Skip to content

Concurrent index: reverse-edge lists grow under update churn until consolidation #469

Open

Description

@eshenayo

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.

Activity

  1. eshenayo commented on Nov 12, 2026

    @eshenayo
    Author

    More data at 50K x 128-d (graph_max_degree 64, FINE_GRAIN at 9330e49), from the benchmark described on #467:

    • With in-place updates (update_points on live ids) and slot reuse, memory_bytes grew linearly by 51.5 MB (+68%) over 25,000 updates and 25,000 inserts, with no plateau, while data_bytes stayed flat.
    • On the same pin, add/delete with periodic consolidate + compact stays bounded (62-79 MB over the same churn).
    • Right after ingest the concurrent index reports 75.2 MB against 51.9 MB for the non-concurrent index at v0.5.0, which is mostly the reverse-edge index.
    • In the profile of the update path, about 65% of update_entries CPU is GraphConsolidator, whose work set is gathered from these lists, so stale entries may also cost time.

    We are prototyping a different approach on the valkey-search side and will report whether it still shows this growth; this is pending that exploration.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions