Apr 8, 2026 · Accepted

Experiment 030: Dedicated Reader Assignment

Date: 2026-04-08

Status: Accepted

Hypothesis

Each Dart pool worker is assigned a fixed C reader index at spawn time. The worker

passes this index directly to new resqlite_stmt_acquire_on() and resqlite_query_bytes()

variants that skip the C pool mutex entirely. Since the Dart pool's busy tracking

already guarantees one-worker-per-reader, the C mutex is redundant overhead.

The old path per query:

1. acquire_reader() — C mutex lock, scan for idle reader, mark in-use, unlock
2. Execute query
3. release_reader() — C mutex lock, mark available, unlock

The new path:

1. resqlite_stmt_acquire_on(db, reader_id, ...) — direct array index, no mutex
2. Execute query
3. No release needed

Investigation

Profiling against sqlite_reactive (43.5K point qps vs resqlite's 30.5K) revealed that

resqlite's per-query overhead included two C mutex round-trips plus native memory

allocation for parameter serialization. sqlite_reactive's workers each own their

connection directly with no pool coordination.

Changes

• Added resqlite_stmt_acquire_on() in C — takes a fixed reader_id, skips pool mutex
• Changed resqlite_query_bytes() to take reader_id instead of acquiring from pool
• Updated Dart worker entrypoint to receive reader_id at spawn time
• Updated _executeQueryImpl to use _resqliteStmtAcquireOn
• Each _WorkerSlot stores its reader index and passes it to the worker

Results

3-run comparison (head-to-head verifier benchmark):

Initial single-run comparison showed apparent CRUD and read-under-write regressions,

but 3-run verification proved these were JIT/thermal warm-up artifacts — both metrics

matched exactly by the third run on both branches.

Bug Found During Merge

Main branch experiments 025-029 changed bind_params from SQLITE_TRANSIENT to

SQLITE_STATIC. The initial dedicated reader code freed params immediately after

resqlite_stmt_acquire_on, which worked with TRANSIENT (SQLite copies during bind)

but caused dangling pointer reads with STATIC (SQLite holds pointer until step

completes). Fix: free params in the outer finally block after stepping, matching

the main branch pattern.

Decision

Accepted — eliminates two C mutex round-trips per query. Point query throughput

improved ~40% (30.5K → ~50K qps), closing the gap with sqlite_reactive. No impact

on writes, large reads, or streaming. Zero code complexity increase — just a new

C function and passing an index instead of acquiring from a pool.