Cooperative embedding#690
Draft
jshook wants to merge 9 commits into
Draft
Conversation
Contributor
|
Before you submit for review:
If you did not complete any of these, then please explain below. |
…uteLayerInfoFromSources computeLayerInfoFromSources called getNodes(0) on each source graph to count live nodes at level 0. getNodes(0) sequentially seeks through every node record on disk to filter out deleted entries. On a cold page cache this touches large amounts of source data before compaction even begins, significantly delaying the start of actual graph merging. Since every live node is present at level 0 by the HNSW invariant, the count is simply liveNodes.get(s).cardinality() — an in-memory popcount requiring no I/O. Also switch PQ retraining from ProductQuantization.compute() (full k-means++ init) to basePQ.refine() (Lloyd's iterations only, warm-started from the existing codebook). The source codebooks are already trained on the same distribution, so warm-starting converges in far fewer passes with no recall loss.
jshook
force-pushed
the
cooperative-embedding
branch
from
fa4de43 to
22d1de3
Compare
…ound record reads
Lets an embedder run graph build/cleanup on the calling thread instead of a ForkJoinPool. Existing pool constructors are preserved as delegating overloads.
…ult pools Facade carrying an embedder's compute/IO executors so build, PQ/NVQ, and compaction all run on it; also closes the PQRetrainer leak that hardcoded the default pools.
Widens the compress/train/encode entry points from ForkJoinPool to ParallelExecutor (keeping ForkJoinPool overloads), so encode/train can run on the calling thread. Encoding is byte-identical across executors; training is quality-equivalent (k-means seeds from ThreadLocalRandom).
…Runs() Now that quantization accepts ParallelExecutor, the facade carries one (plus a merge Executor and IO ExecutorService) instead of a ForkJoinPool, so a memtable flush can run build + PQ/NVQ entirely on its own thread via callerRuns(). PQRetrainer and CompactionContext.computeExecutor move to ParallelExecutor too, which also keeps retrain off the all-core pool when the merge executor isn't a ForkJoinPool.
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
2 participants
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.
This includes some changes we need to make the embedding surface more robust around shared resource usage and scoping. I've shared it here for visibility and will move it from draft status once our integrated testing bears fruit.
This is now rebased on top of Aaron's compaction improvement fixes, so testing of this branch is all-inclusive of both
Synopsis
Make the on-disk graph compactor embeddable (cooperative resource sharing). The shape of the interfaces types provided here were carefully selected to be useful and compatible with an embedding system like Cassandra or OpenSearch, without being overly specific to any. They were also chosen to be compatible with Java 11 onward.
Purpose
OnDiskGraphIndexCompactor currently runs on its own thread pool, is invisible while it works, and always writes to its own file. This PR adds a few small, optional extension points so a host system (e.g. a database's compaction pipeline) can drive the merge cooperatively — on the host's own threads, under the host's observation and throttling, writing straight into the host's own file. Everything is additive and @experimental; with nothing supplied, behavior and output are unchanged.
Key elements
Tests and docs/compaction.md cover the new surface.