make bench_uploads TELEMETRY=on (label baseline) and make bench_queue
TELEMETRY=on QUEUE_LABEL=baseline4 on b3a5ab8 for the coverage
exclusions; record last-step decisions and draft report sections.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Base accepting the saturated full-r3 run on the user's 2026-10-31 host
warning rule instead of the swap decision and ask whether to re-shoot;
complete the report section with limits, metric freshness at 30 s and
per-class and per-component tables; correct derived numbers and the
quantile error wording; align H6 in design with the report; fix the
quantile_error docstring; park S1 and S4 as later proposals.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
21/21 runs on 68e14b6. balanced passes B1-B3 and B5, so no
decision-12 fix is needed; B4 fails (replica CPU per request +27 %,
light p99 +30.7 ms) and H7 with it. H3 and H5 fail on one missing
slow-write trace of 75 under typical tail sampling. Host swap-outs
are ignored by the user's decision of 2026-11-15.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
k6 --console-output appends, so a run killed before cleanup left
<run>.warm.k6 in .work and the resumed warmup counted 24 of 12.
Remove the console file before each k6 run and a stale .work at
start; ignore evidence/**/.work/.
A telemetry storage that exits during the lookup (Tempo in
metrics-r3: name no longer resolvable) now fails the run with the
container state, OOMKilled flag and its log tail saved to raw/;
Tempo requests of the lookup are limited to 3 at a time and
memory peaks after the lookup are recorded.
Raise the light classes of the telemetry cost load by 15 % (read
126.5, write 23, me 34.5, card 34.5 per second): on a quiet host
`full` kept the replicas at 28-29.5 % CPU, below the 30 %
calibration bound.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- [telemetry] trace_sample_ratio (ParentBased(TraceIdRatioBased), 1 keeps
ParentBased(ALWAYS_ON)) and metric_buckets (default|coarse with SLO bounds),
export interval <= 30 s; otel.sdk.span.started, span.live dropped; startup
records name ratio, buckets and interval
- bench: telemetry services on mtusi-bench, TELEMETRY=on switch for uploads
and queue, make bench_telemetry[_quick] with off/full/full-aa/head/tail/
metrics/balanced, rare events, lookup from Tempo/Loki/Prometheus, cost
sampler, verdicts H1-H8 and B1-B5; collector layers and lean Prometheus in
bench/otel with preflight checks
- REPORT: telemetry cost hypotheses and criteria before the label
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Count the first SDK telemetry loss on the loss panel, print pending
moments on every alerts-check step, check the loss panel on a quiet
stand, align two scenarios and the SLI cross-check text with what the
driver verifies.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
HAProxy /metrics scraped by the Collector, recording rules and seven alerts with promtool tests and runbooks (make alerts_test, CI step), byte-sized Collector export queues with 30 s retries, SDK internal metrics, SLO/LB/pipeline dashboard rows, driver commands alerts-check and backend-outage (change slo-alerts).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Peaks under make test + load + student_test left Tempo and Grafana below
1.5x headroom; re-capture telemetry-check and collector-outage evidence on
68f1e97.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Shutdown budget leaves room for SDK shutdown, stand-from-scratch accepts only
its throwaway project, OTLP exporters ignore proxy environment, telemetry-check
covers log/trace links, health series and worker instance id; spec deltas
match what the parser and SDK lifetime allow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Collector (contrib), Prometheus, Loki, Tempo and Grafana with provisioning
from git; SDK in API and worker with level 3 metrics, traces and logs
carrying trace_id/span_id; telemetry off by default outside the stand;
broker healthcheck runs as the broker user so a fresh stand starts once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Commits of lab 5 are not in the history of this repository. The before
image of the uploads profile is now lab 5 code, so its success code is
202 instead of 201.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Source: mtusi2026lab5-safronov, solution/bvt2301-safronov, 046ec9e.
Contracts, public tests and evidence of lab 5 are not carried over.
The compose project, image and CI prefix are renamed to lab6; the first
broker start defect is recorded in KNOWN_ISSUES.md; PLAN.md outlines the
OpenSpec changes of lab 6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>