pg.ddx.io  pgvector/pgvector GitHub issues and pull requests  
help / color / mirror / Atom feed
From: JoeyLYZ666 (@JoeyLYZ666) <noreply+JoeyLYZ666@github.com>
To: pgvector/pgvector <noreply+pgvector-pgvector@github.com>
Subject: [pgvector/pgvector] issue #1052: [Bug] Sparse scalar distances overflow on finite `sparse
Date: Sat, 31 Oct 2026 08:56:55 +0000
Message-ID: <gh-pgvector-pgvector-1052@github.com> (raw)

## Describe the bug

pgvector 0.8.7 returns `Infinity` or `NaN` from sparsevec scalar distance
functions for finite nonzero elements whose exact result is representable by
`double precision`.

## Reproduce

Run `testcase.py` against PostgreSQL with pgvector 0.8.7:

```powershell
python testcase.py
```

The testcase compares two finite `sparsevec(2)` values with opposite `1e38`
components and computes L1, L2, and self cosine distance.

[testcase.py](https://github.com/user-attachments/files/34213348/testcase.py)

## Expected behavior

The expected results are finite: approximately `4e38`, `2.828427e38`, and `0`
for self cosine distance.

## Actual behavior

The affected release returns:

```text
Infinity, Infinity, NaN
```

The sparsevec implementation accumulates intermediate values in `float` before
returning `double precision`.

## Impact

Finite sparse vectors can produce non-finite or invalid distances, breaking
ordering, thresholds, and scalar correctness checks.

## Environment

- pgvector 0.8.7
- PostgreSQL 18
- `sparsevec(2)` values

`vector.c` report because the SQL type and implementation path differ.


view thread (3+ messages)  latest in thread

Message-ID: <gh-pgvector-pgvector-1052@github.com>
Permalink:  ../gh-pgvector-pgvector-1052@github.com/

 · 


This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox