Eventual-Inc/Daft - CodSpeed

Daft

Blog Docs Changelog

Performance History

Latest Results

Merge remote-tracking branch 'origin/main' into colin/plan-caching

colin/plan-caching

4 hours ago

chore: remove debug prints from operator event loops Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

colin/plan-caching

5 hours ago

add image_hash() for image deduplication

chenghuichen:image_hash

10 hours ago

Merge branch 'main' into image_hash

chenghuichen:image_hash

10 hours ago

fix: spawn finalize as background task in streaming sink and join probe Same fix as blocking_sink: complete_input was awaited inline in the event loop, blocking all other inputs during finalization. Now spawned into the shared task set with a Completed variant in the TaskResult/ TaskOutput enum. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

colin/plan-caching

12 hours ago

fix: spawn finalize as background task in blocking sink The blocking sink's event loop awaited finalize_and_send synchronously. While one input was finalizing (running final aggregation on compute pool), the entire event loop blocked — no morsels could be received or task completions processed for ANY input. With 14 concurrent inputs, this caused ~7s of cumulative stall time per plan execution. Now finalize runs in a background JoinSet, keeping the event loop free to continue processing other inputs. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

colin/plan-caching

13 hours ago

fix: disable ordering in blocking sink + add debug prints to all operators BlockingSink used OrderingAwareJoinSet with maintain_order=true, causing head-of-line blocking: fast tasks from one input waited for slow tasks from other inputs to complete first. Blocking sinks don't need ordering since they collect everything before producing output. Also adds per-event-loop debug prints to all operators showing active tasks and per-input-id state. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

colin/plan-caching

14 hours ago

fix: apply backpressure in blocking sink, streaming sink, and join nodes Replace usize::MAX with proper max_concurrency limits in next_event calls. Without backpressure, child pipelines flood sinks with morsels that pile up in pending queues, causing oversized concatenated batches and 2-3x slower aggregate tasks. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

colin/plan-caching

15 hours ago

Latest Branches

0%

feat(swordfish): plan caching with ActivePlansRegistry#6278

4 hours ago

cc196d8

colin/plan-caching

-1%

feat: add image_hash() for image deduplication#6485

10 hours ago

cd819ac

chenghuichen:image_hash

0%

feat(swordfish): introduce PipelineMessage with input_id tracking#6277

2 days ago

ac669fc

colin/pipeline-message

© 2026 CodSpeed Technology

Home Terms Privacy Docs