Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Removing items from the index leaves the index with an inconsistent size...and therefore state #355

Open
2 tasks done
spicymatt opened this issue Feb 26, 2024 · 2 comments
Labels
bug Something isn't working

Comments

@spicymatt
Copy link

spicymatt commented Feb 26, 2024

Describe the bug

The following code (Swift) adds 500 items to the index, and then removes them.
The count of items in the index after removal is 448 (consistently ... when we expect 0)
Search for items does not seem to return any of the original items, but the size of the index is inconsistent.

If we try to add the items back by reserving the expected space (500), it crashes when we add() because of insufficient capacity.
If we reserve more space (448 + 500) and add the 500, it works, and the final count is 500, as expected.

It seems that after a removal, the index has some nodes that are not properly removed.

We tested the compact() method after removal and it does not help...still 448 items in the index

Steps to reproduce

Here is the code:

func testSimpleRemove() async throws {
    // build the vectors
    let numberOfVectors:UInt32 = 500
    let vectorLength:UInt32 = 512
    var idList:[UInt64] = []
    let numberOfInsersions = 20

    var signatures: [USearchSignature] = []
    for index in 0..<numberOfVectors {
        var vector: [Float] = []
        for i in 0..<vectorLength {
            let randomValue = Float(index)
            vector.append(randomValue * Float(i))
        }
        signatures.append(USearchEntry(key: UInt64(index), values: vector))
    }
    let index = USearchIndex.make(metric: USearchMetric.l2sq, dimensions: vectorLength, connectivity: 8, quantization: USearchScalar.F32)
    var clusterSize = index.count
    XCTAssertEqual(clusterSize, 0)

    index.reserve(UInt32(signatures.count))
    clusterSize = index.count
    XCTAssertEqual(clusterSize, 0, "still empty")

    for signature in signatures {
        index.add(key: signature.us_id, vector: signature.us_vector)
    }
    clusterSize = index.count
    XCTAssertEqual(clusterSize, 500)
    var results = index.search(vector: signatures[10].us_vector as [Float], count: 10)
    XCTAssertEqual(results.0.first, 10)
    XCTAssertEqual(results.1.first, 0.0)
    print("Find :\(results)")

    for signature in signatures {
        index.remove(key: signature.us_id)
    }
    XCTAssertEqual(index.contains(key: 0), false)

    index.compact()
    clusterSize = index.count
    // ⚠️ this musn't fail -> it returns clusterSize = 448
    XCTAssertEqual(clusterSize, 0)
    
    // no item from initial set should be found
    for signature in signatures {
        let resultsAfterRemoval = index.search(vector: signature.us_vector as [Float], count: 10)
        XCTAssertTrue(resultsAfterRemoval.0.first ?? 0 > signatures.count,"no item from initial set should be found")
    }
    
    // add the items back --> Crash as the capacity is not sufficient
    index.reserve(UInt32(signatures.count + index.count) ) // we reserve the expected size
    clusterSize = index.count
    // ⚠️ this musn't fail -> it returns clusterSize = 448
    XCTAssertEqual(clusterSize, 0, "should be empty")

    for signature in signatures {
        index.add(key: signature.us_id, vector: signature.us_vector)
    }
    clusterSize = index.count
    XCTAssertEqual(clusterSize, 500)

    results = index.search(vector: signatures[10].us_vector as [Float], count: 10)
    XCTAssertEqual(results.0.first, 10)
    XCTAssertEqual(results.1.first, 0.0)
}

Expected behavior

I would expect the index.count to be 0 after a removal of everything.

USearch version

2.9

Operating System

macOS Sonoma

Hardware architecture

Arm

Which interface are you using?

Other bindings

Contact Details

Matthieu.kopp@gmail.com

Is there an existing issue for this?

  • I have searched the existing issues

Code of Conduct

  • I agree to follow this project's Code of Conduct
@spicymatt spicymatt added the bug Something isn't working label Feb 26, 2024
@spicymatt
Copy link
Author

hi, is there any progress on that task ? It seems that if you remove all vectors from the cluster, and then save the index, there is a crash. It is a side effect of this bug. Looks pretty important as an index with removed items is inconsistent

regards

Screenshot 2024-03-12 at 23 56 54

@ashvardanian
Copy link
Contributor

This snippet looks interesting. I'll try to reproduce it in the C layer and debug there. Thanks for the patience 🤗

ashvardanian added a commit that referenced this issue May 14, 2024
ashvardanian added a commit that referenced this issue Aug 6, 2024
* Improve: Swift test for issue #399 (#400)

* Fix: Integer overflow in aligned-alloc

Fixes ClickHouse/ClickHouse#61780

Co-authored-by: Antonio Andelic <antonio2368@users.noreply.github.com>

* Make: Disable Windows NPM builds

Relates to the #377 and the comment:
#377 (comment)

This temporarily disables the failing CI pipeline to generate
and update docs.

* Fix: Going beyond level 0 in clustering

* Improve: Error handling in `index_dense_gt`

This commit drops `std::vector` dependency,
making compilation time shorter and error
handling universal across abstraction layers.

* Improve: Remove `std::function` calls

* Improve: Remove `std::thread` from `index_dense_gt`

* Improve: `std::vector` -> `buffer_gt` in plugins

* Add: `usearch_change_threads_search`

* Fix: `index_dense_t::make(path)`

* Fix: Exhastive Search

In the past, if we got "too lucky" traversing the graph,
we could exit early before accumulating K top matches,
even if the index had more than K entries. This patch
changes that behavior, making output more predicatable.

* Fix: Replacement leaves isolated nodes

This patch addresses the issue #399, originally
observed in the Swift layer. Reimplementing it in
C++ helped locate the issue and lead to refactoring
the `update` procedure in the lowest-layer `index_gt`.

Now, `add` and `update` share less code. The `add`
is one branch shorter (not that it would be noticeable),
and `update` brings additional logic to avoid spilling
`updated_slot` into top-results and consequently
introducing self-loops.

Closes #399

* Fix: Misc warnings & compilation issues

* Fix: Misc warnings & compilation issues

* Fix: Detect `ring_gt` being full

Relates to #355

* Fix: Re-init `available_threads_` after load

Both `view` and `load` would `reset` the thread contexts.
After that, the very first `search` and `add` would fail, as
no thread-local contexts are initialized. It would require
a `reserve` call with a non-zero second arcgument to
define the number of concurrent threads, for which the
queues & buffers need to be allocated.

That design is counter-intuitive, so this patch re-inits the
same number of threads as before the `load` & `view` or
one, if none existed.

* Fix: `uint32_t` to `uint40_t` cast (#404)

Co-authored-by: Ash Vardanian <1983160+ashvardanian@users.noreply.github.com>

* Docs: Mention `b1` in `README.md`

Co-authored-by: Adolfo Garcia <1250775+adolfogc@users.noreply.github.com>

* Docs: Cover new users

* Improve: Updates stability & catch bug

* Fix: Dereferencing `member_iterator_t`

* Add: Java `get` API (#407)

* Fix: Compilation with `uint40_t` keys

* Add: `AutoClosable` using `c_destroy` for Java (#408)

* Fix: Rare deadlock on tiny collections

* Improve: `enable_key_lookups=false` memory usage

As noted by Robert Schulze, we can avoid populating `slot_lookup_`
during insertions, if `enable_key_lookups` is not set. This would lead
to lower memory consumption for large indexes of tiny vectors,
particularly common in GIS.

Co-authored-by: Robert Schulze <robert@clickhouse.com>

* Fix: Preset `enable_key_lookups=true` in C

* Fix: `std::is_pod` deprecated in C++20

* Fix: Unused type aliases

* Improve: Avoid `#pragma region` pre GCC 13 (#386)

* Do not use #pragma region if not supported by the compiler
* `pragma region` supported by GCC 13+

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85487

---------

Co-authored-by: Mikhail Bautin <mbautin@users.noreply.github.com>
Co-authored-by: Ash Vardanian <1983160+ashvardanian@users.noreply.github.com>

* Fix: Capacity checks in C# tests

* Docs: Add doc-strings in C#

* Make: Verbose C# logging in debug builds

* Docs: Describe serialization in GoLang

* Make: Drop Java & JS API references

Compiling and introspecting docstrings in Sphinx
is extremely flaky. It's safer to simply describe the
usage patterns in the `README.md`.

* Add: Load from path in Java (#410)

* Improve: Avoid sorting on small "refines"

* Docs: Cover JIT-compiled Python examples

* Fix: `index.copy()` trying to `memcpy(*, NULL, *)`

* Docs: UB in C++ Example (#415)

Co-authored-by: Ash Vardanian <1983160+ashvardanian@users.noreply.github.com>

* Improve: Catch UB with tests

* Docs: Rearrange

* Fix: Reserving contexts post-reload

* Improve: Detect more failures in tests

* Improve: Log failing lines

* Fix: Clamp before down-casting to `i8_t` (#422)

Previously we were down-casting floats to the target type (e.g. int8_t), and then
clamping to [-100, 100] range. This means that e.g. 129 would be cast to -127 and
then converted to -100, in stead of becoming 100

The fix does clamping first, and then casts the resulting number
(which is guaranteed to be in range [-100, 100], due to clamping)
from source type to target int8_t. Given the clamping, this will never overflow.

---------

Co-authored-by: Ash Vardanian <1983160+ashvardanian@users.noreply.github.com>

* Fix: Concurrency bug in high-K search

When calling `index_dense_gt`, the thread lock was not propagating
with the `search_result_t`. That is a an error-prone API. When too many
threads are running in parallel (ideally, more than physical CPU cores)
another thread may start reusing the `context_t` before the original
caller finishes exporting entries with `dump_to`.

This solution is backwards compatible and passes the tests.

* Make: Manually bump version to 2.13.0

We can't yet rely on the SemVer tool
semantic-release/release-notes-generator#633 (comment)

* Improve: Attempt to implement batch-metrics

* Add: Batch-capable metrics

* Add: `misaligned_ptr_gt` comparators

* Improve: Separate `error_t` constructors

This makes debugging easier, as it's simpler to trace
where the error message is being set.

* Improve: Support `enum` slots

Tracing implicit conversions of `std::uint32_t` and other
primitive types isn't always easy in concurrent apps. This
commit adds support for `enum` types to be used for
safer implementation of `index_gt` specializations.

* Improve: Ranges-V3 compatibility

* Add: Preliminary support for batch metrics

* Add: Batch-parallel refinements

* Add: `MANIFEST.in` for `py.typed` (#425)

Adding type annotation for Python native modules
solves the `Skipping analyzing "usearch.index" module` 
warning due to `missing library stubs or py.typed marker`.

Closes #424

* Fix: Clear cast buffer before bitwise ORs for `b1x8_t` (#428)

When converting floating point arrays to binary, we use bitwise OR
operations to set the relevant bits in the output buffer to 1. We do
nothing if the bit is zero, so we assume that the bit is zero to start
with. The `memset` statement makes sure this assumption holds.

* Fix: `esm` duplicate import bug in `jest` (#420)

Closes #418
Closes #426

* Fix: build.gradle deprecations

* Fix: ESM build support (#433)

Closes #426
Relates to #420

* Fix: `capacity()` assertion in Rust (#436)

Closes #432

* Fix: Computing `stats(i).max_edges`

* Add: Returning `computed_distances_in_refines`

In high-connectivity graphs, the number of distance
computations can be dominated by the number of "refine"
heuristic computations performed by the core structure.
The extended `add_result_t` now includes both:

- `computed_distances_in_refines`
- `computed_distances_in_reverse_refines`

This commit also extends the documentation.

* Add: Profiling attributes for `index_gt`

* Fix: Preserve thread limits after `fork()`

* Improve: Benchmarks self-recall support and `.bbin`

This patch adds support for partial datasets, without
ground truth neighborhood data. It also adds support
for `.bbin` binary, `.hbin` half-precision, and `.dbin`
double-precision input vector files/

* Fix: Printing progress between exit

* Docs: Fix spelling

* Fix: Wolfram bindings (#437)

Co-authored-by: Ash Vardanian <1983160+ashvardanian@users.noreply.github.com>

* Fix: Pre-reserve enough threads for C users

This indirectly fixes the crash in C# layer

* Revert: Parallel metrics

* Fix: Updating the `entry_slot_` node

* Improve: Enable single-threaded update tests

* Make: Bump version

* Fix: `flat_hash_multi_set_gt::reset` double-free

* Fix: Not enough memory in `fork()`

* Make: Bump to SimSIMD v5

* Fix: Missing SimSIMD v5 capability names

* Improve: Detecting copy & move issues

* Fix: Compilation w. explicit `template class`

* Improve: Bypass UBSan NULL dereferencing warning

* Improve: Minimize alignment issues

* Make: Disable `bf16` on MacOS

* Make: Link to GitHub repo

* Fix: Conditional `call_key` compilation in MSVC

* Fix: Unary minus on unsigned distances

* Make: Disable `bf16` in JS

* Fix: Compatibility with pre-v2.10

Closes #423

* Improve: Test wrong number of dimensions in Rust (#413)

Closes #412

---------

Co-authored-by: Julius Brummack <juliusbrummack@icloud.com>

* Fix: Handle wrong dimensionality in Rust

* Fix: Overwriting `SIMSIMD_DYNAMIC_DISPATCH`

* Make: Upgrade Java version

* Fix: Spelling mistakes

* Fix: `sprintf` deprecated

* Fix: `-Wpass-failed=transform-warning`

* Fix: Memory pinning on `Add` in C#

* Make: Specify Java distribution

* Fix: Pin memory in gets (C#)

* Make: Skip `PersistAndRestore` in CI on MacOS

* Make: Upgrade Docker action

This fixes a GitHub CI warning about the
deprecated NodeJS version.

* Fix: `view_from_buffer` is unsafe in Rust

Closes #453

Co-authored-by: Andrew Dirksen <2702854+bddap@users.noreply.github.com>"

* Fix: `view_from_buffer` is unsafe in Rust

Closes #453

Co-authored-by: Andrew Dirksen <2702854+bddap@users.noreply.github.com>

* Make: Upgrade SimSIMD

* Docs: Index header has no capacity

The lack of capacity data is intended.
Reserving memory is a non-persistent operation
by nature, and we shouldn't save that metadata
on the disk.

Closes #452

Co-authored-by: Christopher Yim <4638193+GoodKnight@users.noreply.github.com>

* Fix: Aggressive neighborhood checks on updates

* Fix: `update()` bug detect with non-POD keys

This bug was tough to spot. Apple Clang was the only one
that caught it. The `-O0` flags were explicitly added to expose
more symbols for debugging. More `uint40_t` tests were added.

* Fix: Align vector type w index in C# (#456)

Co-authored-by: Ash Vardanian <1983160+ashvardanian@users.noreply.github.com>

* Make: Versioning in pre-release CI

* Make: Switch from SemanticRelease to TinySemVer

---------

Co-authored-by: Jaysen Marais <jaysen.marais@gmail.com>
Co-authored-by: Antonio Andelic <antonio2368@users.noreply.github.com>
Co-authored-by: Narek Galstyan <narekg@berkeley.edu>
Co-authored-by: Adolfo Garcia <1250775+adolfogc@users.noreply.github.com>
Co-authored-by: Trevor McCulloch <mccullocht@gmail.com>
Co-authored-by: Robert Schulze <robert@clickhouse.com>
Co-authored-by: Mikhail Bautin <552936+mbautin@users.noreply.github.com>
Co-authored-by: Mikhail Bautin <mbautin@users.noreply.github.com>
Co-authored-by: SheldonFung <88470100+SheldonFung98@users.noreply.github.com>
Co-authored-by: James Braza <jamesbraza@gmail.com>
Co-authored-by: cinchen <eryue0220@gmail.com>
Co-authored-by: Mark Reed <markreed99@gmail.com>
Co-authored-by: John <jjmace01@gmail.com>
Co-authored-by: batracos <giulio.a@gmail.com>
Co-authored-by: Ziyang Hu <2103823+zh217@users.noreply.github.com>
Co-authored-by: Julius Brummack <133630819+jbrummack@users.noreply.github.com>
Co-authored-by: Julius Brummack <juliusbrummack@icloud.com>
Co-authored-by: Andrew Dirksen <2702854+bddap@users.noreply.github.com>
Co-authored-by: Christopher Yim <4638193+GoodKnight@users.noreply.github.com>
Co-authored-by: Britt <brittlewis12@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
bug Something isn't working
Projects
None yet
Development

No branches or pull requests

2 participants