ContextQMD
Back to Plate

NodeId duplicate-id paste should scan the editor once per fragment

docs/solutions/performance-issues/2026-04-03-nodeid-duplicate-paste-should-scan-editor-once-per-fragment.md

53.0.63.0 KB
Original Source

NodeId duplicate-id paste should scan the editor once per fragment

Problem

withNodeId handled duplicate pasted ids correctly, but it paid for that correctness in the dumbest possible way: one editor.api.some(...) scan per candidate id during insert_node.

On a seeded duplicate paste, that turned one fragment insert into a pile of full-editor duplicate checks.

Symptoms

  • Raw import was already cheap.
  • Duplicate-id paste was still the only meaningful withNodeId hotspot.
  • The dedicated nodeid-fragment benchmark showed the live 5k duplicate-paste lane at 20.06 ms, with 199 duplicate lookups costing about 13.89 ms.

What Didn't Work

  • Treating the existing idExistsCache as enough. It only deduplicated repeated checks for the same id. It did nothing for the common case where a pasted fragment contains many distinct ids.
  • More init-time nodeId work. The problem was not initialization anymore.

Solution

Keep the single-pass inserted-node normalization, but replace per-id editor queries with one bounded prepass:

  1. walk the inserted subtree once and collect the candidate ids that might need duplicate checks
  2. scan the editor tree once and record which of those ids already exist
  3. normalize the inserted subtree against that precomputed duplicate-id set

The implementation lives in:

The important constraint is behavioral: the rewrite keeps the same duplicate-id semantics for inserted nodes and _id overrides.

Why This Works

The old path was roughly:

  • O(candidateIds * editorSize) duplicate existence work

The new path is:

  • one inserted-subtree pass
  • one editor scan
  • one inserted-subtree normalization pass

So the repeated full-editor existence checks disappear.

On the same live 5k / 200-block duplicate-paste benchmark:

  • duplicate paste with NodeId: 20.06 ms -> 13.79 ms
  • duplicate lookup calls: 199 -> 0
  • duplicate lookup time: 13.89 ms -> 0

That cut about 6.27 ms from the real duplicate-paste lane without changing the public NodeId API.

Prevention

  • If a pasted fragment can contain many distinct ids, do not call a full-editor existence query once per id.
  • Benchmark raw import and duplicate-id paste separately. They are not the same cost shape.
  • When the benchmark says the time is in duplicate lookup, attack lookup first before touching unrelated initialization or rendering paths.