Repository navigation
Stream checked Zstandard checkpoint unit files - #22
Merged
Merged
Conversation
Signed-off-by: Robert Washbourne <rawashbourne@gmail.com>
Signed-off-by: Robert Washbourne <rawashbourne@gmail.com>
rawsh
marked this pull request as ready for review
October 23, 2026 09:32
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Checkpoint storage already uses the route-aware pickler and runs concurrently across storage units. This change streams each existing unit file through checked Zstandard, keeps legacy pickle restore, and validates complete frame length/checksum/SHA before installing restored data. It preserves async capture, abort and durable publication; it adds no within-unit sharding.
An optional, default-off
route_statsrequest on existing GET_METRICS reports physical route dtype/view bytes and unique backing bytes without restoring tensors or fetching payloads. Unknown rows are explicit.Validation: exact final source passes 53 allocated runtime291 CPU controls, including real two-unit public-KV async durable save/restore with exact logical schema, tags, route/token/mask/logprob hashes; legacy/corruption/write failure; stats without unpack/materialization and shared backing counted once. Compileall passes. The earlier combined ownership regression hit its 530-second bound before completion and is retained as a failed attempt, not a pass. The initial precommit mypy inference error was fixed with a local-variable annotation only.
A synthetic 16-row fixture gives 29,371,827 bytes packed pickle versus 11,587,639 bytes compressed; these are not production size or latency claims. Production rollout, actual save/FULLrestore, and performance qualification remain separate.
AI assistance: implemented and reviewed with Codex.