Skip to content

Feat/tensor compression clean#1

Open
mpb159753 wants to merge 3 commits into
mainfrom
feat/tensor-compression-clean
Open

Feat/tensor compression clean#1
mpb159753 wants to merge 3 commits into
mainfrom
feat/tensor-compression-clean

Conversation

@mpb159753

Copy link
Copy Markdown
Owner

No description provided.

mpb159753 and others added 3 commits May 9, 2026 10:51
Adds an opt-in per-row zstd compression path for the SimpleStorage
manager <-> storage-unit data plane. The storage unit never invokes
the compression backend; compression happens once on PUT (manager
encode), decompression once on GET (manager decode). Configure via
backend.SimpleStorage.compression in config.yaml or
TQ_COMPRESSION_ALGORITHM/LEVEL/MIN_BYTES env vars.

Implementation lives entirely in compression.py + a small Ext(6)
branch in serial_utils.py + per-instance encoder/decoder threaded
through ZMQMessage.serialize/deserialize. Module-level
_encoder/_decoder are never rebound, so manager and SU colocated in
the same process never share compression state.

data_parser is incompatible with row-compressed PUT and is rejected
at put_data entry.

Signed-off-by: lynn <33576786+0lynnlin0@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Signed-off-by: mpb159753 <m.pb@msn.com>
Return raw input directly instead of creating a copy via bytes(). These functions serve as no-op backends when algorithm="none".

Signed-off-by: mpb159753 <m.pb@msn.com>
Signed-off-by: 看我72遍 <m.pb@msn.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant