ContextQMD
Back to Plate

NodeId paste/import work needs a dedicated `insertFragment` benchmark

docs/solutions/performance-issues/2026-04-03-nodeid-paste-import-needs-insertfragment-benchmark.md

53.0.63.5 KB
Original Source

NodeId paste/import work needs a dedicated insertFragment benchmark

Problem

NodeIdPlugin already had a clean init-time story, but the expensive real-world path for copy/paste and import lives inside withNodeId during fragment insertion. The existing init-dissection lane did not touch that path.

That meant we could keep shaving the wrong seam and still have no honest answer about whether withNodeId deserved more surgery.

Symptoms

  • init-dissection only timed construction, initialization, and pure normalizeNodeId(...).
  • The optimized withNodeId insert path still had no dedicated benchmark lane.
  • Any argument about paste/import cost was half evidence and half vibes.

What Didn't Work

  • Treating init-time nodeId numbers as a proxy for paste/import cost. They are not the same path.
  • Guessing from unit tests alone. Tests can prove correctness, not the shape of the runtime bill.
  • Doing more blind withNodeId rewrites before measuring duplicate-id paste directly.

Solution

Add a dedicated nodeid-fragment benchmark lane to /dev/editor-perf.

The new lane times real editor.tf.insertFragment(...) work for four cases:

  • NodeId off, raw import
  • NodeId on, raw import
  • NodeId off, duplicate-id paste
  • NodeId on, duplicate-id paste

It also records the counters that actually explain the cost:

  • ids assigned during insertion
  • duplicate lookup count
  • duplicate lookup time
  • insert_node operation count

The fragment builder intentionally separates two shapes:

  • raw import data with no ids
  • seeded duplicate paste data whose ids already exist in the destination

The focused helper/spec lives in:

Why This Works

It measures the real seam instead of a neighboring seam.

The first live 5k run on http://localhost:3020/dev/editor-perf with a 200-block fragment showed:

  • raw import baseline, NodeId off: 5.32 ms
  • raw import, NodeId on: 5.87 ms
  • duplicate paste baseline, NodeId off: 5.54 ms
  • duplicate paste, NodeId on: 20.06 ms

That means:

  • raw import is basically cheap now; enabling NodeId only adds about 0.55 ms for 199 assigned ids
  • the real remaining bill is duplicate-id paste, not raw import
  • in the duplicate paste case, 199 duplicate lookups cost about 13.89 ms, which explains almost all of the extra runtime

So the benchmark changed the conclusion:

  • do not keep optimizing init-time NodeId because paste/import feels scary
  • only do more withNodeId work if you are targeting duplicate lookup cost

Prevention

  • Do not use init-only benchmarks to justify paste/import rewrites.
  • When a plugin has separate init and live-insert paths, benchmark both.
  • For NodeId specifically, keep two fragment shapes in the benchmark:
    • raw import
    • duplicate-id paste
  • If a future optimization claim does not move the duplicate lookup lane, it is probably not moving the real bottleneck.