Repository navigation
fix(umbp): an in-place SSD read counts as an access - #784
Merged
TianDi101 merged 2 commits intoOct 16, 2026
Merged
Conversation
The segment-file tier ranks eviction by its own index, whose recency only reads through the tier refreshed. LocateRecord, the path GDS reads and in-place promotions take, did not touch it, so under on_read the SSD evicted in write order whatever was read. With promote_mode move a load that promoted a page's c4 objects left their segment mostly dead and still "cold", so it went next, taking the same pages' indexer objects before the load read them (an e2e run lost 1,793 of 1,796 keys in one get and sglang died). LocateRecord now touches the key like any read. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-read-touch # Conflicts: # src/umbp/include/umbp/local/tiers/sharded_ssd_tier.h # src/umbp/include/umbp/local/tiers/ssd_tier.h # src/umbp/local/tiers/sharded_ssd_tier.cpp
TianDi101
added a commit
that referenced
this pull request
Oct 16, 2026
…ect-migration Brings in #775 (segment reclaim) and #784, which this PR now merges after. A reserved record's value is written through its segment's fd outside the lock, so Reserve/Commit/AbortRecord pin that segment (and the replaced record's) until the reservation ends, as Write does; a sharded reserve refreshes its placement hints like a sharded write. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 16, 2026
feat(umbp): SSD demote/promote skip staging; promoting reads land in DRAM; 4 GiB staging preset
#776
Merged
TianDi101
added a commit
that referenced
this pull request
Oct 16, 2026
…-ssd-direct-migration # Conflicts: # src/umbp/distributed/peer/backend/ssd_backend.cpp # src/umbp/distributed/pool/peer_pool.cpp # src/umbp/include/umbp/distributed/peer/backend/page_backend.h # src/umbp/include/umbp/distributed/peer/backend/ssd_backend.h
TianDi101
added a commit
that referenced
this pull request
Oct 16, 2026
fix(umbp): an in-place SSD read counts as an access (#793) SSDTier::LocateRecord gave GDS and in-place migration readers a record's fd without touching the segment index, so segment eviction could drop a segment while a load was still reading it. It now touches the key and its segment like any other read. Re-lands #784, which merged into #775's branch after #775 was squashed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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.
Stacked on #775.
What
With #775 the segment-file SSD tier ranks eviction from its own index, but only reads through the tier (
ReadIntoPtr,ReadBatch*) refreshed that index.LocateRecorddeliberately skipped the touch. It is the zero-copy path used by GDS reads and by every in-place promotion (#776).LocateRecordnow touches the key like any other read.Why
Under
on_readthe SSD serves nothing directly: every read is an in-place promotion (10.8M in one e2e hour), so the SSD evicted in write order regardless of reads. Withpromote_mode: movethis lost data. A load promotes a page'sc4objects first, and the move deletes their SSD records. Their segment is then mostly dead and still looks cold, so(age+1)/live_fractionsends it next, taking the same pages'indexerobjects before the load reads them. In the e2e run one get found 3 of 1,796 keys and sglang's scheduler died.Test
SegmentReclaimTest.InPlaceReadCountsAsAnAccess: after in-place reads of segment 0, segment 1 is evicted first. Without the touch the test fails becausek0comes first.test_umbp_ssd_segment_reclaimpasses 12/12,test_ssd_direct_iopasses, andtest_ssd_backendpasses 32/32.🤖 Generated with Claude Code