ʮΞϥʔςΟϯάʯͷ࿩Λ͠Α͏ — SREcon΍࿦จ౳ͷ࠷ઌ୺ͱͷΪϟοϓΛΈΔ @yuuk1t · Yuuki Tsubouchi ͘͞ΒΠϯλʔωοτݚڀॴ ϗϯϚͰ͔ͬSREษڧձ #1 2026/06/25 

AIͷ࣌୅Ͱ΋ɺΞϥʔςΟϯά͸ඞཁ͔ʁ https://boristane.com/blog/the-software-development-lifecycle-is-dead/ “Monitoring: the last stage standing, and it needs to evolve.” 

Ironies of Automation ґવͱͯ͠ʮΞϥʔτʯ͕ਓؒʹͱΜͰ͘ΔͷͰ͸ Bainbridge, L. "Ironies of Automation." Automatica, 19(6):775–779, 1983. ੍ޚγεςϜ͕ߴ౓ʹͳΕ͹ͳΔ΄ͲɺਓؒͷΦϖ Ϩʔλͷߩݙ͕ΑΓॏཁʹͳΔͱ͍͏ൽ೑ 

Ͳ͏΍ͬͯਐԽ͍͔ͤͯ͘͞ʁ 7೥ͿΓʹΞϥʔτઃܭ͍ͯ͠ΔͷͰɺ࠷ઌ୺ͬͯͲ͜ʁͱͳͬͨ ࠷ઌ୺ ݱࡏ஍ ΪϟοϓΛຒΊΔ LLM Wikiʹબఆͨ͠จݙ Λ50ຊ΄ͲऔΓࠐΈ ΞϥʔςΟϯά͸఻౷త͗ͯ͢ɺओ୊ʹͳΓʹ͍͘ SREcon / ֶज़࿦จ (ICSE, ASE౳) ※1 https://gist.github.com/yuuki/ce843c8089d84c71c56a0706f43fa574 ※1 

ࠓ೔ͷ͸ͳ͠ ࠷ઌ୺ SREcon / ֶज़࿦จ (ICSE, ASE౳) ݱࡏ஍ ΪϟοϓΛຒΊΔ ΞϥʔςΟϯάͷϥΠϑαΠΫϧ Λ4ϑΣʔζʹ෼͚ͯઃܭݪଇΛ நग़ yuuk1͕ؔΘΔεύίϯαʔϏε ͷGrafana Alerting 

ΞϥʔςΟϯάͷ4ϑΣʔζϞσϧ PHASE A อূ → PHASE B ཈੍ → PHASE C ഑৴࠷దԽ → PHASE D ղܾࢧԉ ෆཁͳΞϥʔτΛͲ ͏ݮΒ͔͢ʁ ड͚ख͕ԿΛ ͢Δ͔ʁ ԣஅ — Ξϥʔτ඼࣭ΛͲ͏ଌΔ͔ʁ Phase A ͕յΕ͍ͯΕ͹ɺB Ҏ߱ΛͲΕ͚ͩվળͯ͠΋ҙຯ͕ͳ͍ ԿͰΞϥʔτ͢Δ ͔ʁ Ξϥʔτ͢΂ ͖΋ͷ͕ຊ౰ʹΞϥ ʔτ͢Δ͔ʁ ୭ʹɺͲ͏ಧ͚Δ ͔ʁ 

AIεύίϯαʔϏε Ϣʔβʔ͸ΠϯλϥΫςΟϒϊʔ υʹSSH ෼ࢄϑΝΠϧγεςϜʹPythonε ΫϦϓτ΍σʔληοτΛ഑ஔ δϣϒεέδϡʔϥSlurmͷίϚ ϯυΛ࣮ߦ WebαʔϏεͱ͸͍ͩͿҧ͏ 

4ϑΣʔζʹরΒͨ͠Ϊϟοϓ Phase Gap Status A. อূ SLOϕʔεΞϥʔτ͸ະಋೖ ϧʔϧ݈શੑͷ CI ݕূͳ͠ɺৗઃΧφϦΞͳ͠ɺ… ⾠ ෆ଍ B. ཈੍ ݱࡏͷΞϥʔτ݅਺͸10݅/೔ ҎԼɻ ✓ OK C. ഑৴࠷దԽ Өڹൣғxଈ࣌ੑʹΑΔνϟϯωϧग़͠෼͚ͳͲ ✓ OK D. ղܾࢧԉ ௐࠪࣗಈԽ / RCA / ΤʔδΣϯτະணख ⾠ ෆ଍ ԣஅʢ඼࣭ʣ Quality of Alerts / ίετϞσϧະಋೖ ⾠ ෆ଍ େن໛ϚΠΫϩαʔϏεલఏͷख๏͸আ֎ʢΞϥʔτετʔϜͷ཈੍΍ΤεΧ Ϩʔγϣϯ࠷దԽΛػցֶशͰߦ͏΋ͷͳͲ͕࿦จͰ͸ଟ͘ఏҊ͞Ε͍ͯΔʣ 

4ϑΣʔζʹরΒͨ͠Ϊϟοϓ Phase Gap Status A. อূ SLOϕʔεΞϥʔτ͸ະಋೖ ϧʔϧ݈શੑͷ CI ݕূͳ͠ɺৗઃΧφϦΞͳ͠ɺ… ⾠ ෆ଍ B. ཈੍ ݱࡏͷΞϥʔτ݅਺͸10݅/೔ ҎԼɻ ✓ OK C. ഑৴࠷దԽ Өڹൣғxଈ࣌ੑʹΑΔνϟϯωϧग़͠෼͚ͳͲ ✓ OK D. ղܾࢧԉ ௐࠪࣗಈԽ / RCA / ΤʔδΣϯτະணख ⾠ ෆ଍ ԣஅʢ඼࣭ʣ Quality of Alerts / ίετϞσϧະಋೖ ⾠ ෆ଍ ࠓ೔͸ɺAͱԣஅ ΛΈ͍ͯ͘ 

PHASE Aɿ঱ঢ়ϕʔεΞϥʔτ ✕ ݪҼى఺ CPU > 90% Disk > 85% OOM detected → ΞΫγϣϯͰ͖ͳ͍͜ͱ͕ଟ͍ ✓ ঱ঢ়ى఺ SLOϕʔεͷΞϥʔτ ʮϢʔβʔ͕ӨڹΛड͚͍ͯΔʯ → ΞΫγϣϯ͸ඞਢͰ͋Δ͜ͱΛอূ Multi-burn-rate SLO alerts Thurgood, S., et al. "Alerting on SLOs." In Beyer, B. et al. (eds.), The Site Reliability Workbook, ch. 5. O'Reilly Media, 2018. 4 Golden Signals 

Ϊϟοϓᶃ ঱ঢ়ϕʔεΞϥʔτ ݪҼى఺ ঱ঢ়ى఺ ɾ ݱঢ়͸ɺݪҼى఺ͱ঱ঢ়ى఺͕ࠞࡏ͍ͯ͠Δ ɾ SLI/SLO͸ະಋೖ • δϣϒ౤ೖՄ༻ੑ • GPU ܭࢉ༰ྔ • ετϨʔδՄ༻ੑ • ΠϯλϥΫςΟϒϊʔυ΁ͷ౸ୡੑ ͷΑ͏ͳ΋ͷΛSLIͱ࣮ͯ͠૷͍ͨ͠ Ϣʔβʔͱͷڥք໘͕ଟ͍ͷͰઃܭ͕ ೉͍͠… • δϣϒεέδϡʔϥ systemdىಈ • NTP࣌ࠁಉظ஗Ԇ • IPΞυϨεଟॏઃఆ؂ࢹ • … PHASE DͷΞϥʔτௐࠪࣗಈԽʹΑ Γফ͠΍͘͢ͳΔ ো֐΍ώϠϦϋοτ͕ى͖͔ͯΒ௥Ճ ͢ΔΞϥʔτ͸ݪҼى఺͕ଟ͍ 

PHASE Aɿ Ξϥʔτઃఆ͕ͦ΋ͦ΋ͳ͍໰୊ 40.41% of detection failures ݕ஌ࣦഊͷ࠷େཁҼ͸ ʮͦ΋ͦ΋ରԠ͢Δ Ξϥʔτ͕ॻ͔Ε͍ͯͳ͍ʯ — Microsoft Ͱͷ࣮ূݚڀ [Ganatra+, FSE2023] Ganatra, V. et al. "Detection Is Better Than Cure: A Cloud Incidents Perspective." ESEC/FSE, pp. 310–321, 2023. ڞ௨͢Δґଘؔ܎Λ΋ͭผϚΠΫϩαʔϏεͷΞ ϥʔτϧʔϧ͔Βࣗಈਪન աڈͷΠϯγσϯτνέοτ͔Βࣗಈਪન 

PHASE Aɿͳ͔ͥΞϥʔτͰ͖͍ͯͳ͔ͬͨ໰୊ ໐Γ͗͢໰୊ʢfalse alarmʣ͸ٞ࿦ͭ͘͠͞Ε͍ͯΔ ੩͔ʹյΕΔ໰୊ʢsilent failureʣ͸͋·Γٞ࿦͞Εͳ͍ ↳ ϥϕϧΛม͑ͨॠؒʹΫΤϦ͕ۭΛฦ͢Α͏ͳࣄނ CI — ͦ΋ͦ΋յΕͨϧʔϧΛೖΕͳ͍ Daemon — ಈ͖ग़͔ͯ͠Β઴ਐతʹյΕͨ΋ͷΛݕग़ Lukasz Mierzwa, "Monitoring our Monitoring: How we validate our Prometheus alert rules,” The Cloudflare Blog, 2022, 
 https://blog.cloudflare.com/monitoring-our-monitoring/ cloudflare/pint: PrometheusͷϧʔϧϦϯλʔ ৗறΧφϦΞ vector(1) Λ֤ route ʹྲྀͯ͠ɺ ௨஌ܦ࿏ͷࢮ׆Λৗ࣌ݕূ 

Phase BʙD — ݱঢ়ͱকདྷ B. ཈੍ ✓ pending / keep_firing_for / silence / mute timing Λ࢖͍෼͚͍ͯΔ C. ഑৴࠷దԽ ✓ ඞਢϥϕϧ͕͚ܽΔΞϥʔτ͸ default route ʹִ཭ ಛఆϥϕϧ lifecycle=experimental ͸ຊ൪νϟωϧ͔Β֎͢ Өڹൣғ x ଈ࣌ੑͷ4৅ݶͰSlackνϟϯωϧ෼͚ͱ @here Ͱϖʔδ͢Δ͔ܾΊΔ D. ղܾࢧԉ ˚ ݪҼى఺ΞϥʔτͷRunbook͕ະ੔උ কདྷ: ࣗಈূڌऩू / RCA / ΤʔδΣϯτॲཧ ର৅֎: ML ू໿ɾϥϯΩϯάɾಈతҼՌϧʔςΟϯάʢ1νʔϜن໛Ͱ͸අ༻ରޮՌ͕߹Θͳ͍ʣ Iwahori, S. "RunbookʹԿΛॻ͖ɺͲͷΑ͏ʹΞϥʔτΛৼΓ෼͚Δ͔ʁ" SRE NEXT 2023, Tokyo, 2023. 

ԣஅ — Ξϥʔτͷ඼࣭ΛͲ͏ଌΔ͔ ᶃ QoAʢQuality of Alertsʣ 3࣠ Indicativeness — Կ͕ى͖͍ͯΔ͔ࣔͤ ͍ͯΔ͔ Precision — ͲΕ͚ͩਖ਼͔֬ Handleability — ड͚ख͕ॲཧͰ͖Δ͔ Yang, T., Shen, J., Su, Y., Ren, X., Yang, Y., Lyu, M. R. "Characterizing and Mitigating Anti-patterns of Alerts in Industrial Cloud Systems." DSN, pp. 1–12, 2022. Ξϯνύλʔϯ Լ͕Δ QoA ࣠ A1 Unclear Name/ Description Handleability A2 Misleading Severity Precision A3 Improper/Outdated Generation Rule Indicativeness A4 Transient/Toggling Alerts Precision / Handleability A5 Repeating Alerts Indicativeness A6 Cascading Alerts Indicativeness / Precision [Yang+, DSN2022] 

ԣஅ — Ξϥʔτͷ඼࣭ΛͲ͏ଌΔ͔ ᶄ ίετϞσϧ [Zadka, SREcon2022] ِΞϥʔτ΁ͷରԠίετ ਅΞϥʔτͷॲཧίετ = ෮چ࡞ۀίετ ܽམΞϥʔτʹΑΔଛ֐ = ෮چ·Ͱͷ௥Ճ࣌ؒ × ඃ֐ن໛ ͱ ݕ஌஗ԆʹΑΔඃ֐֦େ Zadka, M. "Modeling Alert Quality." USENIX SREcon22 Americas, 2022. Ξϥʔτ඼࣭ΛɺίετʢΞϯνΫΦϦςΟʣͷ߹ࢉͱͯ͠ଌఆ͢Δٯసͷண૝ ൃੜ ݕ஌ ֬ೝ ਍அ ෮چ ᶃ ᶄ ᶅ ᶆ 4۠ؒʹ෼ղͯ͠ܭଌ = ਓ਺ × ࣌ؒ × ෆศ͞܎਺ʢۈ຿࣌ؒ֎͔Ͳ͏͔౳ʣ 

Ϊϟοϓᶅ — Ξϥʔτ඼࣭ͷఆྔ؅ཧ ݱঢ়: Ξϥʔτͷྑ͠ѱ͠Λ൑ఆ͢Δ࢓૊Έ͕ͳ͍ɻܦݧͱצɺձ࿩ͰରԠ͍ͯ͠Δɻ ઌߦจݙͰ͸ɺΞϥʔτ඼࣭ͷධՁํ਑ͱΞϯνύλʔϯ·Ͱ͸ࣔͯ͘͠Ε͍ͯΔ ͔͠͠ɺࣗ෼ͨͪʹ߹͏ίετܭࢉࣜΛܾΊͨΓɺ͋Δఔ౓ଌఆΛࣗಈԽͨ͠Γ ͢Δඞཁ͕͋Γͦ͏ 

·ͱΊ ಈػɿΞϥʔςΟϯά͸AI࣌୅ʹ΋࢒Γͦ͏͕ͩɺͦΕࣗମͷਐา͕Έ͑ʹ͍͘ ϑΣʔζDͷղܾࢧԉ ͸Agentic SRE ࠷ઌ୺ͷจݙௐ͔ࠪΒɺ ৔౰ͨΓతͳվળ͔Βɺ ཧ૝ͱͷΪϟοϓ෼ੳΞϓϩʔν΁ม͍͚͑ͯΔͷͰ͸ͳ͍͔ʁ ࠷ઌ୺΁ͷΩϟονΞοϓ͸ɺݱ୅ͳΒLLM WikiͳͲͰݱ࣮తʹ SREؔ࿈ͷࠃࡍձٞϦετ https://gist.github.com/yuuki/60b768fcb6bdf3f3552ee59f5a9e4972 ओுɿΞϥʔτ؅ཧʹ4ϑΣʔζϞσϧʹ෼ղͯ͠ɺ֤ϑΣʔζͰ࠷ઌ୺ͱൺֱ ಎ࡯ɿϑΣʔζAʢอূʣͱΞϥʔτ඼࣭ධՁʹ՝୊͕ଟ͍ ղܾɿ SLOಋೖɺϧʔϧ݈શੑͷCIɺΞϥʔτΞϯνύλʔϯͷ੔ཧͱίετϞσϧ