Repository navigation
Conversation
b7r6
force-pushed
the
pr/s3-multipart-64mib
branch
2 times, most recently
from
October 7, 2026 20:40
d2a56e1 to
c5e9460
Compare
The part-size math targeted MIN_MULTIPART_SIZE (5 MiB), so a large blob was split into thousands of tiny parts — a 12 GiB blob became ~2300 concurrent 5 MiB PUTs. That churns HTTP connection pools (aws-smithy "connection never set" storms); on a good link the retries win slowly, on a lossy one they exhaust and the slow-store write fails, which write-through then cascades into dropping the successful fast-store write too. Size parts at a configurable multipart_part_size (default 64 MiB), growing only as needed to stay under the 10,000-part ceiling. 12 GiB -> ~192 parts. The same knob is wired into the ONTAP S3 store, whose previous formula divided the object size by the 5 MiB floor — a part count — and clamped that as a byte size, so every blob used 5 MiB parts. Larger parts mean more buffered bytes, so this also makes the buffering bounds real, in both stores. The existing select loop drained the bounded part channel into an unbounded FuturesUnordered, so the documented per-upload bound was never enforced; the recv arm is now gated on in-flight uploads, bounding one upload's buffer at roughly (2 * multipart_max_concurrent_uploads + 1) * part_size. Across uploads, a store-wide admission semaphore (max_concurrent_multipart_uploads, default 4, whole-upload permits so it cannot deadlock the part loop) bounds the aggregate near 5 GiB at the defaults. The single-request threshold intentionally stays at 5 MiB: the single-PUT path replays from a max_retry_buffer_per_request-sized buffer (default 5 MiB), so raising the threshold would break retries for mid-size uploads. Mid-size blobs pay two extra small requests. The existing multipart mock tests pin multipart_part_size to the old 5 MiB so their part boundaries stay small (which also exercises the knob); a new test locks the default to a single part for a blob between 5 MiB and 64 MiB. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
b7r6
force-pushed
the
pr/s3-multipart-64mib
branch
from
October 8, 2026 15:12
c5e9460 to
002ffc3
Compare
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
1 participant
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.
What and why
Lane A stability fix from the findings catalog (see
nativelink-findings-2026-10-05.csvon the
meta/findings-catalog-2026-10-05branch), rebased onto OSS main@bf4b3aa855. This fork PR exists to validate the change on CI that is
representative of upstream PR lanes (public read-only cache, same endpoints)
before any upstream submission.
How was this verified?
The branch carries its own regression tests (fails-without where the fix
class allows it — see the branch's test commits). This PR's purpose is the
full representative CI run; green here rehearses the exact checks an
upstream PR would face.
Risk
Fork-internal validation PR; nothing lands anywhere by merging this. The
underlying fix's risk assessment happens at upstream-submission time.
AI assistance
Agent-assisted pipeline (Claude) for rebase and PR mechanics under my
direction; the fix itself carries verification in its tests.