Apr 20, 2026 · Rejected (optimization hypothesis disproven; benchmark harness kept)

Experiment 082: Isolate message-graph benchmark for result payloads

Date: 2026-04-20

Status: Rejected (optimization hypothesis disproven; benchmark harness kept)

Problem

After several result-storage experiments, it was still unclear whether the

remaining headroom lived in:

• SQLite decode
• isolate hand-off
• main-isolate row consumption

Those phases were too entangled in normal db.select() benchmarking to

attribute wins cleanly. We needed a benchmark that isolated the message

graph itself.

Hypothesis

If the current ResultSet / Row shape is not near-optimal, then an

isolated hand-off benchmark should show a clearly better alternative graph

shape for one or both of:

• SendPort.send (same-group copy path)
• Isolate.exit (handoff path)

Candidate graphs:

1. Current ResultSet
2. Fully materialized row maps
3. Binary row facade

Measured separately:

• transfer time
• main-isolate consumption time
• total time

Approach

Added reusable benchmark harness:

• benchmark/experiments/message_graph.dart

The harness:

• builds synthetic result-shaped payloads without SQLite
• tests both numeric-heavy and mixed-schema data
• tests 100 / 1000 / 10000 rows
• runs both SendPort.send and Isolate.exit
• reports p50 / p90 for transfer, consume, and total
• runs in both JIT and AOT

This gives a direct answer to “is the result object shape itself still a

frontier?” without conflating decode cost.

Results

Artifacts:

• JIT: /tmp/resqlite-exp097-jit.txt
• AOT: /tmp/resqlite-exp097-aot.txt

The AOT run is the most important signal because it matches shipped

Flutter runtime behavior.

AOT examples

Numeric-heavy, SendPort.send, 1000 rows

Mixed-schema, SendPort.send, 10000 rows

Mixed-schema, Isolate.exit, 10000 rows

Decision

Rejected as an optimization direction, but the benchmark harness should be

kept.

The main result is that the current ResultSet / Row shape is already

very strong for the shipped select() contract:

• lazy rows on the receiving isolate
• one shared flat backing list
• Map-compatible API

Both alternatives we tested lost on end-to-end total time once actual row

consumption was included:

• materialized maps lose on transfer/object count
• binary row facades lose on main-isolate access cost

This is a useful negative result. It means future performance work should

not keep trying to replace generic select() result objects unless the

API contract changes.

More promising directions from here are:

• specialized byte/typed-data APIs
• threshold / hand-off policy work
• optimizing stream/reactive work rather than generic row payloads

The harness remains valuable because it gives a clean way to test any

future result-graph idea without conflating SQLite decode, isolate

transport, and row consumption.