Spindle vs the field

github-actions/ubuntu-24.04 (shared runner) · AMD EPYC 7763 64-Core Processor, 4 cores, kernel 6.17.0-1022-azure · 3 rounds, 2026-09-06 · spindle @ 35988f5c5705

measured: spindle 0.1.0 @ 35988f5c5705, synapse 1.160.0 (b=main,35988f5) on postgres, continuwuity 26.8.1 (ab3c05d), tuwunel 1.9.0, dendrite 0.15.2+e546df2

Measured over 3 rounds per server. The large figure in a cell is the median over rounds against the median over rounds; the small range under it is the band the rounds allow — their fastest round against our slowest, to their slowest against our fastest. A cell is called only when that band excludes 1.0×, which is the same as saying the two servers' rounds separate: win our slowest round beat their fastest · overlapping the ranges cross, so the difference is not resolved by this many rounds, whatever the medians say · loss our fastest round lost to their slowest — every such cell links to its investigation, because the roadmap treats it as a defect until explained. Hover any cell for both servers' medians and observed ranges in milliseconds.

That rule bounds the false-call rate for one cell, and this table has 105. Two identical servers separate by luck 10.0% of the time at 3 rounds, so about 10.5 of these cells should be called by chance alone — against 100 actually called. The arithmetic cannot say which ones, so a call marked † is one that stands alone: the same operation is not called the same way at any other size. A cost that is real in a per-item measure normally shows across the size axis, so an isolated call is the shape a chance separation takes. Read those as unconfirmed rather than as results (#183).

Measured over 3 rounds per server. The large figure in a cell is the median over rounds against the median over rounds; the small range under it is the band the rounds allow — their fastest round against our slowest, to their slowest against our fastest. A cell is called only when that band excludes 1.0×, which is the same as saying the two servers' rounds separate: win our slowest round beat their fastest · overlapping the ranges cross, so the difference is not resolved by this many rounds, whatever the medians say · loss our fastest round lost to their slowest — every such cell links to its investigation, because the roadmap treats it as a defect until explained. Hover any cell for both servers' medians and observed ranges in milliseconds.

That rule bounds the false-call rate for one cell, and this table has 81. Two identical servers separate by luck 10.0% of the time at 3 rounds, so about 8.1 of these cells should be called by chance alone — against 75 actually called. The arithmetic cannot say which ones, so a call marked † is one that stands alone: the same operation is not called the same way at any other size. A cost that is real in a per-item measure normally shows across the size axis, so an isolated call is the shape a chance separation takes. Read those as unconfirmed rather than as results (#183).

1 round(s) per server — not resolved. Three rounds a side is the minimum that means anything: if two servers were identical, all of one side's rounds landing below all of the other's happens by chance 2/C(2n, n) of the time — one in three at two rounds, one in ten at three. With no spread to read, the cells below are coloured by the repeatability this host was measured to have: six rounds of the same binary moved the median cell 1.38×, and 21 of 21 cells varied by more than the ±10% band this page used to use. So: ≥1.38× · 0.72–1.38× · ≤0.72×. A grey cell is not a tie — it is a difference this sitting cannot see, and only more rounds can (#171). The large ratios are unaffected. Hover any cell for the raw milliseconds.

1 round(s) per server — not resolved. Three rounds a side is the minimum that means anything: if two servers were identical, all of one side's rounds landing below all of the other's happens by chance 2/C(2n, n) of the time — one in three at two rounds, one in ten at three. With no spread to read, the cells below are coloured by the repeatability this host was measured to have: six rounds of the same binary moved the median cell 1.38×, and 21 of 21 cells varied by more than the ±10% band this page used to use. So: ≥1.38× · 0.72–1.38× · ≤0.72×. A grey cell is not a tie — it is a difference this sitting cannot see, and only more rounds can (#171). The large ratios are unaffected. Hover any cell for the raw milliseconds.

Nothing in this sitting reads as a loss under the band above. The investigations below were written when it did, and are kept because they are the evidence for the conclusion, not a footnote to it.

Investigated — Sliding sync (MSC4186): 0.77× vs Continuwuity at 3,200 events, and 0.85× in the same day's discarded loaded run — repeatable by the two-sitting rule, so it got the full second look. Both servers were probed live minutes after the sitting, same client, same instant, two shapes: like-for-like the gap does not exist (creator shape 0.844 ms vs 0.854 ms, parity; the driver's exact observer shape 0.785 ms vs 0.874 ms, Spindle 1.11× faster). The two sitting legs caught opposite sides of the machine's same-day swing. No fix ships because no defect was found; the cell keeps its measured value and links here — the honest kind of red.

Investigated — Initial /sync: 0.89× vs Tuwunel at 200 events in this sitting — and 1.21× in the same day's discarded run, with 800 and 3,200 in noise both times. A cell that flips sign between sittings hours apart is run-to-run variance, published as measured.

1 round(s) per server — not resolved. Three rounds a side is the minimum that means anything: if two servers were identical, all of one side's rounds landing below all of the other's happens by chance 2/C(2n, n) of the time — one in three at two rounds, one in ten at three. With no spread to read, the cells below are coloured by the repeatability this host was measured to have: six rounds of the same binary moved the median cell 1.38×, and 21 of 21 cells varied by more than the ±10% band this page used to use. So: ≥1.38× · 0.72–1.38× · ≤0.72×. A grey cell is not a tie — it is a difference this sitting cannot see, and only more rounds can (#171). The large ratios are unaffected. Hover any cell for the raw milliseconds.

Nothing in this sitting reads as a loss under the band above. The investigations below were written when it did, and are kept because they are the evidence for the conclusion, not a footnote to it.

Investigated — Sliding sync (MSC4186): 0.90× and 0.88× vs Continuwuity at 800 and 3,200 events — and the same cell sat at 0.83× in the M2 close-out, so two sittings agreed this was repeatable, not noise. A component probe on the live bench server found the room list's recency sort reading each room's head event body from the store and parsing its JSON on every request, for one i64. #126 moved the sort key into memory, refreshed by the append that changes it; re-measured on the same idle machine the cells recover to 0.96× and 1.00×, and Spindle's own growth curve flattens from 1.28× to 1.13× across a 16× room-size increase.

1 round(s) per server — not resolved. Three rounds a side is the minimum that means anything: if two servers were identical, all of one side's rounds landing below all of the other's happens by chance 2/C(2n, n) of the time — one in three at two rounds, one in ten at three. With no spread to read, the cells below are coloured by the repeatability this host was measured to have: six rounds of the same binary moved the median cell 1.38×, and 21 of 21 cells varied by more than the ±10% band this page used to use. So: ≥1.38× · 0.72–1.38× · ≤0.72×. A grey cell is not a tie — it is a difference this sitting cannot see, and only more rounds can (#171). The large ratios are unaffected. Hover any cell for the raw milliseconds.

Nothing in this sitting reads as a loss under the band above. The investigations below were written when it did, and are kept because they are the evidence for the conclusion, not a footnote to it.

Investigated — Sliding sync (MSC4186): 0.87× vs Continuwuity at 3,200 events was the one real loss of the M2 close-out, and the only curve growing with room size. Bisecting a live server pinned it in one probe: the unread counter was reading every event body after the receipt floor. #113 replaced that walk with two binary searches over a per-room sender index — 11.79 ms → 1.00 ms on the pathological case — and the M3 rows below show the cell recovered.

Investigated — Read room state: 0.85× vs Tuwunel at 200 events, and 0.93× again at M3 progress — the one cell Tuwunel held across two sittings, so the investigation started in their tree. Their /state serves each event through RocksDB's block cache; ours paid a room-lock acquisition, a body read and a JSON parse per state event, per request. Component probes showed their state machinery was never actually faster — their per-request pipeline is just leaner. #129 caches the rendered /state body under its BLAKE3 state root (content-addressed, so a hit is provably current and a root mismatch is the only invalidation); re-measured against the live Tuwunel binary the cell flips to 1.91× and 1.33× in Spindle's favour.

1 round(s) per server — not resolved. Three rounds a side is the minimum that means anything: if two servers were identical, all of one side's rounds landing below all of the other's happens by chance 2/C(2n, n) of the time — one in three at two rounds, one in ten at three. With no spread to read, the cells below are coloured by the repeatability this host was measured to have: six rounds of the same binary moved the median cell 1.38×, and 21 of 21 cells varied by more than the ±10% band this page used to use. So: ≥1.38× · 0.72–1.38× · ≤0.72×. A grey cell is not a tie — it is a difference this sitting cannot see, and only more rounds can (#171). The large ratios are unaffected. Hover any cell for the raw milliseconds.

1 round(s) per server — not resolved. Three rounds a side is the minimum that means anything: if two servers were identical, all of one side's rounds landing below all of the other's happens by chance 2/C(2n, n) of the time — one in three at two rounds, one in ten at three. With no spread to read, the cells below are coloured by the repeatability this host was measured to have: six rounds of the same binary moved the median cell 1.38×, and 21 of 21 cells varied by more than the ±10% band this page used to use. So: ≥1.38× · 0.72–1.38× · ≤0.72×. A grey cell is not a tie — it is a difference this sitting cannot see, and only more rounds can (#171). The large ratios are unaffected. Hover any cell for the raw milliseconds.

1 round(s) per server — not resolved. Three rounds a side is the minimum that means anything: if two servers were identical, all of one side's rounds landing below all of the other's happens by chance 2/C(2n, n) of the time — one in three at two rounds, one in ten at three. With no spread to read, the cells below are coloured by the repeatability this host was measured to have: six rounds of the same binary moved the median cell 1.38×, and 21 of 21 cells varied by more than the ±10% band this page used to use. So: ≥1.38× · 0.72–1.38× · ≤0.72×. A grey cell is not a tie — it is a difference this sitting cannot see, and only more rounds can (#171). The large ratios are unaffected. Hover any cell for the raw milliseconds.

2 round(s) per server — not resolved. Three rounds a side is the minimum that means anything: if two servers were identical, all of one side's rounds landing below all of the other's happens by chance 2/C(2n, n) of the time — one in three at two rounds, one in ten at three. With no spread to read, the cells below are coloured by the repeatability this host was measured to have: six rounds of the same binary moved the median cell 1.38×, and 21 of 21 cells varied by more than the ±10% band this page used to use. So: ≥1.38× · 0.72–1.38× · ≤0.72×. A grey cell is not a tie — it is a difference this sitting cannot see, and only more rounds can (#171). The large ratios are unaffected. Hover any cell for the raw milliseconds.

1 round(s) per server — not resolved. Three rounds a side is the minimum that means anything: if two servers were identical, all of one side's rounds landing below all of the other's happens by chance 2/C(2n, n) of the time — one in three at two rounds, one in ten at three. With no spread to read, the cells below are coloured by the repeatability this host was measured to have: six rounds of the same binary moved the median cell 1.38×, and 21 of 21 cells varied by more than the ±10% band this page used to use. So: ≥1.38× · 0.72–1.38× · ≤0.72×. A grey cell is not a tie — it is a difference this sitting cannot see, and only more rounds can (#171). The large ratios are unaffected. Hover any cell for the raw milliseconds.