vortex-data/vortex - CodSpeed

vortex

Blog Docs Changelog

Performance History

Latest Results

break[array]: mask compute to lazy mask execution (#6485)

develop

9 hours ago

fix[layout]: snapshot nbytes at push time in ChunksBuffer to avoid underflow (#6520) ## Does this PR closes an open issue or discussion? - N/A ## What changes are included in this PR? - Changed ChunksBuffer.data from VecDeque<ArrayRef> to VecDeque<(ArrayRef, u64)>, snapshotting each chunk's nbytes() at push time. - pop_front now subtracts the stored snapshot instead of re-querying chunk.nbytes(). - Added a regression test that reproduces the underflow via SharedArray state transition. ## What is the rationale for this change? SharedArray slices sharing an Arc<Mutex<SharedState>> can transition from Source to Cached when any one of them is canonicalized. This is interior mutation — it happens without the ChunksBuffer knowing. Because nbytes() traverses the current array tree, a chunk that reported a small nbytes() at push time (compressed/source-era) can report a much larger value at pop time (decompressed/cached-era). The running nbytes total was accumulated with the smaller values, so subtracting the larger value causes an integer underflow panic (attempt to subtract with overflow). By snapshotting nbytes() once at push time and storing it alongside the chunk, the accounting stays self-consistent regardless of any interior state changes between push and pop. ## How is this change tested? A new unit test chunks_buffer_pop_front_no_panic_after_shared_execution reproduces the exact failure scenario: 1. Creates a ConstantArray (tiny nbytes) wrapped in a SharedArray. 2. Slices it into two chunks and pushes both into a ChunksBuffer. 3. Pops the first chunk (succeeds, no state change yet). 4. Triggers canonicalization via shared_handle.get_or_compute(), which transitions all slices to Cached state with much larger nbytes. 5. Pops the second chunk — before this fix, this panicked with underflow. ## Are there any user-facing changes? No. This is an internal bug fix in the repartition writer. No public APIs are changed. --------- Signed-off-by: Brian Hart <brian@withforge.com>

develop

11 hours ago

CR Signed-off-by: Brian Hart <brian@withforge.com>

brainhart:2026-02-15-vortex-repartition-bytes-underflow

11 hours ago

bracket CUB filters Signed-off-by: Andrew Duffy <andrew@a10y.dev>

aduffy/gpu-scan-measure

12 hours ago

hardest Signed-off-by: Andrew Duffy <andrew@a10y.dev>

aduffy/gpu-scan-measure

14 hours ago

address harder Signed-off-by: Andrew Duffy <andrew@a10y.dev>

aduffy/gpu-scan-measure

14 hours ago

fix Signed-off-by: Joe Isaacs <joe.isaacs@live.co.uk>

ji/mask-compute-fn-remove

15 hours ago

fix Signed-off-by: Joe Isaacs <joe.isaacs@live.co.uk>

ji/mask-compute-fn-remove

15 hours ago

Active Branches

CLI for measuring execute_cuda encoding perf

last run

12 hours ago

0%

pvalue::cast returns VortexResult

last run

15 hours ago

0%

Move `vortex-scalar` into `vortex-array`

last run

16 hours ago

+16%

© 2026 CodSpeed Technology

Home Terms Privacy Docs