| row |
|---|
Legend: ✅ pass — conformance check passed · ❌ fail — an error was thrown · ⏱ timed out — this row exceeded its time budget · ⏭ not run — an earlier row ended this column · ⃠ not applicable — this VFS's capabilities declare this row out of scope · ⃠ blocked — the row ran, but the concurrent read was never served; the pool-acquire blocked at the scheduler while a write transaction held the VFS's single access handle — this tells you the VFS serializes the pool under a write transaction · — no result — we ran it but could not produce a trustworthy number.
Hover any of these for the reason; the downloaded JSON carries the same reasons under reasons.
read-burst-concurrency shows pool concurrency gains; a ratio is
omitted when the median serial total falls below 4× the clock’s resolution —
reads too fast to time reliably produce jitter, not a signal.
A pool-1 VFS near 1.00× is declared behaviour, not a failure.
can read during a long running query answers yes or no, not
how much: a short read is raced against a long statement holding one of
the pool's workers. ✅ means the read was served without waiting for it,
❌ means it only completed once that statement had finished. It does not
apply to pool-1 VFS (cell shows ⃠).
Reading the cells: the bar behind a measurement is its rank in that
row, not its size — best ÷ value, so a full bar is the row's
winner. Green is within 20% of the best, amber within half, red below
half. A cell in italics behind a ⚠ is past what its measurement can
physically produce — a concurrency gain above the pool that produced it —
so it is ranked at that bound instead of winning, while the figure shown
and the exported one stay what was actually read. Hover it for the detail.
A browser clock is coarsened to 0.1 ms (Chromium) or 1 ms (Safari and
Firefox), which is slower than the operations below it. So
point-read-p50 times its reads in groups of 20, and the
single INSERT pair and one page of a list in
groups of 5, all reporting the per-operation figure — the group is how the number was measured, not what
it means. The smaller group is deliberate: a group mean smears a stall
across its members, and catching stalls is what the p95 is for, so its
sensitivity is coarsened fivefold rather than twentyfold.
overwrite-throughput is read against
transaction-throughput: same count, same single transaction,
but the writes dirty existing pages scattered through the file instead of
appending at its end. The memory VFS are the control — they have no file,
so a ratio near 1 there is what says the difference elsewhere is storage
and not SQLite.