fix(streaming): worker-exact resume arithmetic and multi-worker resume test

The fast-forward skip assumed every DataLoader worker delivers batches;
workers that own no shards yield nothing and are stopped, so the batch
round-robin runs over min(num_workers, num_shards) active workers. Use
that effective count (shard-less workers skip nothing). Adds a resume
test under num_workers=2 asserting exact continuation.

Note: the test fixtures write a single parquet file regardless of
data_files_size_in_mb, so worker-splitting tests exercise the degenerate
single-shard layout; multi-shard behavior is covered by the rank-level
split_dataset_by_node tests.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
Pepijn
2026-06-11 15:11:00 +02:00
parent 1050c2fb6c
commit a7b7f4964e
3 changed files with 46 additions and 5 deletions
+1 -1
View File
@@ -74,7 +74,7 @@ frames/s/node at `--num_workers 3` (3 cameras, fps 20).
`frames_per_s_node`, `samples_per_s`, `first_batch_latency_s`, `p50/p95/p99_sample_latency_ms`,
`wallclock_s`, and `video_decoder_cache` (`hits`, `misses`, `evictions`, `hit_rate`, `size`). A low
cache `hit_rate` with high `p99` is the decoder-thrash signature — raise `--video_decoder_cache_size`
or `--buffer_size`, or reduce `num_workers`.
or `--episode_pool_size`, or reduce `num_workers`.
## Bucket sources & prewarming (manual)