ad](b02d0ad)) ### Test * Set vector alignment ([0acb54a](0acb54a)) * Wrong buffer size caused illegal access ([830e280](830e280))
ashvardanian
force-pushed
the
main-dev
branch
5 times, most recently
from
9163d79 to
bc26b4e
Compare
ashvardanian
force-pushed
the
main-dev
branch
from
d880374 to
2306ff2
Compare
ashvardanian
force-pushed
the
main-dev
branch
5 times, most recently
from
871b788 to
6461df2
Compare
ashvardanian
force-pushed
the
main-dev
branch
10 times, most recently
from
06c4d75 to
74e98fa
Compare
ashvardanian
force-pushed
the
main-dev
branch
2 times, most recently
from
d79b296 to
5ff8e83
Compare
ashvardanian
force-pushed
the
main-dev
branch
6 times, most recently
from
ff5cf1d to
7d17eab
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
I am not sure you guys want this in upstream usearch. But, we use this for the following:
For our use-case, we only need vector insertions and search.
Currently, usearch always builds a lookup table for vector labels at index creation time (e.g. here) to support
contains, get, removequeries. I think there are many use-cases that do not need these (including our use-case) so it would be good to not pay the cpu and memory cost of this lookup table creation.This PR feature-gates the index creation functions and all the methods depending on them so one can easily disable them and remove their runtime cost.
Perhaps the one-time lookup table creation cost and the small maintenance cost is trivial for most use-cases. But, for our case, we need to re-initialize the usearch context for each search query so it is crucial to minimize its memory footprint. This is likely not the typical case for usearch, but this is one case to consider.
I considered leaving the code in and just adding a runtime flag to enable lookups. But preferred this approach because with this the compiler and the IDE did more to help ensure available APIs are used.