Origin-shield saturation
on a correlated cache-miss storm.
A working postmortem on the na-east weekday primetime broadcast where a correlated upstream cache-miss storm fanned into the CDN origin-shield tier past its pre-provisioned ceiling — origin.replay_origin_origin_fetch_qps climbed to 4.2× within a 90-second window, cdn_shield.origin_segment_cache_miss_fanin lifted from 0.11 to 0.62 as the warm edge POPs' cache layer evicted correlated segments, and cdn_shield.queue_depth_concurrent pegged to 1.00 of the slot ceiling — while cohort.reconnect_storm_ratio read at 0.04 baseline (the cohort is calm — the driver is upstream) and edge.egress_kbps sat at 8214 Kbps, FLAT at the expected rate. The Streamwake agentic ops layer classified it as origin_shield_saturation_under_correlated_cache_miss_storm at 84% confidence with cdn_edge_pop_warmup_under_eviction ruled out by name on warm edge POP segment-leg cache-hit, and remediated with an authorization-tiered policy: Tier 0 autonomous under gate (who can shed traffic), Tier 1 surfaced to humans (who can fail open), Tier 2 surfaced for operator-team approval forward (who can escalate) — recovery verified cohort-side + shield-side on the four staged gates T+30 s → T+15 m, NOT infrastructure-green alone.
Book a technical demo for Origin shield saturation · correlated cache-miss storm
Read the postmortem — then bring your own incident to Streamwake.
Two ways to engage on this exact failure pattern: book a 30-minute technical demo where we walk through the probe cascade on your source, or hand us an archived incident and watch the agent diagnose it end-to-end.
Both routes land on the scoping intake form — no SDR gate.
What the cohort saw
The first things to read on any real correlated upstream cache-miss storm incident are the cohort-level numbers on the affected na-east tier — and crucially, the contrast against a cohort that stays calm: cohort.reconnect_storm_ratio at 0.04 baseline (the cohort is not driving the fan-in) and edge.egress_kbps flat at the expected rate (the cohort IS delivering traffic — the upstream origin-shield is the pinch, NOT the cohort's egress shape). Three numbers did the heavy lifting here: the replay-origin cache-miss fan-in, the shield queue depth, and the warm edge POPs' cache layer eviction posture paired with the cohort's connected-state (calm) signal. The figures below are simulated telemetry — the disclosure above applies to every figure on this page.
Roughly 12% of the na-east weekday primetime cohort — session-level late-join stalled behind the shield queue fan-in on the affected cdn-A/atl01 tier, where origin.replay_origin_origin_fetch_qps climbed to 4.2× the tier's pre-provisioned ceiling and cdn_shield.queue_depth_concurrent pegged at 1.00 of the slot ceiling. The cohort itself was calm — cohort.reconnect_storm_ratio read at 0.04 baseline (NOT 1.00 like the marquee live-event viewer-storm page) and edge.egress_kbps sat at 8214 Kbps, FLAT at the expected rate. The reader couldn't get through the shield's backpressure curve on first join, dragging time-to-first-frame and pushing cohort_join_stall_ratio to 0.094 over a 90-second window before the Tier 0 shed landed.
~36 minutes between the first replay-origin cache-miss fan-in at T+0m and the cohort-side close-out probes settling within tolerance at T+22m. The postmortem window goes longer than the natural cohort-side close-out settle because the Tier 1 fail-open posture on the shield's circuit-breaker and the Tier 2 operator-team codification of the new origin-failover policy ship after the close-out probes land — so the postmortem window is two-tier: T+0 m → T+22 m on the cohort-side + shield- side close-out, and shaped forward by the Tier 1 fail-open approval + Tier 2 codification that bound forward to the next weekday primetime broadcast.
origin.replay_origin_origin_fetch_qps: 4.2× on shield-NA-3 (na-east affected tier · warm edge POP cohort) — the upstream replay-origin fanned into the shield tier past its pre-provisioned ceiling, with cdn_shield.queue_depth_concurrent → 1.00 and cdn_shield.origin_segment_cache_miss_fanin → 0.62 (vs 0.11 pre-event baseline). The discriminator is that the cohort is calm — cohort.reconnect_storm_ratio: 0.04 (BASELINE), edge.egress_kbps: 8214 (flat). The symbol of failure was the upstream origin-shield queue + cache-miss fan-in, NOT the cohort's egress, NOT a viewer-driven reconnect storm.
How Streamwake classified this incident
Three ranked hypotheses: the top one filing the timeline as origin_shield_saturation_under_correlated_cache_miss_storm · dominant, the second explicitly tagged cdn_edge_pop_warmup_under_eviction · ruled out by so the warm-edge-POP cache layer lane reads correctly (warm edge POPs' segment-leg cache-hit is GREEN — the eviction posture is the correlated upstream pressure, NOT a fresh-cache warm-up on the warm cohort), and the third filed as an alternate: viewer_driven_reconnect_storm_saturation with a low confidence that captures the cohort-reconnect-storm shape before cohort.reconnect_storm_ratio baseline +edge.egress_kbps flat remove it.
origin_shield_saturation_under_correlated_cache_miss_storm — a correlated upstream cache-miss storm fanned into the origin-shield tier past its pre-provisioned ceiling; the warm edge POPs' cache layer evicted correlated segments (cache.eviction_posture on the warm cohort read fail), the replay-origin took the fan-in uphill, and the shield's queue stopped accepting fan-in. The cohort itself was calm — the discriminator against the marquee live-event viewer-storm page. Four signals line up: origin.replay_origin_origin_fetch_qps 4.2× vs 1.05× pre-event baseline, cdn_shield.origin_segment_cache_miss_fanin 0.62 vs 0.11 pre-event baseline (correlated fan-in), cdn_shield.queue_depth_concurrent 1.00 of slot ceiling vs 0.41 pre-event baseline, cdn_shield.queue_p99_wait_ms 4.21s vs 0.21s baseline (exceeds the 1.5s slot budget by ~3×).
cdn_edge_pop_warmup_under_eviction was the secondary signal ranked at 31% — but it's tagged ruled out by so the warm-edge-POP cache-layer lane reads correctly. The warm edge POPs' cdn_shield.segment_leg_cache_hit reads GREEN at the ladder on the affected cohort; the eviction posture is the correlated upstream pressure on the warm cohort, NOT a fresh-cache warm-up on the warm cohort's ladder itself. The dismissal rule was "rank the cause on the warm-edge-POP segment-leg cache-hit posture, not on the player-visible cache-eviction symptom alone"; the upstream-origin-shield-fan-in nature of the failure is the decisive signal pattern.
- Region: na-east weekday primetime broadcast (cdn-A/atl01 affected tier · warm edge POP cohort)
- Status: resolved (window closed; Tier 1 fail-open approved; Tier 2 operator-team codification staged)
- Opened: 2026-08-20 18:42 UTC
- Spread: contained to the na-east affected tier — apac and eu-west weekday primetime cohorts unaffected; adjacent non-primetime cohorts on the same vendor reads at baseline; warm edge POPs on the adjacent tiers ship green on segment-leg cache-hit (the eviction posture was CORRELATED to the na-east affected tier only).
Above the 80% threshold the agent treats as a confident top-hypothesis filing. cdn_edge_pop_warmup_under_eviction · 0.31 was cleared explicitly because the warm edge POPs' cdn_shield.segment_leg_cache_hit reads GREEN at the ladder on the affected cohort — the warm-edge-POP cache layer lane is healthy at the affected cohort, ruling out a fresh-cache warm-up on the warm cohort. The cohort.reconnect_storm_ratio: 0.04 (baseline) + edge.egress_kbps: 8214 (flat) pairing clears viewer_driven_reconnect_storm_saturation as the alternate — the discriminator against the marquee live-event viewer-storm page on this incident.
Who clears the gate
The governed-action policy on this incident is split across three authorization tiers — the agent sheds correlated cache-misses to warm edge POPs autonomously under an explicit gate (Tier 0 — who can shed traffic), surfaces the fail-open posture on the shield's circuit-breaker to on-call + reliability team (Tier 1 — who can fail open), and hands the new origin-failover policy + warm-edge-POP cache-routing logic codification to operator-team sign-off forward (Tier 2 — who can escalate). Each tier has an explicit gate; each gate names the threshold before the action lands, and the lane the action belongs to is what makes the split distinct from a generic postmortem.
shed_correlated_cache_misses_to_warm_pops_via_origin_failover — autonomous under gate: top_confidence >= 0.80 AND cdn_shield.origin_segment_cache_miss_fanin >= 0.35 AND cohort_demand.in_window_constraint == true on the affected cohort. The agent sheds the correlated cache-miss storm fan-in to a warm edge POP that has the segments cached — origin.replay_origin_origin_fetch_qps drops below 1.0× (within pre-provisioned envelope); the cohort_join_stall_ratio collapses below 0.05 within ten consecutive 30s windows. The cohort itself is untouched — the upstream fan-in is rerouted, the cohort's egress shape is preserved (FLAT at expected rate by definition on this incident).
fail_open_on_shield_ceiling_via_origin_circuit_breaker — surfaced to humans: cdn_shield.queue_concurrency_ceiling_pct == 1.00 AND viewer_visible_config_change_required. The fail-open posture on the shield's circuit-breaker is surfaced to on-call + reliability team — the trip thresholds are operator-team-owned config, NOT probe-driven. The agent holds off emitting the fail-open; the operator team code-approves the shield's circuit-breaker posture before the next correlated cache-miss storm window. A misanchor that surfaces on a future primetime isn't masked by a "we reconciled" close-out.
escalate_origin_shield_quota_codification_to_playbook — surfaced for operator-team approval forward: incident_resolved AND cohort_side close-out signals cleared. This is the learn-loop act — once the lane is closed, the operator team signs off on codifying the new origin-failover policy (correlated cache-miss-storm fan-in handling + origin.replay_origin_origin_fetch_qps ladder ceiling + warm-edge-POP failover policy) into the cohort probe fan-in for future correlated cache-miss storm windows; the new probes are pinned for the next four weekday primetime windows so the false-positive rate can be measured before promotion.
Incident timeline
Ten events: detection on the na-east affected tier, classification across three ranked hypotheses (with the cdn_edge_pop_warmup_under_eviction lane tagged ruled out by and theviewer_driven_reconnect_storm_saturation lane tagged alternate), five acts the agent took under the authorization-tiered governed-action gates (Tier 0 autonomous under gate + the four staged verification gates T+30 s → T+15 m), three acts it surfaced to humans (Tier 1 shielded fail-open approved, Tier 2 operator-team sign-off forward, plus the operator ack), and the resolution. The right-hand "act" + "tier" tags are what makes this postmortem distinct from a generic write-up — they pin the split between autonomous agentic ops, the shield's circuit-breaker operator approval, and the operator-team sign-off forward. All times below are simulated telemetry — the disclosure at the top of this page applies to every minute offset on the timeline.
Today
- T+0mDetectionby cohort agent · na-east (atl01 shield tier · warm edge POP cohort)act · autonomous
Upstream replay-origin origin_fetch_qps climbing; cdn_shield.origin_segment_cache_miss_fanin lifts; cohort REJECTED as a viewer-storm driver
origin.replay_origin_origin_fetch_qps climbed to 4.2× the tier's pre-provisioned ceiling within a 90-second window; cdn_shield.origin_segment_cache_miss_fanin read 0.62 vs 0.11 pre-event baseline (correlated fan-in from the warm edge POPs into the replay-origin); cache.eviction_posture on the warm edge POPs read fail (correlated eviction posture across the warm cohort — segment-leg cache-hit failing on the SAME tier the shield is queueing on); cdn_shield.queue_depth_concurrent crossed 0.86 → 1.00 of the slot ceiling; cdn_shield.queue_p99_wait_ms climbed to 4.21s (exceeds the 1.5s slot budget by ~3×); cohort.cohort_join_stall_ratio lifted to 0.094 (vs 0.018 baseline) and cohort.cohort_rebuffer_ratio to 0.062 (vs 0.012 baseline); cohort.cohort_startup_time_p95_ms read 7430ms — BUT: cohort.reconnect_storm_ratio reads at 0.04 baseline AND edge.egress_kbps 8214 Kbps flat at expected rate. The cohort is calm. The driver is upstream.
Aug 20, 06:42:11 PM - T+1mClassificationby Streamwake reliability agentact · autonomous
Ranked: origin_shield_saturation_under_correlated_cache_miss_storm · dominant (0.84) · cdn_edge_pop_warmup_under_eviction · ruled out by (0.31) · viewer_driven_reconnect_storm_saturation · alternate (0.22)
Top hypothesis reads 84% confidence. cdn_edge_pop_warmup_under_eviction is ruled out because the warm edge POPs on the backend tier read segment-leg cache-hit GREEN on the affected cohort at the ladder — the eviction posture is the correlated upstream pressure, NOT a fresh-cache warm-up on the warm cohort. viewer_driven_reconnect_storm_saturation is the alternate (cohort.reconnect_storm_ratio reads at 0.04 baseline AND edge.egress_kbps is flat at 8214 Kbps — the discriminator against the marquee live-event viewer-storm page on this incident).
Aug 20, 06:43:11 PM - T+2mAutomated actionby Streamwake reliability agentact · autonomousTier 0
Governed · Tier 0 autonomous under gate: shed_correlated_cache_misses_to_warm_pops_via_origin_failover (who can shed traffic)
top_confidence 0.84 ≥ 0.80 gate cleared; cdn_shield.origin_segment_cache_miss_fanin 0.62 ≥ 0.35 gate cleared; cohort.demand.in_window_constraint == true gate cleared (marquee broadcast window shape). The agent sheds correlated cache-misses to warm edge POPs via origin failover — the upstream points the warm cohort to a warm edge POP that has the segments cached, the correlated cache-miss storm fan-in routes around the affected tier, and cdn_shield.origin_segment_cache_miss_fanin clears below 0.18 within 90 seconds.
Aug 20, 06:44:11 PM - T+5mStatus changeby Streamwake reliability agentact · autonomous
T+30 s staged gate — edge POP cache-hit re-anchor (NOT a close-out by itself)
cdn_shield.warm_edge_pop_segment_leg_cache_hit flipped pass on the routed warm edge POP within the first same-length cohort window; cdn_shield.origin_segment_cache_miss_fanin dropped from 0.62 → 0.18 within the same window; cohort.demand.in_window_constraint still true. Edge-side observable is GREEN — but this is the FIRST gated verification, NOT a close-out. The T+2 m shield queue depth collapsed gate is next.
Aug 20, 06:47:11 PM - T+7mStatus changeby Streamwake reliability agentact · autonomous
T+2 m staged gate — shield queue depth collapsed (NOT a close-out by itself)
cdn_shield.queue_depth_concurrent dropped from 1.00 → 0.41 (within pre-event baseline tolerance on the affected tier); cdn_shield.queue_p99_wait_ms cleared to 0.32s (within the 1.5s slot budget); origin.replay_concurrency_ceiling_pct re-anchored at 0.62 (well within ceiling). Shield-side gate cleared — but this is the SECOND gated verification, NOT a close-out. The T+5 m cohort close-out probes are next.
Aug 20, 06:49:11 PM - T+9mSurfaced to humanby agent → on-call + reliability teamact · surfaced to humansTier 1
Governed · Tier 1 surfaced to humans: fail_open_on_shield_ceiling_via_origin_circuit_breaker (who can fail open)
Viewer-visible config change required. The fail-open posture on the shield's circuit-breaker is surfaced to on-call + reliability team — the trip thresholds are operator-team-owned config, NOT probe-driven. The agent holds off emitting the fail-open and the operator team code-approves the shield's circuit-breaker posture before the next correlated cache-miss storm window.
Aug 20, 06:51:11 PM - T+10mAutomated actionby Streamwake reliability agentact · autonomous
T+5 m staged gate — cohort join-stall + rebuffer + startup-time within tolerance (PARTIAL close-out)
cohort.cohort_join_stall_ratio settled at 0.022 (within baseline tolerance 0.05); cohort.cohort_rebuffer_ratio settled at 0.013 (within tolerance); cohort.cohort_startup_time_p95_ms settled at 2010ms (within startup-time baseline tolerance). The cohort side reads GREEN — but this is the THIRD gated verification, a PARTIAL close-out, NOT the full close-out. The T+15 m full cohort re-anchor gate is next; cohort.reconnect_storm_ratio still at 0.04 baseline; edge.egress_kbps still 8214 Kbps flat.
Aug 20, 06:52:11 PM - T+12mSurfaced to humanby on-callact · surfaced to humans
On-call ack; fail-open posture on the shield circuit-breaker approved
On-call + reliability team acked within 88s; the shield's circuit-breaker posture approved against the next correlated cache-miss storm window; reviewer confirmed the agent's rank and that the Tier 0 shed_correlated_cache_misses_to_warm_pops had already prevented the cohort from doubling into a secondary cohort-side stall before the upstream fan-in dropped.
Aug 20, 06:54:11 PM - T+22mAutomated actionby Streamwake reliability agentact · autonomous
T+15 m staged gate — full cohort re-anchor + replay-origin concurrency back to baseline (FULL close-out)
cohort.cohort_join_stall_ratio, cohort.cohort_rebuffer_ratio, cohort.cohort_startup_time_p95_ms settle within tolerance over the next same-length cohort window; origin.replay_concurrency_ceiling_pct re-anchors at 0.41 (well within ceiling — back to pre-event baseline); origin.replay_origin_origin_fetch_qps re-anchors at 1.05x (within pre-provisioned envelope); cdn_shield.queue_depth_concurrent at 0.41 (within pre-event baseline). Cohort-side + shield-side close-out signals cleared on all four gated verifications — NOT on cdn_shield.queue_depth_concurrent returning to baseline alone.
Aug 20, 07:04:11 PM - T+36mResolutionby Operator + agentact · surfaced to humansTier 2
Incident resolved; Tier 0 shed landed autonomously; Tier 1 fail-open approved; Tier 2 operator-team codification pass staged
Cohort-side + shield-side close-out signals cleared on the na-east affected tier (cohort_join_stall_ratio + cohort_rebuffer_ratio + cohort_startup_time_p95_ms + origin.replay_concurrency_ceiling_pct + cdn_shield.queue_depth_concurrent all within tolerance over the next same-length cohort window) — NOT on cdn_shield.queue_depth_concurrent returning to baseline alone, NOT on edge.egress_kbps staying FLAT at expected baseline (cohort is calm by definition on this incident — that's the discriminator, not a close-out). Tier 2 operator-team sign-off forward staged for the next four weekday primetime windows — the learn-loop codification of the correlated cache-miss-storm probe + the new origin-failover policy.
Aug 20, 07:18:11 PM
- classify · ranked three hypotheses with confidence in 90 s; cdn_edge_pop_warmup_under_eviction · ruled out by tagged (warm edge POPs' segment-leg cache-hit GREEN at the ladder)
- Tier 0 shed · shed_correlated_cache_misses_to_warm_pops_via_origin_failover under the top_confidence ≥ 0.80 + cdn_shield.origin_segment_cache_miss_fanin ≥ 0.35 + cohort_demand.in_window_constraint gate (who can shed traffic)
- T+30 s status · edge POP cache-hit re-anchor observed (NOT close-out — first staged gate)
- T+2 m status · shield queue depth collapsed (NOT close-out — second staged gate)
- T+5 m status · cohort close-out probes within tolerance (PARTIAL close-out — third staged gate)
- T+15 m status · full cohort re-anchor + replay-origin concurrency back to baseline (FULL close-out — fourth staged gate)
- on-call + reliability team · Tier 1 surface — code-approved the fail_open_on_shield_ceiling_via_origin_circuit_breaker posture for the next correlated cache-miss storm window (who can fail open)
- operator team · Tier 2 surface — sign-off forward on codifying the new origin-failover policy + warm-edge-POP cache-routing logic into the cohort probe fan-in (who can escalate)
- reliability team · assigned the public postmortem write-up (this page) — the learn-loop section codifies the audit step on this incident
What "recovery" actually looks like on this incident
Recovery on this incident is verified across the four staged gates T+30 s → T+15 m on the affected cohort + on the upstream origin-shield tier — NOT on infrastructure-green alone. The cache-layer / warm-edge-POP / shield-lane posture is part of the recovery lens, but it does not by itself prove the cohort is delivering green playback; the cohort-side close-out probes do. The audit step writes the close-out signal into the playbook as a four-gate verification — edge POP cache-hit re-anchor at T+30 s · shield queue depth collapsed at T+2 m · cohort close-out probes within tolerance at T+5 m · full cohort re-anchor + replay-origin concurrency back to baseline at T+15 m.
Recovery on this incident is verified by a four-gate verification ladder: T+30 s — edge POP cache-hit re-anchor clears the warm-edge-POPs'cdn_shield.segment_leg_cache_hit on the routed warm POP (first staged gate, NOT a close-out by itself); T+2 m — shield queue depth collapsed clears cdn_shield.queue_depth_concurrent below 0.80 AND cdn_shield.queue_p99_wait_ms under 1.5s on the affected tier (second staged gate, NOT a close-out by itself); T+5 m — cohort close-out probes within tolerance (cohort_join_stall_ratio + cohort_rebuffer_ratio + cohort_startup_time_p95_ms) clears below tolerance on the affected cohort (third staged gate, PARTIAL close-out); T+15 m — full cohort re-anchor (origin.replay_concurrency_ceiling_pct re-anchors at 0.41 on the affected tier + origin.replay_origin_origin_fetch_qps re-anchors at 1.05× within pre-provisioned envelope) — full close-out (fourth staged gate).
cdn_shield.queue_depth_concurrent returning to its pre-event baseline alone is NOT a close-out — the shield queue clears geometrically (the slot has headroom again) without necessarily clearing the cohort's late-join stall. The T+30 s edge POP cache-hit re-anchor is NOT a close-out — the warm-edge-POPs' cache layer clears before the cohort lands downstream. The T+2 m shield queue depth collapsed read is NOT a close-out — the shield side cleared ahead of the cohort side, as expected. The cohort egress staying FLAT at expected baseline is NOT a close-out — the cohort is calm by definition on this incident (calm cohort = upstream driven), but calibration-vs-shape is not the same as calibration-vs-cohort-delivering-green-playback.
A fix that normalizes origin.replay_origin_origin_fetch_qps + cdn_shield.origin_segment_cache_miss_fanin + cdn_shield.queue_depth_concurrent without clearing the affected cohort's join stall, rebuffer, or startup-time within tolerance is a fix that didn't reach the cohort. The cohort was calm by definition on this incident — meaning the upstream origin-shield fan-in is the pinch, NOT a viewer-side reconnect storm — so a read that declares recovery on the upstream gauges alone is the read that misses the lane forward. The audit step codifies it: cohort-side + shield-side four-gate verification, NOT infrastructure-green alone.
Anatomy of the evidence packet
The two packets on the failing source — an origin-shield probe packet on the na-east affected tier + warm edge POP cohort (with the upstream replay-origin origin_fetch_qps climbing past the tier's pre-provisioned ceiling, the cache-miss fan-in lifting correlated across the warm cohort, and the cohort itself staying CALM on reconnect-storm + edge egress shape — the discriminator against the marquee live-event viewer-storm page), and the agent timeline response with the three ranked hypotheses, the authorization-tiered remediation policy, and the explicit four staged gates T+30 s → T+15 m. The probe packet is what the agent decided on; the timeline response is what the agent emitted.
GET /live/event/stream.m3u8 HTTP/1.1
host: cdn.example.com
accept: application/vnd.apple.mpegurl
----- cycle 0 (T+0m, before shed_correlated_cache_misses_to_warm_pops + fail_open_on_shield_ceiling) -----
# manifest returned clean by the shield tier, queue depth climbing — UPSTREAM-driven
HTTP/2 200
content-type: application/vnd.apple.mpegurl
server-timing: manifest-fetch;dur=58
x-cdn: cdn-A/atl01 ← affected tier (origin-shield target)
x-packager: pkg-prime-04
x-shield-tier: shield-NA-3 ← the affected replay-origin tier
xo-content-digest: sha256=...
#EXTM3U
#EXT-X-VERSION:9
#EXT-X-TARGETDURATION:6
#EXT-X-MEDIA-SEQUENCE:84217
#EXT-X-MAP:URI="init.mp4"
#EXTINF:6.000,
084217.ts
# origin.replay_origin_origin_fetch_qps: 4.2× (correlated cache-miss storm uphill into the tier's pre-provisioned ceiling)
# cdn_shield.origin_segment_cache_miss_fanin: 0.62 (vs 0.11 pre-event baseline — correlated fan-in
# from the warm edge POPs into the replay-origin)
# cache.eviction_posture on warm edge POPs: fail (correlated eviction posture across the warm cohort;
# segment-leg cache-hit failing on the SAME tier the shield is queueing on)
# cdn_shield.queue_depth_concurrent: 1.00 (slot ceiling — vs 0.41 pre-event baseline)
# cdn_shield.queue_p99_wait_ms: 4.21 (vs 0.21s baseline; exceeds the 1.5s slot budget)
# cdn_shield.queue_concurrency_ceiling_pct: 1.00 (the shield has stopped accepting fan-in)
# origin.replay_concurrency_ceiling_pct: 0.94 (replay-origin playing out — JUST UNDER ceiling;
# upstream is queueing, replay-origin is hot)
# cdn_shield.warm_edge_pop_segment_leg_cache_hit: fail (warm edge POPs' segment-leg green fails
# on the same tier — the warm cohort shares the fate)
# edge.egress_kbps: 8214 (FLAT at expected baseline — cohort IS delivering traffic)
# edge.manifest_route_5xx_ratio: 0.000 (no 5xx; the shield is queueing, not erroring)
# cohort.cohort_join_stall_ratio: 0.094 (vs 0.018 baseline, 90s cohort window — late-join
# stalled behind the queue tail on the affected tier)
# cohort.cohort_rebuffer_ratio: 0.062 (vs 0.012 baseline — playback stalled on the affected tier)
# cohort.cohort_startup_time_p95_ms: 7430 (vs 1840 baseline — cohort start-up time hitting the 8s threshold)
# cohort.reconnect_storm_ratio: 0.04 ← BASELINE — NOT a viewer-driven storm. The cohort is calm.
# cohort.demand.in_window_constraint: true (cohort is in the marquee broadcast window — demand is constrained)
# encoder.bitrate_target_delta_pct: 0.18 (within baseline — encoder profile reads clean)
# origin.replay_orig_egress_kbps: 4211 (stable; origin egress not the pinch — shield queue is)
# cdn_pop.partial_segment_warmup_state: fail (warm edge POPs NOT primed on this tier — eviction posture
# is the correlated upstream pressure)
# cdn_pop.affected_region_pop_id: cdn-A/atl01
# isp.cell_load.on_affected_region: pass (cell-load 0.43 vs 0.51 baseline reference)
----- cycle 1 (T+~3m, after shed_correlated_cache_misses_to_warm_pops_via_origin_failover) -----
HTTP/2 200
content-type: application/vnd.apple.mpegurl
server-timing: manifest-fetch;dur=32
x-cdn: cdn-A/atl01
x-shield-tier: shield-NA-3-shedded ← correlated fan-in routed to a warm POP
xo-content-digest: sha256=...
#EXTM3U
#EXT-X-VERSION:9
#EXT-X-TARGETDURATION:6
#EXT-X-MEDIA-SEQUENCE:84221
#EXT-X-MAP:URI="init.mp4"
#EXTINF:6.000,
084221.ts
# origin.replay_origin_origin_fetch_qps: 0.91× (correlated cache-miss storm at bay — fan-in routed
# to a warm edge POP that has the segments cached)
# cdn_shield.origin_segment_cache_miss_fanin: 0.18 (within pre-event baseline)
# cache.eviction_posture on warm edge POPs: pass (correlated eviction posture cleared on the warm cohort)
# cdn_shield.queue_depth_concurrent: 0.41 (within the slot's pre-event baseline)
# cdn_shield.queue_p99_wait_ms: 0.32 (within slot budget)
# origin.replay_concurrency_ceiling_pct: 0.62 (replay-origin playing out — within ceiling)
# cdn_shield.warm_edge_pop_segment_leg_cache_hit: pass (warm edge POPs' segment-leg green once routed)
# edge.egress_kbps: 8214 (FLAT at expected baseline — cohort continues to deliver traffic)
# cohort.cohort_join_stall_ratio: 0.022 (within baseline tolerance — within cohort_baseline)
# cohort.cohort_rebuffer_ratio: 0.013 (within baseline tolerance)
# cohort.cohort_startup_time_p95_ms: 2010 (within startup_time baseline tolerance)
# cohort.reconnect_storm_ratio: 0.04 ← STILL BASELINE — the cohort has not changed shape;
# it was never the driver
# cohort.demand.in_window_constraint: true (still in marquee broadcast window)
# governed_action_emitted: shed_correlated_cache_misses_to_warm_pops_via_origin_failover,
# fail_open_on_shield_ceiling_via_origin_circuit_breaker,
# surface_origin_shield_quota_codification_to_operator_team- origin.replay_origin_origin_fetch_qps →
4.2× (correlated cache-miss storm uphill into the tier's pre-provisioned ceiling — vs 1.05× pre-event baseline) - cdn_shield.origin_segment_cache_miss_fanin →
0.62 vs 0.11 pre-event baseline (correlated fan-in across the warm edge POPs into the replay-origin) - cdn_shield.queue_depth_concurrent →
1.00 of slot ceiling (vs 0.41 pre-event baseline) - cdn_shield.queue_p99_wait_ms →
4.21 (exceeds 1.5s slot budget by ~3×) - cohort.reconnect_storm_ratio → 0.04 (BASELINE) the cohort is calm. edge.egress_kbps → 8214 Kbps, FLAT at expected baseline — the cohort IS delivering traffic. cdn_shield.warm_edge_pop_segment_leg_cache_hit → fail on the SAME tier, but pass at the ladder when routed — ruling out a fresh-cache warm-up on the warm cohort.
{
"stream_id": "ckshieldsatcorrcachemiss6185",
"source": "https://cdn.example.com/live/event/stream.m3u8",
"protocol": "HLS / CMAF / shield-tier / replay-origin / warm-edge-POP cache layer",
"checked_at": "2026-08-20T18:42:11Z",
"ranked_hypotheses": [
{
"rank": 1,
"hypothesis": "origin_shield_saturation_under_correlated_cache_miss_storm",
"tag": "dominant",
"confidence": 0.84,
"evidence_signals": [
"origin.replay_origin_origin_fetch_qps → fail (4.2× the tier's pre-provisioned ceiling — correlated cache-miss storm fan-in uphill into the replay-origin)",
"cdn_shield.origin_segment_cache_miss_fanin → fail (0.62 vs 0.11 pre-event baseline — CORRELATED fan-in across the warm edge POPs into the replay-origin)",
"cache.eviction_posture on warm edge POPs → fail (correlated eviction posture across the warm cohort — segment-leg cache-hit failing on the SAME tier the shield is queueing on)",
"cdn_shield.queue_depth_concurrent → fail (1.00 of slot ceiling vs 0.41 pre-event baseline — the shield has stopped accepting fan-in)",
"cdn_shield.queue_p99_wait_ms → fail (4.21s vs 0.21s baseline; exceeds the 1.5s slot budget)",
"cdn_shield.queue_concurrency_ceiling_pct → fail (1.00 — slot reads at ceiling; origin-failover eligible)",
"origin.replay_concurrency_ceiling_pct → warn (0.94 — replay-origin playing out — JUST UNDER ceiling; upstream is queueing, replay-origin is hot)",
"cohort.cohort_join_stall_ratio → fail (0.094 vs 0.018 baseline, 90s cohort window — late-join stalled behind the queue tail)",
"cohort.cohort_rebuffer_ratio → fail (0.062 vs 0.012 baseline — playback stalled on the affected tier)",
"cohort.cohort_startup_time_p95_ms → fail (7430ms vs 1840ms baseline — start-up time hitting the 8s threshold)",
"cohort.reconnect_storm_ratio → pass at 0.04 (BASELINE — discriminator against a viewer-driven reconnect storm; the cohort is calm, the upstream origin is the driver)",
"edge.egress_kbps → pass at 8214 Kbps (FLAT at expected baseline — discriminator: cohort IS delivering traffic; the upstream origin-shield fan-in is the pinch, NOT the cohort's egress)",
"edge.manifest_route_5xx_ratio → pass (no 5xx — the shield is queueing, not erroring)",
"encoder.bitrate_target_delta_pct → pass (0.18 within baseline — encoder profile reads clean; failure is upstream)")
]
},
{
"rank": 2,
"hypothesis": "cdn_edge_pop_warmup_under_eviction",
"tag": "ruled_out_by",
"confidence": 0.31,
"evidence_signals": [
"warm edge POPs' cdn_shield.segment_leg_cache_hit reads pass AT BASE TIER (the warm cohort has the segments in cache)",
"correlated eviction posture on the warm edge POPs is a STAGE-shaped signal — but on backend warm edge cohorts, segment_leg cache-hit is GREEN at the ladder, so the failure cannot be a fresh-cache warm-up on the warm cohort's ladder",
"the segment-leg cache-hit posture on cdn_pop.partial_segment_warmup_state is fail (warm edge POPs NOT primed on this tier — eviction posture is the correlated upstream pressure, NOT a clean cache eviction on the warm cohort)"
]
},
{
"rank": 3,
"hypothesis": "viewer_driven_reconnect_storm_saturation",
"tag": "alternate",
"confidence": 0.22,
"evidence_signals": [
"cohort.reconnect_storm_ratio reads at 0.04 — BASELINE, NOT 1.00 like the marquee live-event viewer-storm page",
"edge.egress_kbps reads 8214 Kbps FLAT at expected baseline — the cohort is NOT feeding the queue",
"if a viewer-driven reconnect storm were stateful, the late-join cohort on the affected tier would also misfire — it is not; the cohort reconnect-storm ratio is at baseline, the cohort_join_stall_ratio is high because of the shield queue tail (NOT a viewer-side reconnect storm)"
]
}
],
"governed_actions": [
{
"action": "shed_correlated_cache_misses_to_warm_pops_via_origin_failover",
"decision_lane": "autonomous",
"authorization_tier": "Tier 0",
"account_name": "who can shed traffic",
"type": "governed",
"gating": "top_confidence >= 0.80 AND cdn_shield.origin_segment_cache_miss_fanin >= 0.35 AND cohort_demand.in_window_constraint == true",
"evidence": "top-hypothesis confidence 0.84; cdn_shield.origin_segment_cache_miss_fanin 0.62; cohort.demand.in_window_constraint == true (marquee broadcast window shape); origin.replay_origin_origin_fetch_qps 4.2x (correlated cache-miss storm uphill into the tier's pre-provisioned ceiling)",
"expected_effect": "replay_origin_origin_fetch_qps drops below 1.0x (within pre-provisioned envelope); cohort_join_stall_ratio collapses below 0.05 within ten consecutive 30s windows; the correlated cache-miss storm fan-in routes to a warm edge POP that has the segments cached"
},
{
"action": "fail_open_on_shield_ceiling_via_origin_circuit_breaker",
"decision_lane": "surfaced_to_humans",
"authorization_tier": "Tier 1",
"account_name": "who can fail open",
"type": "governed",
"gating": "cdn_shield.queue_concurrency_ceiling_pct == 1.00 AND viewer_visible_config_change_required",
"evidence": "cdn_shield.queue_concurrency_ceiling_pct pegged at 1.00 on the affected tier; the circuit-breaker trip thresholds are operator-team-owned config, NOT probe-driven; the agent holds off emitting and surfaces the fail-open posture to on-call + reliability team",
"expected_effect": "on-call + reliability team code-approve the fail-open posture on the shield's circuit-breaker; the shield's queue_concurrency_ceiling_pct re-anchors below 1.00 on the next triggered probe cycle"
},
{
"action": "escalate_origin_shield_quota_codification_to_playbook",
"decision_lane": "surfaced_to_humans",
"authorization_tier": "Tier 2",
"account_name": "who can escalate",
"type": "governed",
"gating": "incident_resolved AND cohort_side close-out signals cleared",
"evidence": "cohort.cohort_join_stall_ratio + cohort.cohort_rebuffer_ratio + cohort.cohort_startup_time_p95_ms within tolerance over the next same-length cohort window; the operator-team codifies the new origin-failover policy into the playbook after the close-out probe lands; THIS IS THE LEARN-LOOP ACT",
"expected_effect": "operator-team signs off on pinning the new origin-shield quota policy (correlated cache-miss-storm fan-in handling + replay_origin_origin_fetch_qps ladder ceiling + warm-edge-POP failover policy) into the cohort probe fan-in for future correlated cache-miss storm windows; the false-positive rate is measured over the next four weekday primetime windows before promotion"
}
],
"verification_window": {
"close_out_signal": "staged gated verification — T+30 s edge POP cache-hit re-anchor · T+2 m shield queue depth collapsed · T+5 m cohort close-out probes within tolerance · T+15 m full cohort re-anchor + replay-origin concurrency back to baseline",
"close_out_probes": [
"cohort.cohort_join_stall_ratio within tolerance over the next same-length cohort window",
"cohort.cohort_rebuffer_ratio within tolerance",
"cohort.cohort_startup_time_p95_ms within tolerance",
"edge.egress_kbps flat at expected baseline (CORRELATED fan-in did NOT alter cohort shape)"
],
"NOT_close_out_signal": [
"cdn_shield.queue_depth_concurrent returning to baseline alone — PASSES geometrically, does NOT pass cohort-side",
"origin.replay_concurrency_ceiling_pct returning below 1.00 alone — same caveat",
"edge.egress_kbps staying FLAT at expected baseline — this discriminator is "what we expected", NOT "what is closing out" (cohort is calm by definition)"
],
"explicit_note": "'replay-origin queue normalized' != 'cohort is delivering green playback'. Recovery is verified cohort-side + shield-side — staged T+30 s → T+15 m. NOT infrastructure-green alone."
},
"surfaced_to_humans": [
{"owner": "on-call", "task": "approve the fail_open_on_shield_ceiling_via_origin_circuit_breaker posture before the next correlated cache-miss storm window"},
{"owner": "reliability team", "task": "page for the origin-shield saturation root-cause review"},
{"owner": "operator team", "task": "codify the new origin-failover policy + warm-edge-POP cache-routing logic into the playbook (Tier 2 · operator-team sign-off forward)"},
{"owner": "reliability team", "task": "assign the public postmortem write-up (this page)"}
]
}- Tier 0 →
shed_correlated_cache_misses_to_warm_pops_via_origin_failover(autonomous under gate: top_confidence ≥ 0.80 + cdn_shield.origin_segment_cache_miss_fanin ≥ 0.35 + cohort_demand.in_window_constraint — who can shed traffic) - Tier 1 →
fail_open_on_shield_ceiling_via_origin_circuit_breaker(surfaced: cdn_shield.queue_concurrency_ceiling_pct == 1.00 + viewer_visible_config_change_required; on-call + reliability team code-approve the shield's circuit-breaker posture — who can fail open) - Tier 2 →
escalate_origin_shield_quota_codification_to_playbook(surfaced for operator-team approval forward: incident_resolved + cohort-side close-out signals cleared; the learn-loop act — who can escalate) - surfaced → paged on-call for the origin-shield saturation root-cause review
- close-out signal → four staged gates on the cohort-side + shield-side:
cohort_join_stall_ratio,cohort_rebuffer_ratio,cohort_startup_time_p95_ms,origin.replay_origin_origin_fetch_qpscdn_shield.queue_depth_concurrentcdn_shield.warm_edge_pop_segment_leg_cache_hitsettle within tolerance on the four staged gates T+30 s → T+15 m — NOTcdn_shield.queue_depth_concurrentreturning to baseline alone
Detection, classify, mitigate, recover (simulated telemetry)
Four timing windows on the postmortem timeline, each read off the cohort probe cadence. The figures are simulated telemetry — the disclosure near the top of this page applies to every figure on this list. Note that the close-out window is verified cohort-side + shield-side on the na-east affected tier for the affected window (NOT infrastructure-green alone, NOT cohort egress shape alone).
~5 s
origin.replay_origin_origin_fetch_qps crossed 1.05× → 4.2× within a 30 s window; the agent surfaced the detector from the upstream replay-origin fan-in at T+5 s on the affected tier.
~1 m
origin_shield_saturation_under_correlated_cache_miss_storm · dominant ranked at 0.84 confidence with three ranked hypotheses at T+1 m — discriminator is cohort.reconnect_storm_ratio at 0.04 baseline + edge.egress_kbps flat at expected rate.
~7 m
Tier 0 shed_correlated_cache_misses_to_warm_pops_via_origin_failover queued at T+2 m; cdn_shield.queue_depth_concurrent re-anchored below 0.80 by T+7 m; on-call + reliability team ack on the fail-open posture by T+12 m.
~22 m
cohort.cohort_join_stall_ratio + cohort.cohort_rebuffer_ratio + cohort.cohort_startup_time_p95_ms + origin.replay_concurrency_ceiling_pct cleared within tolerance at T+22 m — four staged gates closed cohort-side + shield-side, NOT infrastructure-green alone.
The loop on this incident
Three steps close the lane on a na-east correlated upstream cache-miss storm incident that fans into the origin-shield tier. The fix is split explicitly into the authorization-tiered governed-action branches — Tier 0 acts on its own under the shed-correlated-cache-misses gate (who can shed traffic), Tier 1 surfaces the fail-open posture on the shield's circuit-breaker (who can fail open), Tier 2 surfaces for operator-team sign-off forward on the new origin-failover policy + warm-edge-POP cache-routing logic (who can escalate).
The probe set fans in across the upstream replay-origin + warm edge POP cohort connected-state lane. The upstream reads origin.replay_origin_origin_fetch_qps → 4.2×, cdn_shield.origin_segment_cache_miss_fanin → 0.62, and cdn_shield.queue_concurrency_ceiling_pct → 1.00; the warm-edge-POPs' cache layer reads cache.eviction_posture → fail on the affected tier; the cohort reads cohort.cohort_join_stall_ratio → 0.094, cohort_rebuffer_ratio → 0.062, and cohort_startup_time_p95_ms → 7430ms. The cohort is CALM: reconnect_storm_ratio: 0.04 (BASELINE), and edge.egress_kbps: 8214 (flat) at expected baseline — the cohort is not driving the fan-in.
cdn_edge_pop_warmup_under_eviction · ruled out by (0.31) and viewer_driven_reconnect_storm_saturation · alternate (0.22).The discriminator against the warmup lane is the warm edge POPs' cdn_shield.segment_leg_cache_hit — reads GREEN at the ladder on the affected cohort, so the failure is NOT a fresh-cache warm-up on the warm cohort's ladder. The discriminator against the viewer-driven lane is cohort.reconnect_storm_ratio at 0.04 baseline AND edge.egress_kbps flat at expected baseline — so the cohort is calm and the upstream origin-shield fan-in is the pinch.
The Tier 0 autonomous branch sheds correlated cache-misses to warm edge POPs via origin failover — clears the upstream fan-in within T+30 s and the cohort side within T+10 m on the staged gates. The Tier 1 surface branch surfaces the fail-open posture on the shield's circuit-breaker for on-call + reliability team approval before the next correlated cache-miss storm window. The Tier 2 forward branch codifies the new origin-failover policy + warm-edge-POP cache-routing logic into the cohort probe fan-in for future correlated cache-miss storm windows. Recovery is verified cohort-side + shield-side across the four staged gates T+30 s → T+15 m on the na-east affected tier.
Shed to a warm POP — or fail-open on the shield's circuit-breaker.
On a na-east weekday primetime broadcast where a correlated upstream cache-miss storm fans into the origin-shield tier past its pre-provisioned ceiling, the governed fix is a three-arm branch: an autonomous arm that sheds correlated cache-misses to warm POPs under explicit gates (Tier 0 — who can shed traffic), a surfaced-to-humans arm that fails-open on the shield's circuit-breaker (Tier 1 — who can fail open), and a forward-surfaced arm that codifies the new origin-failover policy into the cohort probe fan-in (Tier 2 — who can escalate). The three arms close the lane in the same incident window and forward.
Emit shed_correlated_cache_misses_to_warm_pops_via_origin_failover — the upstream origin-shield route the correlated fan-in to a warm edge POP that has the segments cached; the warm-edge-POPs' cache layer clears the fail and the correlated fan-in drops below the origin shield's pre-provisioned ceiling. The fix is Tier 0 autonomous under gate (top_confidence >= 0.80 AND cdn_shield.origin_segment_cache_miss_fanin >= 0.35 AND cohort_demand.in_window_constraint == true on the affected cohort) and surfaced for operator-team sign-off forward outside that gate so the operator team can review the false-positive rate before the shed lands. The cohort itself is untouched — the upstream fan-in is rerouted, the cohort's egress shape is preserved (FLAT at expected rate by definition on this incident).
cdn_shield.warm_edge_pop_segment_leg_cache_hit clears within the first same-length cohort window (T+30 s staged gate — first gate, NOT close-out); cdn_shield.queue_depth_concurrent settles below 0.80 within ten consecutive 30s windows (T+2 m staged gate — second gate, NOT close-out); cohort.cohort_join_stall_ratio + cohort.cohort_rebuffer_ratio + cohort.cohort_startup_time_p95_ms settle within baseline tolerance on ten consecutive 30s windows (T+5 m staged gate — third gate, PARTIAL close-out); origin.replay_concurrency_ceiling_pct + origin.replay_origin_origin_fetch_qps re-anchor below ceiling / within pre-provisioned envelope on a same-length cohort window (T+15 m — four staged gate, FULL close-out).
NOT a close-out signal: cdn_shield.queue_depth_concurrent returning to its pre-event baseline alone. The shield queue clears geometrically (the slot has headroom again) without necessarily clearing the cohort's late-join stall — recovery is verified cohort-side + shield-side on the four staged gates, NOT shield lane alone.
Surface fail_open_on_shield_ceiling_via_origin_circuit_breaker — on-call + reliability team code-approves the fail-open posture on the shield's circuit-breaker before the next correlated cache-miss storm window; the trip thresholds are operator-team-owned config, NOT probe-driven. Surfaced to humans because it requires viewer-visible config change the agent cannot perform — a misanchor that surfaces on a future primetime isn't masked by a "we reconciled" close-out. The Tier 1 surface arms the shield's circuit-breaker posture forward so future correlated cache-miss storms don't peg the queue at ceiling before the upstream fan-in reroutes.
cdn_shield.queue_concurrency_ceiling_pct re-anchors below 1.00 on the next triggered probe cycle (≤ 1 m after operator approval on the shield's circuit-breaker posture); the fail-open posture ships configured for the next correlated cache-miss storm window so future storms are absorbed before the shield queue pegs at ceiling; on-call + reliability team gets a single pane-of-glass view of the shield's circuit-breaker posture for the next four weekday primetime windows.
NOT a close-out signal: cohort.cohort_reconnect_storm_ratio reading at 0.04 baseline on this incident. The cohort is calm by definition on a correlated upstream cache-miss storm — that's the discriminator, not the close-out. Reading the cohort's connected-state signal as a close-out misses the lane forward.
A fix that normalizes origin.replay_origin_origin_fetch_qps + cdn_shield.origin_segment_cache_miss_fanin + cdn_shield.queue_depth_concurrent without clearing the affected cohort's join stall, rebuffer, or startup-time within tolerance is a fix that didn't reach the cohort. The cohort was calm by definition on this incident — meaning the upstream origin-shield fan-in is the pinch, NOT a viewer-side reconnect storm — so a read that declares recovery on the upstream gauges alone is the read that misses the lane forward. The audit step on this incident writes the close-out signal into the playbook as "verify the four staged gates on cohort_join_stall_ratio + cohort_rebuffer_ratio + cohort_startup_time_p95_ms + origin.replay_origin_origin_fetch_qps + cdn_shield.queue_depth_concurrent settle within tolerance over the next same- length cohort window" — not "verify the upstream origin-shield returned to baseline", not "verify the warm-edge-POPs' cache layer cleared", not "verify the cohort egress shape is steady".
Codifying the new origin-failover policy + cache-miss probe
The Tier 2 surface for operator-team sign-off forward on this incident codifies three probes + one policy change into the cohort probe fan-in — origin.replay_origin_origin_fetch_qps pinned into the cohort fan-in for the next weekday primetime broadcast window (operator-approved; surfaces on every na-east affected tier hit); cdn_shield.origin_segment_cache_miss_fanin pinned into the cohort fan-in for the next four weekday primetime windows so the false-positive rate can be measured before promotion; cdn_shield.warm_edge_pop_segment_leg_cache_hit pinned into the cohort fan-in to track the warm-edge-POPs' cache-layer posture; and the new origin-failover policy (correlated cache-miss-storm fan-in handling + replay-origin-failover ladder ceiling) operator-team approved forward.
The audit step codifies that origin.replay_origin_origin_fetch_qps +cdn_shield.origin_segment_cache_miss_fanin + cdn_shield.queue_depth_concurrent + cdn_shield.warm_edge_pop_segment_leg_cache_hit + cohort.cohort_join_stall_ratio + cohort.cohort_rebuffer_ratio + cohort.cohort_startup_time_p95_ms settle within tolerance over the four staged gates — meaning the upstream fan-in + cohort-side close-out fan-in tracks across every correlated cache-miss storm window, the operator-team sign-off goes on the new origin-failover policy + warm-edge-POP cache-routing logic codification, and the cohort fan-in carries the correlated-cache-miss signal lane forward. The cache-leg / shield / warm-edge-POP / cohort egress readings stay CALM through the entire window — those are geometric, NOT cohort-side close-out signals.
Want Streamwake to catch this on your upstream origin-shield?
Sign up, register a na-east affected-tier probe, and the same origin.replay_origin_origin_fetch_qps · cdn_shield.origin_segment_cache_miss_fanin · cache.eviction_posture · cohort.cohort_join_stall_ratio · cohort.reconnect_storm_ratio probes that produced the timeline above run on every prime-cohort refresh — and surface in a Slack channel, a webhook, or the streams dashboard.
Synthetic Incident — This scenario uses simulated telemetry constructed from documented streaming behaviors. It does not represent a Streamwake customer outage.
- Pick a recent on-call incident — manifest stall, edge miss, player-side stall, or peer congestion.
- We replay it through the same reliability-agent probe cascade used on the postmortem above.
- You walk away with a written what-could-have-been-Automated readout, not a sales deck.
Related writeups
The closest siblings cover the player-visible symptom lands and the lower-level failure-mode shapes: the existing /incident-lab/origin-shield-saturation-live-event-spike write-up (the marquee live-event viewer-spike form of the SAME origin-shield failure category — the parallel where the cohort itself IS the driver); the ABR ladder drift postmortem (a multi-CDN apac/teal mid-stream failover where the failure is on the cross-CDN ladder manifest, NOT origin-shield queue); the ISP-vs-CDN recovery criteria entry (cdn_failure · ruled out by name + cohort-side + last-mile close-out); the broader ISP-vs-CDN triage guide at /troubleshooting/isp-congestion-vs-cdn-failure (eight player-visible symptoms, one Streamwake probe pattern per row); the manifest fetch timeout storm at a regional edge POP (a regional edge POP returning manifest- timeouts above baseline during a quiet pre-peak window); and the live-event scale-out buffering postmortem (a marquee broadcast with a viewer-storm beyond the pre-provisioned capacity envelope).