codeflash-agent/.codeflash/netflix/metaflow/status.md

metaflow Status

Last updated: 2026-04-10

Current state

First PR open upstream. Waiting for maintainer feedback.

Target repo

~/Desktop/work/netflix_org/metaflow — fork remote: KRRT7/metaflow

VM

Azure Standard_D2s_v5, IP: 20.112.32.177, RG: metaflow-BENCH-RG (deallocated)

• Python 3.12, pip editable install, lz4/xxhash/numpy/hyperfine installed
• Baseline + realistic benchmarks complete in ~/results/

PRs

PR	Branch	Status	Description
Netflix/metaflow#3090	perf/lz4-artifact-compression	Open, waiting for review	Replace gzip with lz4 in CAS — 7-18x on realistic data
KRRT7/metaflow#1	perf/lz4-artifact-compression	Draft (mirror)	Same, on fork

Key results (realistic artifacts)

Payload	Pickled Size	gzip total	lz4 total	Speedup
Small dict (config)	233B	0.341ms	0.218ms	1.6x
Metrics dict (feature stats)	52KB	2.278ms	0.327ms	7.0x
Numpy float64 (embeddings)	800KB	29.111ms	1.557ms	18.7x
Numpy float64 (model weights)	8MB	289.234ms	15.792ms	18.3x
Random bytes (opaque model)	5MB	118.315ms	9.646ms	12.3x

Open questions on PR

• Hard vs soft dependency for lz4
• Forward compat story (old metaflow can't read cas_version=2)
• Benchmark scripts to be reverted before merge

Next steps (pending maintainer response)

1. If approach accepted: make lz4 optional, revert benchmark scripts, address feedback
2. If rejected on dependency grounds: explore zlib.compress directly (no new dep, smaller win)
3. Open SHA1 discussion issue (data in data/sha1-proposal.md)
4. Multicore polling improvement (low priority, marginal impact)

Blockers

Waiting on Netflix/metaflow#3090 review.