Concurrent index: reverse-edge lists grow under update churn until consolidation

オープン
  • 何か問題がありますか？
#469 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 5 日以内に返信

まだ誰も着手していません。

評価

難易度
4/5
見積もり時間
3〜5日
初心者へのやさしさ
42/100
issue の種類
バグ
明瞭さ
おおむね明確
活発さ
活発
技術スタック
cpp
領域
databases

調査の方向性

Read include/svs/concurrent/reverse_edges.h and include/svs/concurrent/graph.h, especially replace_node_locked and the update_entries path, then inspect the C API memory accounting in bindings/c/src/svs_c.cpp. Reproduce the reported single-threaded update churn and track reverse-edge memory; done means reverse-edge storage stays bounded under repeated slot reuse without breaking concurrent writers or edge repair.

索引モデルが issue の本文から書いたものです。

説明

Summary

With SVS_SYNC_KIND_FINE_GRAIN, the concurrent index's reverse-edge index
(ReverseEdges, include/svs/concurrent/reverse_edges.h) only ever gains
entries during inserts and updates. Stale entries are pruned only by
consolidation. A caller that reuses deleted slots with
svs_index_dynamic_update_points + svs_index_dynamic_replace_id (#467, #468)
rarely needs to consolidate, so these lists keep growing. The growth shows in
svs_index_get_memory_usage and in graph_bytes of
svs_index_get_memory_breakdown, because the C API adds reverse_edges_bytes
to graph_bytes (bindings/c/src/svs_c.cpp:678).

Where it comes from

The code below is on main (added in #369) and unchanged by #467 or #468;
line numbers are at 9330e49, the head of #468.

  • replace_node_locked (include/svs/concurrent/graph.h, around line 503)
    records every new out-edge and, for every dropped out-edge, records the
    reverse pairing. It never removes an entry. The exact variant is commented
    out ("[Exact list] ... reverse_edges list would be exact for single
    thread. Need investigation for concurent case.").
  • ReverseEdges stores each list as a std::vector, so capacity is never
    returned even when consolidation removes entries.

Measurement

Single-threaded, through the C API. L2, FLOAT32, dimension 8,
graph_max_degree 64, construction window 128, 500 live vectors. Each round
deletes 100 vectors, then inserts 100 new ones into the deleted slots
(update_points on the deleted id, then replace_id to the new id). No
consolidation runs.

Inserts Memory growth since the initial fill
0 0 (the index uses 226,598 bytes)
5,000 252,580
20,000 260,004
40,000 265,124

data_bytes stays flat, so every slot is reused. All of the growth is in
graph_bytes, about 130 extra entries per node, roughly twice
graph_max_degree. After that it keeps rising slowly, by about 1.6 KB per
5,000 inserts.

Request

Could the reverse-edge index stay bounded without a full consolidation?
For example:

  • remove the reverse entry when an edge is dropped (the commented-out exact
    variant), if it can be made safe for concurrent writers; or
  • prune and shrink_to_fit a node's list when update_entries rewrites that
    node, since it already holds the node's lock and repairs its edges.

Without one of these, an index whose inserts reuse deleted slots grows to
about twice its graph size under steady churn, even though it never adds a
slot.

主要言語
C++
スター
241
フォーク
48
平均マージ
4日 21時間
マージ済み PR（30日）
11

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

intel/ScalableVectorSearch のほかの issue

intel/ScalableVectorSearch の issue をすべて見る

似ている issue

C++ の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。