Eventual-Inc/Daft - CodSpeed

Daft

Blog Docs Changelog

Performance History

Latest Results

fix(io): default https scheme in S3 endpoint normalization

plusplusjiajia:fix-iceberg-endpoint-scheme

7 hours ago

fix(flotilla): keep min_cpu_per_task default at 1.0 (no-regression) Before this PR, the flotilla scheduler fell back to a hardcoded 1.0 CPU when ResourceRequest had no num_cpus; the configured min_cpu_per_task default of 0.5 was inert. Wiring the field through (1d05985e8) without changing the default would silently halve the per-task CPU floor for every existing user, which Greptile's review flagged as a behaviour change risk on capacity-tuned clusters (e.g. KubeRay packs 2x more tasks before scale-out, increasing OOM risk). Move the default to 1.0 so the no-explicit-num_cpus path matches the pre-wiring behaviour exactly. Users who want a smaller floor set it explicitly via daft.set_execution_config(min_cpu_per_task=...) or DAFT_MIN_CPU_PER_TASK, which is the original purpose of the knob.

XiaoHongbo-Hope:fix/min-cpu-per-task-wiring

21 hours ago

Merge remote-tracking branch 'origin/main' into perf/inline-agg-any-value # Conflicts: # src/daft-recordbatch/src/ops/inline_agg.rs

BABTUNA:perf/inline-agg-any-value

1 day ago

ci: retrigger flaky HF ConnectError

BABTUNA:perf/inline-agg-bool

1 day ago

test: cover min_cpu_per_task wiring; clarify field doc - Comment said 'Floor' but the implementation is 'Default-when-None' (explicit num_cpus is honored as-is). Reword to match. - Add two unit tests: * num_cpus falls back to min_cpu_per_task when ResourceRequest is empty * explicit num_cpus passes through unchanged

XiaoHongbo-Hope:fix/min-cpu-per-task-wiring

1 day ago

fix(flotilla): wire min_cpu_per_task into TaskResourceRequest The min_cpu_per_task execution config field had no readers in the distributed scheduler: TaskResourceRequest::num_cpus() returned a hardcoded 1.0 when the plan's ResourceRequest had no num_cpus. This was wired up in #4506 for the legacy ray runner only; #5375 removed that runner along with its lone reader, and the flotilla scheduler was never wired in. Result: setting min_cpu_per_task via daft.set_execution_config or DAFT_MIN_CPU_PER_TASK had zero effect on autoscaler bundle requests. Plumb the value from DaftExecutionConfig (already in scope at SwordfishTaskBuilder::build) into TaskResourceRequest, and use it as the fallback in num_cpus() instead of the literal 1.0. Closes #7123

XiaoHongbo-Hope:fix/min-cpu-per-task-wiring

1 day ago

Merge branch 'main' into fix-actor-udf-blocking-sink-deadlock

XiaoHongbo-Hope:fix-actor-udf-blocking-sink-deadlock

1 day ago

Merge branch 'main' into fix-actor-udf-blocking-sink-deadlock

XiaoHongbo-Hope:fix-actor-udf-blocking-sink-deadlock

1 day ago

Latest Branches

+1%

fix(io): default https scheme for scheme-less S3 endpoint_url#7111

7 hours ago

228efac

plusplusjiajia:fix-iceberg-endpoint-scheme

0%

fix(flotilla): wire min_cpu_per_task into TaskResourceRequest#7125

21 hours ago

a0f07e5

XiaoHongbo-Hope:fix/min-cpu-per-task-wiring

0%

feat(inline-agg): add AnyValue accumulator#7036

1 day ago

f18a9b4

BABTUNA:perf/inline-agg-any-value

© 2026 CodSpeed Technology

Home Terms Privacy Docs