The WP-4 runs measured the stream path only up to the bounded send channel;
the probe deliberately substituted the socket hop with a controllable reader,
which is exactly where the R2 SiteStreamEventBatcher lives — so batching had
never been in any benchmark's hot path. The harness now also attaches per-site
wire probes: a real Kestrel h2c listener hosting the production
SiteStreamGrpcServer, dialed by the production SiteStreamGrpcClient with
batching negotiated, batch sizes read from the production batch_size histogram
via a MeterListener (--wire-probes-per-site, default 10; 0 restores the old
shape). Full-scale run (10x500x75, 20-min window, results doc §8 + raw JSON):
899,807 events in 12,462 frames (mean 72.2 ev/frame, p50=p99=75), wire P99
43.4 ms vs the 250 ms budget — every batch pays the full 25 ms linger since a
75-event burst never hits the 100 cap — zero drops, zero faults, offered load
100.0% of nominal with fan-out doubled. CI smoke Fact now asserts the wire
path carries traffic and observes frames.
HarnessRunResult now carries every ResourceSample, not just the window summaries.
Motivated by finding F1: the definitive run recorded ZERO gen-2 collections across
45M events, so a positive least-squares heap slope cannot be told apart from
gen-2 garbage that was simply never collected. The summary alone cannot settle
that; the series can. The reported run predates this field -- noted as such in the
results doc rather than implied otherwise.
Also records:
- the second full-scale run was DISCARDED, not reported: a verification build
overlapped the start of its measurement window, and a contaminated measurement
is not evidence.
- finding F4, a pre-existing test-isolation flake in
QueueDepthGaugeTests.Gauge_TracksBufferedDepth_AcrossEnqueueDrainAndPark
(fails in a full-suite run, passes in isolation -- shared static gauge carrying
state across tests). It cannot originate here: this branch changes zero src/
files vs its base 986e6e7a. Left unfixed deliberately; filed for separate triage.
Verified: full slnx build clean; SiteRuntime 604/604, Communication 691/691 pass;
TargetScaleHarnessSmokeTests passes (78s).
Full-scale run executed on this machine: 10 sites x 500 instances x 75 tags =
375,000 live tag subscriptions, 37,518 updates/s achieved vs 37,500 nominal,
45,021,375 updates over a 20-minute steady-state window (M4 Pro, 14 cores, 48 GB,
with the 8-node docker rig still running so the figures are pessimistic).
11 clean passes, 1 pass with a caveat, 0 failures:
tag latency P50 0.88ms P99 4.57ms max 37.41ms (1.1M samples, end-to-end)
stream 900,675 delivered, 0 dropped at 100 live subscribers
health collect+ingest P99 0.31ms, 10/10 sites tracked
debug view P99 2.19ms, 264 completed, 0 timeouts
deploy 500 instances to a site in 2.6s
cpu 41% of ONE core = 2.9% of the box
memory working-set slope +8.83 MB/min
Three findings, reported rather than tuned away:
F1 (Low) 20 min with ZERO gen-2 collections cannot fully settle the leak
question; the heap demonstrably sawtooths but an uncompacted gen-2 makes a
positive slope ambiguous. The 1-hour run would settle it. Not tuned.
F2 (informational, by design) a deferred S&F backlog sits for one full
DefaultRetryInterval (28.9s measured) before anything drains --
EnqueueAsync(attemptImmediateDelivery:false) stamps LastAttemptAt. Easy to
misread as slow drainage, so drain is reported as two numbers: retry wait,
then 3,533 msg/s of actual capacity.
F3 (positive) slow-subscriber isolation is TOTAL: 4 healthy subscribers at
100.00% with zero drops while a peer lost 197,028/200,000 events entirely
within its own bounded channel. Mechanism recorded link by link.
Also records what the run does NOT prove (not clustered, not real-network, not
real OPC UA, not 1 hour) and the four WP-4 sub-criteria this harness does not
cover, so the evidence is not over-read.
Register rows 25 and 50 -> RESOLVED 2026-08-15; remediation execution-log
residual 7 -> resolved; phase-8-checklist WP-4 section replaced with the
measured numbers.
Joseph Doherty