Eventual-Inc
Daft
Blog
Docs
Changelog
Blog
Docs
Changelog
Overview
Branches
Benchmarks
Runs
Performance History
Latest Results
feat(grouped-agg): shard PartitionOnly execution per morsel
BABTUNA:perf/sharded-partition-only
1 hour ago
docs(lerobot benchmark): commit the A/B driver that swaps reader revisions run_real_datasets.sh runs real_datasets.py once per reader revision (merge-base vs branch) and compares the outputs hash-for-hash
yk/lerobot-batch-decode
5 hours ago
fix(telemetry): satisfy ruff lazy imports
FANNG1:fix/telemetry-exit-segfault
6 hours ago
fix: handle supported object writers generically
aaron-ang:write-single-parquet
10 hours ago
fix(telemetry): keep SSL context setup best-effort Addresses review: building the shared SSLContext happens on the main thread (in track_import_on_scarf / track_runner_on_scarf), outside the worker's try/except. If create_default_context() fails (e.g. system CA store cannot be loaded), the error would propagate into the user's import or query execution, breaking telemetry's best-effort contract. Wrap the setup in try/except and skip the telemetry event on failure (return no thread), instead of raising. Add a test asserting that an SSL setup failure is swallowed and no request thread is started. Signed-off-by: fanng <“fanng@apache.org”>
FANNG1:fix/telemetry-exit-segfault
18 hours ago
fix(telemetry): prevent interpreter-shutdown segfault from Scarf telemetry The Scarf telemetry worker runs in a daemon thread that makes HTTPS requests via urllib. At interpreter shutdown that thread can still be inside the SSL/OpenSSL stack (creating an SSLContext and loading CA certificates) while the main thread finalizes, and the two race on OpenSSL global state, crashing the process with SIGSEGV *after* all user work has already completed successfully. It reproduces with plain repeated daft usage (e.g. a few `DataFrame.write_lance` / `read_*` calls in one process) and is easy to hit on interpreters that statically link OpenSSL, such as the python-build-standalone builds used by uv/pyenv. It surfaces as exit code 139, which CI and batch schedulers treat as a failure even though the data was written correctly. Fix: build the SSLContext once on the main thread and reuse it for the telemetry requests, so the daemon worker never initializes OpenSSL global state itself and cannot race teardown. Also pass an explicit request timeout so a slow or unreachable endpoint can't keep the worker alive inside the SSL stack. Telemetry stays non-blocking (still a daemon thread), so process shutdown is not slowed. Opting out via DAFT_ANALYTICS_ENABLED=0 / DO_NOT_TRACK=1 also avoids it. Signed-off-by: fanng <“fanng@apache.org”>
FANNG1:fix/telemetry-exit-segfault
20 hours ago
Update the tests
lhoestq:document-s3-alternative-for-cc
1 day ago
perf(video): tracking doc for read_video_frames streaming gaps
yk/video-read-frames-perf
1 day ago
Latest Branches
CodSpeed Performance Gauge
+11%
feat(grouped-agg): shard PartitionOnly execution per morsel
#7205
2 hours ago
d69c328
BABTUNA:perf/sharded-partition-only
CodSpeed Performance Gauge
+1%
feat(lerobot): batch video-frame decode by shard
#7184
6 hours ago
f652689
yk/lerobot-batch-decode
CodSpeed Performance Gauge
0%
fix(telemetry): prevent interpreter-shutdown segfault from Scarf telemetry
#7203
7 hours ago
023de44
FANNG1:fix/telemetry-exit-segfault
© 2026 CodSpeed Technology
Home
Terms
Privacy
Docs