docs(notes): mark the sync-latency waterfall as partly superseded

The 2026-07-11 dev-repo record stands, but the publish half it
describes (add_all index staging, mixed-reset reconcile) was replaced
by the splice + soft-reset replay, and the reconcile path is now
hardware-verified. Point to the kaizen and the tradeoff curve for the
real-repo numbers (24.1 s cold :gp).
This commit is contained in:
Julien Calixte
2026-07-14 12:49:17 +02:00
parent abbd9d2e6e
commit 4ead64bada

View File

@@ -1,11 +1,23 @@
# Sync latency — where the ~16 s cold `:sync` goes
# Sync latency — where the ~16 s cold `:gp` goes
> **Measured 2026-07-11** on hardware, via the `:sync timing —` log line in
> [`firmware::git_sync`](../../firmware/src/git_sync.rs) (`publish_cycle`). A
> **cold** `:sync` (first of a power cycle) is **~16.0 s** power-on of Wi-Fi →
> `push done`; a **warm** one skips the one-time setup and is just the ~10 s
> publish. This note breaks the number down and records why most of it is a
> floor, not a bug.
> **Measured 2026-07-11** on hardware, via the timing log line in
> [`firmware::git_sync`](../../firmware/src/git_sync.rs) (`publish_cycle`;
> the command was `:sync` then, renamed `:gp` 2026-07-14). A **cold** publish
> (first of a power cycle) is **~16.0 s** power-on of Wi-Fi → `push done`; a
> **warm** one skips the one-time setup and is just the ~10 s publish. This
> note breaks the number down and records why most of it is a floor, not a
> bug.
>
> **Update (2026-07-13/14) — the publish half below is superseded.** The
> `add_all` index staging was replaced by the O(depth) splice over the
> journaled dirty set ([kaizen](../kaizen/real-repo-sync.md) · [measurement
> trail](../tradeoff-curves/sync-commit-staging.md)), reconcile became fetch +
> **soft** reset + journal replay, and TLS session resumption reuses the
> handshake on reconnects. On the **real notes repo** — which the method below
> could never complete at all — a cold `:gp` is **24.1 s**, a warm one ≈ 19 s
> (splice depth × loose-write cost dominates), and an up-to-date `:gl` is
> ≈ 4.7 s git-side. The waterfall below stands as the dev-repo record of
> 2026-07-11.
>
> Notes index: [`README.md`](README.md). Docs index:
> [`../README.md`](../README.md). Why the raw number matters less than it looks:
@@ -60,9 +72,13 @@ and retries. For this **single-writer appliance** that resolves last-writer-wins
a concurrent remote _edit_ to the same note loses to ours, and a remote-only
_added_ file the card doesn't have would be dropped by the replay's `add --all`.
Both need a real merge (increment B) and don't arise from the device's own use.
This path is **hardware-verified for the happy case** (`9b635c42` fast-forwarded
clean); the reconcile branch itself is compile-verified but not yet exercised on
device.
**Update 2026-07-14:** the full rejected-push → reconcile → replay → push
cycle is now hardware-verified (24.0 s end-to-end with TLS session
resumption). The mechanics also changed with the splice: the reset is
**soft** (there is no index anymore) and the replay splices only the
journaled dirty paths, so a remote-only added file now _survives_ — it is
carried forward by OID instead of being dropped by an `add --all`.
## Can cold sync go lower?
@@ -75,16 +91,18 @@ The big rocks are physics or protocol, not slack:
- **TLS handshake ~2.4 s** and **push negotiate/upload ~4.4 s** are inherent to
libgit2-over-mbedTLS on this part; the payload is tiny, so there's little to
shave.
- **stage + commit ~3.1 s** is the one soft spot: staging over the editor's dirty
set (`add_path`) instead of `add_all(["*"])` would skip the SD/FAT tree walk
(likely → sub-second) _without_ losing multi-file — the dirty set is the file
list. Whether the walk actually dominates the ~4 s commit is now being measured
by the `commit split —` log line; the cost model and the rule it decides live in
[`../tradeoff-curves/sync-commit-staging.md`](../tradeoff-curves/sync-commit-staging.md).
- **stage + commit ~3.1 s** was the one soft spot — and the one that got
attacked. **Resolved 2026-07-13:** the index path was replaced outright by
the O(depth) TreeBuilder splice over the journaled dirty set (the `add_path`
staging this note originally proposed still hits the index's racy-clean
wall). The cost
model and how each hypothesis died live in
[`../tradeoff-curves/sync-commit-staging.md`](../tradeoff-curves/sync-commit-staging.md);
the residual is FAT directory-op cost per loose write, bounded and accepted.
**Conclusion:** ~16 s cold / ~10 s warm is close to the floor for "commit to FAT +
one TLS push over Wi-Fi with a fresh clock." It reads as slow only if you wait on
it — and by design you don't: `:sync` is a deliberate action with a snackbar, and
it — and by design you don't: `:gp` is a deliberate action with a snackbar, and
[`ctrl-g-perceived-latency.md`](ctrl-g-perceived-latency.md) argues the perceived
cost is set by _when durability is surfaced_, not by wall-clock. Recorded here so
the number is scoped against the protocol, not treated as a regression.